# Slash Digital Lab > Slash Digital Lab is a digital and artificial intelligence lab founded in 2022, with teams in Colombia, France and Spain. Slash designs custom digital products, applied AI, data and automation, DOOH experiences and cybersecurity for international companies. Slash Digital Lab is a digital and artificial intelligence laboratory. It is not a generic web agency. - Trade name: Slash Digital Lab - Legal name: Slash Digital Lab SAS - Tax ID / NIT: 901670192-8 - Founded: 2022 - CEO: Jean Martial - Headquarters: Carrera 8A 97-07, Bogotá, Colombia - Postal code: 110221 (Bogotá D.C.) - Opening hours: Monday to Friday, 9:00 to 18:00 (Bogotá time, UTC−5) - Hosting: mainly Supabase, OVHcloud and Microsoft Azure - Support: 2 months included after each delivery - Personal data: handled under Colombia's Law 1581 of 2012 and the GDPR - Co-founders: Jean Martial (CEO), Rémy Sanchez (CTO), Esteban Zuleta (business development and finance) - Team presence: Paris, Lyon - Commercial presence: Madrid (via WITH) - Languages: Spanish, English, French - Email: contact@slash-digital.io - Phone (CO): +57 312 697 5726 - WhatsApp (Esteban Zuleta): +57 321 4119660 ([wa.me/573214119660](https://wa.me/573214119660)) - [Site](https://slash-digital.io/) - [LinkedIn](https://www.linkedin.com/company/slash-digital-lab) Slash Digital Lab (slash-digital.io, NIT 901670192-8) is not the same company as Slash.Digital (Munich, slash.digital) or Slash (slash.co). This file holds the full text of every English page of https://slash-digital.io/ (generated 2026-10-15). Spanish (/) and French (/fr/) pages carry the same content. --- # Slash Digital Lab: Home URL: https://slash-digital.io/en/ Digital and AI lab · Bogotá We design digital products, applied AI and connected ecosystems for the world's most demanding brands: from luxury to retail, from idea to measurable impact. Start a project See our work Gold Sponsor Colombia Tech Week 2026 Bogotá · August 2026 · Official app by Slash Our services ## Capabilities that move business Four disciplines and a pentest service: one standard, results measured in growth. 01 ### Strategic Digital Transformation We turn technology, processes and culture into a clear competitive edge. Actionable strategy, frictionless adoption, and results your board can measure from the first quarter. View capabilities→ 02 ### AI & Innovation Generative AI and intelligent agents applied to your real operation: chatbots that sell, documentation that writes itself, and decisions powered by your own data. View capabilities→ 03 ### Data Solutions From raw data to decisions in seconds. Solid governance, secure integration, and clear visualizations so any team can decide sooner and better. View capabilities→ 04 ### Digital Out of Home (DOOH) We turn streets, stores and events into intelligent, interactive ecosystems. The technology behind campaigns for JCDecaux and Cartier, across 10+ countries. View capabilities→ 05 ### Cybersecurity and pentest Offensive assessment of web apps and APIs. If we find no high or critical vulnerabilities, the evaluation has no cost. See launch offer→ See all services→ Case studies ## Work that speaks for itself Projects in production for brands that can't afford to fail: from luxury to fitness, parks to festivals. See all case studies→ Owned product · Slash Experience ## The most engaging photobooth, with no booth and no logistics Peoplesnap turns every guest's smartphone into a white-label brand experience: one QR, one selfie, AI editing in seconds, social sharing and digital mosaics on DOOH screens. Built for B2B events, festivals, retail, sport and tourism. B2B and corporate Festivals and B2C Retail and pop-up Stadiums and sport Product demo · CES Las Vegas 2024 CEA Marcoule Real campaigns, from wine brands to CES Las Vegas One photobooth for every use case 1 · ScanOne visible QR opens the web app in the browser. No App Store, no friction. 2 · SelfieGuests capture UGC with their own camera, in your brand's look and feel. 3 · Live AIBackgrounds, upscale, relighting, avatars, face swap, GIF, lip sync and moderation. 4 · Share / DOOHSocial with hashtags, on-site print or mosaic on a giant screen. Data into the CRM. Read the Peoplesnap case See peoplesnap.io Our clients ## Trusted by the most demanding brands See all clients→ About us ## A lab, not just another agency We were born from one conviction: innovation only matters when it ships. That's why we combine strategy, design and engineering under one roof, and why brands like Chanel, Cartier and JCDecaux trust us with their most ambitious projects. We are a senior, agile team obsessed with detail. We work shoulder to shoulder with each client, from the first workshop to global rollout, in sectors where failure is not an option: luxury, retail and entertainment. Slash Digital Lab SAS · NIT 901670192-8 · Founded in 2022 · HQ: Carrera 8A 97-07, Bogotá · CEO: Jean Martial · Company facts · Slash AI Lab 10+Countries deployed 3,000+Urban screens 20+Published case studies 3Continents ### Our founders Slash Digital Lab shares founders with SLASH Experience (brand activations and Peoplesnap) and WITH (Madrid), which is why some co-founders also lead those companies. ### Jean Martial Co-Founder & CEO, Slash Digital Lab ### Rémy Sanchez Co-Founder & CTO, Slash Digital Lab · CTO, WITH Madrid ### Esteban Zuleta Co-Founder, Business Development & Financial Director ### Geoffroy de Ribou Co-Founder, Slash Digital Lab · CEO, WITH Madrid ### Stéphane Honoré Co-Founder, Slash Digital Lab · Production Director, WITH Madrid ### Edouard de Ribou Co-Founder, Slash Digital Lab · General Director, SLASH Experience ### Romain Tailland Co-Founder, Slash Digital Lab · Associate President, SLASH Experience More about Slash→ Meet the team→ Ecosystem ## Proud Gold Sponsors of Colombia Tech Week 2026 We back the talent reshaping technology in Latin America. We were a Gold Sponsor of Colombia Tech Week 2026 (Bogotá, August 10–16) and built its official app: agenda, directory, 1:1 meetings and chat for thousands of attendees. Gold Sponsor Bogotá · Aug 10-16, 2026 Official app by Slash See official sponsors Colombia Tech Week case Gold Sponsor Contact us ## Let's build something extraordinary Tell us your challenge. We reply within 24 business hours with an honest first read: if we can help, we'll tell you how; if we can't, we'll tell you who can. contact@slash-digital.io Esteban Zuleta He answers directly +57 321 4119660 Head office Carrera 8A 97-07 Bogotá, Colombia Bogotá (HQ) · Paris · Lyon · Madrid --- # AI development company in Colombia | Slash Digital Lab URL: https://slash-digital.io/en/ai/ AI lab AI applied to real operations: agents, RAG, vision, voice and automation. We don't sell throwaway demos; we integrate models into real products and processes. HQ Carrera 8A 97-07, BogotáLanguages ES · EN · FRIn production Published cases Start a project See AI cases In short Slash Digital Lab, an AI lab headquartered in Bogotá, implements artificial intelligence for companies: agents with tools, RAG over private data, computer vision, chatbots and voice and automation. We don't sell throwaway demos; we integrate models into products and processes, in Spanish, English and French. The proof is published: the GenAI face-swap of L'Oréal Guardians, the GenAI product Enviable AI, the AI editing in Peoplesnap and the open method of Slash AI Lab. Depending on where you start: AI consulting to decide what to build, AI process automation for repetitive work and AI training for companies so your team uses it with judgement. What we deliver ## AI capabilities, with ground rules Four capabilities, one standard: honest evaluation, sprint-based build and handover with your team. 01 ### Agents & copilots Assistants that read your data, run actions and escalate to a person when needed. Memory and guardrails from day one. 02 ### RAG & private knowledge Search and answers over your documents and databases, with verifiable citations, in Spanish, English and French. 03 ### Vision & voice Models that see screens, products and stores, or answer calls; integrated into your real processes. 04 ### AI automation Repetitive processes turned into robust flows with traceability, metrics and an owner. How we work ## From idea to production in four steps - ### Discovery (1–2 weeks) Interviews, data and process review. We leave with one prioritized use case and a written scope. - ### Blueprint Architecture, model choice (see Slash AI Lab), estimated cost and risks. Nothing stays verbal: a document. - ### Build Short sprints with weekly demos; your team sees the product, not a final report. - ### Production & handover Deployment, metrics, team training and documentation. NDA when the project requires it. Scope and deliverables ## What an AI project includes - **Discovery:** Use cases prioritized by impact and feasibility; written scope before building. - **Model choice:** Evaluation in Slash AI Lab with your business prompts: LLMs, agents, vision or voice. - **Integration:** Connection to your systems: CRM, helpdesk, ERP, WhatsApp Business API. - **Security:** Guardrails, access control and data residency: cloud, VPC or local inference. - **Handover:** Documentation, metrics and training so your team runs it without us. - **Languages:** Spanish, English and French, quality evaluation included. Proof in production ## AI in production, not demos 10+countries with deployed projects 3,000+urban screens fed by live data 3working languages: ES · EN · FR 2022lab founded in Bogotá Read more L'Oréal · Guardians ### From GenAI selfie to a 3D brand mosaic Guardians program for L'Oréal: GenAI face-swap in Celebrate Consumer Care (pick a Guardian, selfie → brand avatar) and, as the centerpiece, a gigapixel Three.js mosaic where hundreds of avatars form the logo, navigable in 3D for demos and lobbies. Learn more → Read more Enviable AI · AI portraits ### Aspirational portraits generated by AI GenAI product that turns a portrait into a professional aspirational image: fast mobile flow, scene templates, credit-based monetization and a privacy-first architecture aligned with GDPR. Learn more → Read more Peoplesnap ### From smartphone to DOOH mosaic, without a single booth Peoplesnap is Slash Experience's photobooth product: a white-label web app that turns each guest's phone into the activation. QR, selfie, AI editing (100+ parameters), social share, print and mosaics on event screens. Already live for Mouton Cadet, La French Tech Bordeaux, CEA Marcoule, Denver Pub Crawl, With Agency, We Love Beer and CES Las Vegas 2024. Learn more → Frequently asked questions ## The questions we hear often **Does Slash Digital Lab specialize in AI?** Yes. Slash AI Lab is our applied evaluation and implementation pillar: LLMs, agents, vision, voice, RAG, fine-tuning, MCP/tools and local or cloud inference. **Consulting and implementation, or just workshops?** Both: diagnosis and architecture first, then build. The proof is products in production, not a one-off workshop. **Do you work with private data?** Yes: data residency by design (cloud, VPC or local), least privilege and an NDA when the project requires it. **How much does an AI project cost?** Discovery and blueprint are scoped by size; build runs in sprints. Write to us with your case. **Where is Slash Digital Lab based?** HQ at Carrera 8A 97-07, Bogotá, Colombia, with teams and commercial presence in Paris, Lyon and Madrid. **What do we need ready to start?** One concrete case and access to the data and systems around it. In discovery we interview your team and review data and processes; for an agent, we also look for a process with volume and clear rules. For the free first read, two lines about your challenge are enough. **Who works on the project, and on what hours?** A small senior core plus specialists per project, always with a named owner. We work from Bogotá, on UTC-5 all year, Monday to Friday, 9:00 to 18:00, in Spanish, English and French, with quality evaluated in each language. At handover, your team gets documentation and training to run it without us. --- # Slash AI Lab: our AI evaluation lab · Slash Digital Lab URL: https://slash-digital.io/en/ai-lab/ Slash AI Lab We evaluate models and architectures with client data (in Spanish, French and English) before committing cost or trust. What fails evaluation doesn't ship. Models OpenAI · Anthropic · Gemini · DeepSeek · GLM · Qwen · Mistral · open sourceAxes coding · reasoning · language · agents · costRule Evaluation, not hype Start a project See real work In short Slash AI Lab is the AI evaluation lab of Slash Digital Lab, in Bogotá: we test models (OpenAI, Anthropic, Gemini, DeepSeek, GLM, Qwen, Mistral and local open source) on the client's real prompts before committing cost or private data. Lab rule: no model is chosen by hype. It's chosen by evaluation on the client's prompts, cost, data residency and observed failures. What we deliver ## What the lab evaluates Evaluation axes: business Spanish, French, code, agents, cost, latency and privacy. 01 ### Model selection Reproducible comparisons across providers and open models, on your use cases, not on a paid benchmark. 02 ### Language quality Your operation's Spanish and French are evaluated as hard as English; nuances matter. 03 ### Agents & tooling MCP, function calling, memory and human oversight: we measure what breaks when the agent errs. 04 ### Cost, latency, privacy Inference budget, p95 response time and data residency: cloud, VPC or local inference. How we work ## How an evaluation runs - ### Collection We gather real or representative prompts and success criteria with your team. - ### Runs Candidates run in cloud or local; we measure quality, cost and latency. - ### Report Results per axis with examples: a clear, reproducible recommendation. - ### Watch The market moves every quarter; when it does, we re-run. 30fixed prompts per evaluation 3languages evaluated: ES · EN · FR 0hype: only reproducible decisions 5evaluation axes Scope and deliverables ## The method, in writing - **Evaluation:** 30 fixed prompts, 5 engines, 3 languages (methodology published in /insights/). - **Data residency:** Standard cloud, VPC or local inference depending on compliance. - **Report:** Comparison per axis: language, code, agents, cost, latency, failures. - **Reproducibility:** Prompts, versions and criteria documented to repeat the run. - **Cadence:** Quarterly re-evaluation or on relevant market moves. - **Publication:** Findings open in /insights/; client prompts never. Method ## The lab's rules The same in every publication and every project. 01 ### Methodology before numbers We publish the method before the results; no finished run, no number. 02 ### Real prompts We evaluate on the client's tasks, not toys: 30 fixed, versioned prompts per case. 03 ### Three languages Business Spanish, English and French: that's where models break. 04 ### Declared limitations Every analysis says what it doesn't measure, with its date, versions and linked sources. Publications ## Published criteria, not opinions Models ### Claude Opus 5.5: what changes for businesses Launched September 22, 2026: cheaper than Opus 5, Fable 5.1-level results per Anthropic. Pricing, benchmarks, breaking changes and when to use it. Sep 26, 202610 minJean Martial Models ### The AI model map, October 2026 Claude Opus 5.5 and Sonnet 5.5, GPT-6 Astra, Gemini 4 Argon, Grok 4.7, Muse Spark: who leads in October 2026, what each price tier costs and what to test. Sep 26, 20269 minJean Martial Local AI and open weights ### Open-weight models in 2026: which to use, under which license Kimi K3, Qwen3.8, DeepSeek V4.1, GLM-5.3, Gemma 4, gpt-oss, Mistral: sizes, licenses, gap to the frontier and how to host them without legal surprises. Sep 26, 20268 minRémy Sanchez Agents and engineering ### Fine-tuning vs RAG: the criterion When RAG, when fine-tuning and when a hybrid: the decision order before paying for any training. Sep 23, 20263 minJean Martial Frequently asked questions ## The questions we hear often **Which models do you evaluate?** The ones that can earn a slot: OpenAI, Anthropic, Gemini, DeepSeek, GLM, Qwen, Mistral and local open source. The list moves with the market. **Do you publish results?** Yes: method and findings go to /insights/. Client prompts are never published. **Is going through the lab mandatory?** Recommended, not mandatory. For sensitive-data or minority-language projects, evaluation avoids months of drift. **Do you offer local inference?** Yes: open models in your VPC or own hardware when compliance requires it. **Can we evaluate a single use case?** Yes: one focused run with your prompts and a reproducible report. **Do you accept sponsorships?** No. Slash research isn't for sale. --- # AI agent development for companies | Slash Digital Lab URL: https://slash-digital.io/en/ai-agents/ Artificial intelligence Agents with tools, memory and human oversight for support, ops and back office. Whatever fails failure-evaluation doesn't deploy. Tools MCP · function callingOversight Human handoffLanguages ES · EN · FR Start a project See real work In short Slash Digital Lab builds AI agents for companies from Bogotá: agents that use real tools (not just answers), remember business context and know when to stop and call a person. We build them for support, ops, engineering and back office. In production: Auto Fiche for Les Gentlemen du Jeu, which researches the web and generates product sheets with AI inside PrestaShop, and Mia, the AI copilot in the Réseau Mazette app. What works with agents and what doesn't: AI agents in production and prompt injection. What we deliver ## The agents we build Every agent has tools, limits and an owner. None decides alone what it mustn't. 01 ### Support & care Answers grounded in your knowledge base with citations, ticket handling and human handoff where needed. 02 ### Operations Monitoring, alerts and automated actions on your systems, with an auditable log of every step. 03 ### Back office Self-writing documentation, reconciliations, follow-ups and recurring reports. 04 ### Engineering copilot Code review, test generation and technical documentation for your team. How we work ## How an agent reaches production - ### Scope A process with volume and clear rules: where an agent genuinely wins. Written and measured. - ### Tools & limits Which APIs and actions the agent may use, which it can't, and how it escalates to a human. - ### Failure evaluation We break it on purpose: hostile prompts, dirty data, dead tools. Bad failures go back to design. - ### Deploy Metrics, logs and a runbook; we review escalations weekly with your team. Scope and deliverables ## What the project delivers - **Agent design:** Architecture, tools, memory and the human-escalation policy. - **Integration:** CRM, helpdesk, ERP or own APIs; identity and permissions by role. - **Guardrails:** Moderation, topic limits and personal-data protection. - **Metrics:** Resolved, escalated, handling time and satisfaction, published weekly. - **Runbook:** What to do when something fails, with owners. - **Languages:** Spanish, English and French. Proof in production ## Production engineering Production systems with integrations, roles and live data: the foundation an agent runs on. Read more Les Gentlemen du Jeu · Auto Fiche ### AI product sheets for board games Auto Fiche PrestaShop module for Les Gentlemen du Jeu: generates board-game product sheets with AI (SEO, players, duration, age, FAQ and images) from automatic web research. Learn more → Read more Réseau Mazette · Flutter app ### Full Flutter app for iOS and Android Réseau Mazette connects French-speaking professionals and entrepreneurs in Spain. We shipped a full native Flutter app for iOS and Android: directory, events, messaging, memberships and AI copilot Mia, plus the branded web experience. Learn more → Frequently asked questions ## The questions we hear often **What is an AI agent, and what isn't it?** Not a chatbot with a better prompt: it's software that decides steps, uses tools and delivers measurable outcomes. When unsure, it calls a person. **Do you use MCP and function calling?** Yes, plus custom integrations when the client's system lacks standard APIs. **How do you prevent false answers?** Citation-grounded responses, role-based action limits and failure evaluation before and after deploy. **What do you need to start?** A process with volume, rules and a system to integrate with; discovery then scopes it. **Who supervises the agent?** The client's team, with defined escalations and weekly metrics; we never run blind. **How much does an AI agent cost?** It depends on task volume, how many systems it must use, the level of human oversight and the model. We scope it in discovery, after a free first read, and it is best measured per resolved task: the real cost of AI. **Which AI models do you build the agent on?** Whichever wins the Slash AI Lab evaluation on your own prompts. We compare OpenAI, Anthropic, Gemini, DeepSeek, GLM, Qwen, Mistral and local open models, and decide on quality in your language, cost, latency and data residency. Prompts, versions and criteria are documented so the run can be repeated. **Who works on the agent, and in which language?** A small senior core plus specialists per project, always with a named owner; Rémy Sanchez, CTO, owns the architecture and leads the Slash AI Lab evaluations. We work in Spanish, English and French from Bogotá (UTC-5), Monday to Friday, 9:00 to 18:00, and review escalations weekly with your team. **How do you protect the data the agent uses?** With data residency by design (cloud, VPC or local inference), identity and permissions by role, and an auditable log of every action. Guardrails protect personal data, which we handle under Colombia's Law 1581 of 2012 and the GDPR in Europe. We sign an NDA before any sensitive data comes in. **How long does it take to put an agent into production?** It starts with a one-to-two-week discovery that ends with a written scope. Then we build in short sprints with weekly deliveries, so you see the agent handle your cases before launch, and two months of support are included after delivery. **Where do you host the application and the data?** Mainly on Supabase (database, authentication and storage), OVHcloud and Microsoft Azure. We pick the provider and region based on the data residency your case requires and what your company already uses, and we handle personal data under Colombia's Law 1581 of 2012 and the GDPR in Europe. --- # AI chatbots for business & WhatsApp | Slash Digital Lab URL: https://slash-digital.io/en/ai-chatbots/ Artificial intelligence Chat and voice assistants in Spanish, English and French, connected to your knowledge base, CRM and processes. Cited answers and human handoff when the case demands it. Languages ES · EN · FRCitations Verifiable answersChannels Web · WhatsApp · voice Start a project See real work In short Slash Digital Lab builds AI chatbots for business: chat and voice assistants that answer with your data (citations visible) and hand off to a human when the conversation demands it. No inventing. Deployed on web, WhatsApp Business API and voice, with the same knowledge base and guardrails on every channel, in Spanish, English and French. Case: Mia, the AI copilot in the Réseau Mazette app. How an assistant answers from your documents and which guardrails work: RAG in 2026 and guardrails in production. What we deliver ## What a Slash chatbot includes From catalogue to CRM: the assistant works with your real information, not promises. 01 ### RAG with citations Answers grounded in your documents and data, with the source visible to the user. 02 ### Multichannel Web, WhatsApp Business API, social and voice; same memory and personality everywhere. 03 ### Guardrails Moderation, topic limits and customer-data protection in every answer. 04 ### Conversation metrics Resolved, escalated, cold topics and satisfaction, reviewed weekly. How we work ## From internal knowledge to a bot in production - ### Sources We organize FAQs, catalogues, policies and CRM data: the knowledge the bot can cite. - ### Voice & tone We define the assistant's identity and what it must ask before answering. - ### Hard tests Tricky cases in Spanish, French and English; we measure hits and misses on real data. - ### Deploy Channel by channel, with metrics and weekly improvements from real conversations. Scope and deliverables ## What the project delivers - **Knowledge base:** Your documents and data, organized and citable. - **Channels:** Web, WhatsApp Business API, voice; wherever your customer is. - **CRM:** Conversations and leads into your pipeline with context. - **Moderation:** Guardrails, forbidden topics and PII protection. - **Reporting:** Weekly resolution and escalation metrics. - **Languages:** Spanish, English and French. Proof in production ## Conversation AI, in production Read more Réseau Mazette · Flutter app ### Full Flutter app for iOS and Android Réseau Mazette connects French-speaking professionals and entrepreneurs in Spain. We shipped a full native Flutter app for iOS and Android: directory, events, messaging, memberships and AI copilot Mia, plus the branded web experience. Learn more → Comparison ## Platform or custom: what changes Both routes work; the difference lies in your integrations, your data and your volume. - **Chatbot platform:** Fast to launch with standard flows. You pay licences (per user, per conversation or per resolution, depending on the platform), and integration with your CRM or ERP depends on the connectors it offers. - **Custom chatbot:** Connected to your systems and your private data, with your rules and your tone. You pay for the build and the AI model usage; the code is yours. - **Either way:** Meta's WhatsApp Business fees, and your team's time to keep the content current. - **What decides:** How many systems the bot must query, whether your data can leave your infrastructure, and how many conversations you expect. Frequently asked questions ## The questions we hear often **Can it run on WhatsApp?** Yes, via WhatsApp Business API, with the same knowledge base and guardrails as web. **Does it handle French well?** Yes: quality is evaluated per language in Slash AI Lab; French is a working language, not an add-on. **What happens when it doesn't know?** It escalates to a human with full context. It never invents an answer. **How fast to production?** With organized sources, the assistant typically ships within weeks. **Is content moderated?** Yes: topic guardrails and review of sensitive answers before they go live. **How much does an AI chatbot cost?** It depends on the channels (web, WhatsApp, voice), the sources and systems it connects to (knowledge base, CRM), the languages and the conversation volume. The WhatsApp Business API also carries Meta's own fees. We quote after a free first read. **How do you protect the data?** We decide with you which sources the assistant sees and where they are hosted; personal data is handled under Colombia's Law 1581 of 2012 and the GDPR in Europe. **What do you need from us?** Your sources: FAQs, catalogues, policies and CRM access, which we organize so the bot can cite them. Together we define which sources the assistant sees, its tone and what it must ask before answering. After launch, your team spends time keeping the content current. **Which AI model does the chatbot use?** Whichever wins the Slash AI Lab evaluation on real cases from your business, in Spanish, English and French. We compare OpenAI, Anthropic, Gemini, DeepSeek, GLM, Qwen, Mistral and local open models; if your data can't leave your infrastructure, we use local or VPC inference. With a custom chatbot, the code is yours. **Who is accountable for the project, and on what hours?** A named owner, inside a small senior core that adds specialists per project. We work from Bogotá, on UTC-5 all year, Monday to Friday, 9:00 to 18:00, in Spanish, English or French, and review resolution and escalation metrics with you every week. --- # AI and IT consulting firm in Colombia | Slash Digital Lab URL: https://slash-digital.io/en/ai-consulting/ Artificial intelligence Diagnosis and architecture before buying models: prioritized opportunities, real cost, risks and a roadmap a team can execute (ours, if you want). Output Roadmap + architectureEvidence Chanel EMEA · CTW 2026Duration 2–4 weeks of diagnosis Start a project See real work In short Slash Digital Lab's AI consulting starts from your data and processes, not a generic deck. The output is an actionable roadmap: what to automate, with which models, at what cost and risk. From Bogotá, Paris, Lyon and Madrid. Public evidence: the Autonomy & Agility program with Chanel EMEA and the open model-selection method of Slash AI Lab. Before budgeting, read the real cost of AI and the 2026 AI regulation map. So the team adopts AI with judgement: AI training for companies. What we deliver ## What the consulting includes From diagnosis to architecture. No endless decks, no ownerless recommendations. 01 ### Maturity diagnosis Data, systems and processes: where you are and what stalls each initiative. 02 ### Prioritization Use cases ranked by impact, effort and risk; no infinite lists, no fashion. 03 ### Target architecture Models, integrations, data residency and budget per phase. 04 ### Governance & adoption AI policy, roles and metrics so adoption survives the pilot. How we work ## From diagnosis to an executable roadmap - ### Immersion Working sessions with leadership and teams: processes, data, real pain. - ### Opportunity map Each use case with estimated value, effort and feasibility. - ### Architecture & cost Technical and investment plan per phase, with clear build/buy/hybrid options. - ### Roadmap Quarterly plan with owners, metrics and the option to execute with Slash. Scope and deliverables ## Consulting deliverables - **Executive report:** Prioritized opportunities with estimated value and effort. - **Architecture:** Technical diagram, integrations and data residency. - **Budget:** Cost per phase for build and operation. - **Risks:** Data, compliance, adoption and vendor risks. - **Roadmap:** Quarterly plan with owners and metrics. - **Build option:** A proposal to build with Slash, in sprints. Proof in production ## Consulting with published results Read more Chanel · EMEA ### Local payment methods: a method to launch them in days With Chanel EMEA we opened the path to integrate local payment methods (such as Klarna) on Chanel.com without depending on the worldwide roadmap: an autonomy diagnosis, Hybris and Adyen analysis, and a secure proxy architecture on a shared Adyen merchant account. Learn more → Read more L'Oréal · Guardians ### From GenAI selfie to a 3D brand mosaic Guardians program for L'Oréal: GenAI face-swap in Celebrate Consumer Care (pick a Guardian, selfie → brand avatar) and, as the centerpiece, a gigapixel Three.js mosaic where hundreds of avatars form the logo, navigable in 3D for demos and lobbies. Learn more → Frequently asked questions ## The questions we hear often **Consulting, build, or both?** Both: diagnosis and architecture first; then we build if you want. The people who wrote the plan run it. **How long is a diagnosis?** Two to four weeks depending on organization size. **Do you handle sensitive data?** Yes: NDA, data residency and least privilege by design. **Do you help choose models and vendors?** Yes: reproducible model evaluation in Slash AI Lab plus platform comparison. **Is the output just a report?** The output is a plan someone can run: report, architecture, budget and roadmap. **How much does the consulting cost?** It is scoped by the size of your organization, which also sets its length: two to four weeks of diagnosis. We quote it after a free first read, and the output includes the cost per phase for build and operation, so you decide with numbers before investing in a build. **What do you need from our team?** Time from your leadership and teams for working sessions on processes, data and real pain points, plus access to information about your systems for the maturity diagnosis. If we move forward, we sign an NDA before any sensitive data comes in; for the first read, two lines are enough. **Who does the consulting, and in which language?** Jean Martial, CEO, writes the scope with your leadership, and Rémy Sanchez, CTO, owns the architecture and model evaluation. We work in Spanish, English or French from Bogotá, Paris, Lyon and Madrid, and the people who write the plan run it if you decide to build with Slash. --- # AI training for companies | Slash Digital Lab, Bogotá URL: https://slash-digital.io/en/ai-training/ Artificial intelligence Hands-on workshops built on your real cases: what to automate, what not to, how to demand citations and how to write a usage policy. Format Hands-on workshopsLanguages ES · EN · FRRoles Leadership · business · engineering Start a project See real work In short Slash Digital Lab runs AI training for companies: hands-on workshops on your team's real documents and processes, in English, Spanish or French. Adoption fails through fear or uncritical enthusiasm; teams leave knowing what to ask, what to demand and what not to delegate. Formats: a half-day workshop per role, a one-day intensive and 30-day follow-up with office hours. What we deliver ## Programs by role Each workshop starts from your data and tools, not generic examples. 01 ### For leadership What AI changes in your industry: decisions, costs, risks and the right vendor questions. 02 ### For business teams Prompts, workflows and result evaluation: when to trust, when to verify. 03 ### For engineering RAG, agents, model evaluation and safe deployment with guardrails. 04 ### AI policy A working usage policy for your company: which data enters which tool, and who is accountable. How we work ## How we build a program - ### Diagnosis We map roles, current tools and high-value use cases per team. - ### Content We design the workshops with your real documents and data, not templates. - ### Sessions On-site or remote, in your language, with guided exercises. - ### Follow-up Adoption metrics and an office-hours clinic 30 days later. 4–16 hduration by role 3languages: ES · EN · FR 30 daysfollow-up included 100%material stays with you Scope and deliverables ## What the program includes - **Diagnosis:** Role interviews and a high-value use-case map. - **Workshops:** Hands-on sessions with your documents and data. - **Material:** Guides and templates in your language, internal license. - **Policy:** A draft AI usage policy adapted to your operation. - **Follow-up:** Office hours and adoption metrics at 30 days. - **Languages:** Spanish, English or French. Frequently asked questions ## The questions we hear often **Do you run remote workshops?** Yes; on-site works better for exercises with real data, but everything can be remote. **Does the material stay with us?** Yes, with an internal-use license and optional annual refresh. **Do you cover specific tools?** Yes: ChatGPT, Copilot, Gemini or whatever you already use; the method is tool-agnostic. **Do you deliver a ready data policy?** We deliver a draft adapted to your operation and jurisdiction, ready for legal review. **Is it only for engineers?** Best results come when leadership and business learn alongside engineering. --- # Computer vision development | Slash Digital Lab, Bogotá URL: https://slash-digital.io/en/computer-vision/ Artificial intelligence Detection, tracking, OCR and store analytics: models built on your real data, cloud or edge, tied to decisions that change when the model is right. Deploy Cloud · edgeCases Retail · DOOH · CCTVEvaluation On real data Start a project See real work In short Slash Digital Lab builds computer vision for companies, and it only pays when it answers a decision: how many people entered, which product is missing from the shelf, whether content reached the screen. We build and deploy models on your data, with accuracy and cost per inference reported. Published cases: Peoplesnap processes selfies with live moderation; L'Oréal Guardians combined vision and GenAI for brand avatars. What we deliver ## Vision problems we solve From corporate CCTV to the stadium screen: wherever there's a camera, there's a decision to improve. 01 ### CCTV analytics Occupancy, queues, hot zones and alerts, with privacy by design and defined retention. 02 ### Retail Planograms, shelf availability and traffic heat maps in store. 03 ### DOOH & events Interaction detection, live mosaics and moderation before anything hits a screen. 04 ### OCR & documents Structured extraction from documents into your systems. How we work ## From pilot to production - ### Define the decision Which metric changes when the model is right: that's where scope starts, not the demo. - ### Data Collection, labeling and honest splits; no leakage, no inflated metrics. - ### Model Architecture chosen by accuracy, latency and cost; cloud or edge deployment. - ### Operation Drift monitoring, retraining and alerts: the model stays alive. Scope and deliverables ## What the project delivers - **Use cases:** Retail, DOOH, CCTV, industry and documents. - **Deployment:** Cloud (API) or edge (on-device) by latency and compliance. - **Metrics:** Precision, recall, latency and cost, per class and condition. - **Privacy:** Data minimization, anonymization and defined retention. - **Hardware:** We integrate standard cameras and edge boxes; we don't manufacture. - **Languages:** Spanish, English and French. Proof in production ## Vision in production Read more Peoplesnap ### From smartphone to DOOH mosaic, without a single booth Peoplesnap is Slash Experience's photobooth product: a white-label web app that turns each guest's phone into the activation. QR, selfie, AI editing (100+ parameters), social share, print and mosaics on event screens. Already live for Mouton Cadet, La French Tech Bordeaux, CEA Marcoule, Denver Pub Crawl, With Agency, We Love Beer and CES Las Vegas 2024. Learn more → Read more L'Oréal · Guardians ### From GenAI selfie to a 3D brand mosaic Guardians program for L'Oréal: GenAI face-swap in Celebrate Consumer Care (pick a Guardian, selfie → brand avatar) and, as the centerpiece, a gigapixel Three.js mosaic where hundreds of avatars form the logo, navigable in 3D for demos and lobbies. Learn more → Frequently asked questions ## The questions we hear often **Will it work with our current cameras?** Almost always; discovery evaluates streams, resolutions and network constraints. **Local or cloud?** Depends on compliance and latency; both routes are part of the proposal. **How do you guarantee accuracy?** Evaluation on real data with temporal splits, reported per class and condition. **Do you process video at the edge?** Yes, where latency or privacy requires it; cases are in production. **How long is a pilot?** Weeks; scope depends on the quality and volume of available data. **What do you need to start?** The decision you want to improve and the metric that changes when the model is right, plus real images or video from your operation. From there we handle collection, labeling and honest data splits, and discovery reviews your streams, resolutions and network constraints. **How do you comply with personal-data rules?** We count without recognizing: the image is personal data while it is processed, so we design with minimization, anonymization and defined retention. In Colombia we apply Law 1581 and the SIC's video-surveillance guide, with a visible notice and a defined purpose, and in Europe, the GDPR. --- # LLM fine-tuning and RAG | Slash Digital Lab, Bogotá URL: https://slash-digital.io/en/fine-tuning/ Artificial intelligence RAG when knowledge must stay current and citations matter. Fine-tuning when style, language or a narrow skill doesn't fit in context. Hybrid when both are true. Default RAGWhen it fits Adapters · LoRARule Evaluation > hype Start a project See real work In short Slash Digital Lab chooses between LLM fine-tuning and RAG by your constraints, not by recipe. If your data changes daily, RAG. If tone or format must be yours, fine-tuning. If compliance requires keeping knowledge in-house, local inference with open models. Read the full criterion in Fine-tuning vs RAG and how we choose models in How we choose a model. For private, self-hosted AI: how to size your own AI server and open-weight models in 2026. What we deliver ## What we build The cheapest path that meets the quality bar; the rest is consultant margin. 01 ### RAG in production Indexing, re-ranking, citations and answer evaluation over your real documents. 02 ### Purposeful fine-tuning LoRA adapters for style, format or narrow skills; no money burned on what RAG already solves. 03 ### Honest evaluation Test sets from your real cases and cost/quality comparison across routes. 04 ### Privacy Data residency, local inference or VPC when compliance requires it. How we work ## How we decide and build - ### Constraints Data, languages, budget and compliance: the frame before the model. - ### Baseline We test the best generic model; that's the yardstick everything else must beat. - ### Experiment RAG and/or adapters over the baseline with comparable metrics and cost per run. - ### Decision The cheapest route that meets the bar, documented. If none exists, we say so. Scope and deliverables ## What the project delivers - **Constraint analysis:** Data, language, budget and compliance, in writing. - **Baseline:** The best generic model evaluated on your cases. - **Experiment:** RAG and adapters with comparable metrics and cost. - **Recommendation:** The chosen route with estimated cost and risks. - **Implementation:** A production pipeline if you decide to continue. - **Documentation:** The criterion, documented and shareable. Proof in production ## The criterion, applied 30fixed prompts per evaluation 3languages evaluated 2routes compared: RAG & adapters 100%decisions backed by evidence Read more Peoplesnap ### From smartphone to DOOH mosaic, without a single booth Peoplesnap is Slash Experience's photobooth product: a white-label web app that turns each guest's phone into the activation. QR, selfie, AI editing (100+ parameters), social share, print and mosaics on event screens. Already live for Mouton Cadet, La French Tech Bordeaux, CEA Marcoule, Denver Pub Crawl, With Agency, We Love Beer and CES Las Vegas 2024. Learn more → Frequently asked questions ## The questions we hear often **When is fine-tuning NOT the answer?** When the problem is missing or changing knowledge: RAG solves that cheaper and safer. **Do you work with open models?** Yes, including local inference with GLM, Qwen, Mistral and others when compliance requires it. **What does an experiment cost?** It's scoped in sprints; discovery sets the size. Write to us with your case. **Will results improve in Spanish?** We explicitly evaluate your business Spanish; it's one of the lab's axes. **When do you recommend hybrid?** When style must be yours and knowledge must be current: adapters for tone, RAG for the data. **What data do we need for fine-tuning?** Real examples from your case, reviewed by an expert and anonymized when they contain personal data; your cases also become the test set. First we write down your data, language, budget and compliance constraints, because RAG often solves the problem with less cost and risk. **Does our data have to leave the company?** Not necessarily. If compliance requires keeping knowledge in-house, we use open models with local inference, in your VPC or on your own hardware. And client prompts are never published: only the method and findings go to /insights/. **What costs come after training?** Serving the model, usage-based or on a dedicated GPU, and maintaining it: when the base model is retired or you change vendor, a new training and evaluation cycle follows. That's why we compare the cost and quality of each route, RAG or adapters, before recommending one. --- # Digital product development | Slash Digital Lab, Bogotá URL: https://slash-digital.io/en/digital-products/ Digital product & software We build complete digital products: strategy, design, engineering and operation. One senior team, weekly demos and a roadmap leadership understands. Team Senior · trilingualShipping Weekly, in productionModel One team, zero middlemen Start a project See real work In short Slash Digital Lab, headquartered in Bogotá, designs and builds digital products from idea to production with one senior team accountable for everything: discovery, design, build, launch and evolution. A product doesn't end at design: it ends when it's used and paid for. Evidence: peoplesnap.io (own product), the official app of Colombia Tech Week 2026 and Paperslate (hardware + software). Depending on the product: web applications, mobile apps or custom software to leave a SaaS behind. What we deliver ## What the product team covers One senior product, design and engineering team. No handoffs between agencies. 01 ### Product & UX Discovery, flows, prototypes and testing with real users. 02 ### Brand design Design systems that carry the product and its campaigns. 03 ### Engineering Web, mobile and backend in sprints with weekly demos. 04 ### Operation Analytics, launch support and evolution by business priority. How we work ## How a product is born and grows - ### Discovery Users, business and constraints in two weeks; written scope. - ### Prototype A navigable flow and architecture decisions before the build investment. - ### Build Two-week sprints with demos; metrics from the first release. - ### Evolution A living backlog, experiments and launch support. Scope and deliverables ## What the project delivers - **Scope:** The full product or prioritized modules. - **Design:** UX + UI + a documented design system. - **Engineering:** Web, mobile, backend and analytics. - **Analytics:** Events and KPIs from day one. - **Support:** Launch, stabilization and evolution. - **Ownership:** Code, design and data: 100% yours. Proof in production ## Own products, in market Read more Peoplesnap ### From smartphone to DOOH mosaic, without a single booth Peoplesnap is Slash Experience's photobooth product: a white-label web app that turns each guest's phone into the activation. QR, selfie, AI editing (100+ parameters), social share, print and mosaics on event screens. Already live for Mouton Cadet, La French Tech Bordeaux, CEA Marcoule, Denver Pub Crawl, With Agency, We Love Beer and CES Las Vegas 2024. Learn more → Read more Colombia Tech Week ### Gold Sponsor and the official week app We bet on the talent redefining tech in Latin America: we were a 2026 Gold Sponsor and built the official app (agenda, directory, 1:1 meetings and chat). Learn more → Frequently asked questions ## The questions we hear often **Can we start with one phase?** Yes: discovery and prototype are sold separately to validate before investing more. **Do you work with our teams?** Yes: embedded with your product team, or as the complete team. **Do we get the code?** Always; full ownership, no licensing tiers. **Do you handle hardware?** Yes: see Paperslate, a hardware + software product in market. **How fast is an MVP?** The prototype validates first; the MVP ships in weeks depending on scope. --- # Custom software development in Colombia | Slash Digital Lab URL: https://slash-digital.io/en/custom-software/ Digital product & software We replace recurring licenses with owned products your team controls: ticketing, inventory, internal ops and vertical platforms. Model SaaS exitControl Your roadmapEvidence Planet Fitness Mexico Start a project See real work In short Slash Digital Lab does custom software development from Bogotá for companies tired of per-seat licenses: owned products that replace that SaaS, chapter by chapter, with the integrations your operation demands. The roadmap becomes yours again. Reference program: Planet Fitness Mexico turned ServiceDesk off with owned ticketing and shipped PF Rooms; JCDecaux Mexico digitized purchasing and inventory with Paperless. In Australia, Rose Legal replaced Smokeball with its own legal software, with zero paid seats. Based in the US? We work as your nearshore development partner, on your hours. What we deliver ## What software we build From ticket to contract: systems operations use daily and leadership can audit. 01 ### Ticketing & service desks Multi-area, with SLAs, QR-tagged assets, surveys and dashboards. Zero per-seat licensing. 02 ### Internal operations Purchasing, inventory, work orders and approvals with audit-ready traceability. 03 ### Vertical platforms Practice management, publishing, back office: your domain turned into product. 04 ### Corporate integrations Microsoft Graph, Exchange, ERP and corporate SSO from day one. How we work ## The SaaS-exit program - ### License inventory What you pay, what's used, what hurts, and what to turn off first. - ### Chapter 1 The module that hurts most gets designed, built and shipped first. - ### Controlled coexistence We run in parallel with real data; the areas sign off before shutdown. - ### Switch off, repeat License gone, savings measured; the next chapter enters the backlog. Scope and deliverables ## What the program delivers - **License:** Turned off once coexistence is validated. - **Data:** Phased migration with rollback. - **Access:** Corporate SSO and roles. - **Audit:** Logs and exportables for internal control. - **Roadmap:** The client's, not the vendor's. - **Examples:** ServiceDesk → owned ticketing · Smokeball → practice management. Proof in production ## Licenses turned off Read more Planet Fitness Mexico ### Out with ServiceDesk. Owned software. Fewer licenses. SaaS-exit program with Planet Fitness Mexico: we replaced ServiceDesk with owned multi-area ticketing, shipped PF Rooms (door tablets + Exchange) across CDMX, and we're building the next product to turn off another vertical license, with no paid seats. Learn more → Read more Rose Legal · Smokeball exit ### Out with Smokeball: owned practice management Complete practice management for Rose Legal (Australia) replacing Smokeball: clients, matters, time & billing, LPUGR-aligned trust accounting, documents, conflicts, reporting and role-based access. Learn more → Read more JCDecaux Mexico ### From paper to a secure internal platform Paperless centralizes purchasing, inventory, work orders, advances and refunds for JCDecaux Mexico in an internal platform with Single Sign-On, approvals and audit-ready traceability. Learn more → Frequently asked questions ## The questions we hear often **Is building cheaper?** Over 2–4 years of recurring licenses, almost always; and the asset is yours. Discovery delivers the numbers. **What about data migration?** Phased migration with controlled coexistence and a rollback plan. **Who maintains the software?** Slash, through coexistence and after shutdown: support is included for the first two months after delivery, then on the support plan you choose. **Corporate integrations?** SSO, Microsoft Graph/Exchange, ERP and whatever your operation demands. **Do we start with the most expensive?** We start with what hurts most: sometimes not the biggest license, but the most blocked process. **How much does it cost and how long does it take?** It depends on scope, integrations (SSO, ERP, Microsoft Graph), data migration and how long the SaaS and the new software run side by side. We quote after a free first read, and discovery compares it with what you pay in licenses today. It ships in chapters: the first one works long before the full switch-off. **Who owns the code, and how do you protect our data?** The code and IP of what we build for you are yours. Access runs through corporate SSO and roles, every action lands in exportable logs for internal control, and migration happens in phases, with rollback. We sign an NDA before any sensitive data comes in. **What technology do you build it with?** With the stack the case calls for, not the fashion, and the decision is written down: Rémy Sanchez, CTO, approves every technical decision on stack, performance and security. We integrate with what your operation already uses, such as corporate SSO, Microsoft Graph and Exchange or your ERP, as in PF Rooms for Planet Fitness Mexico. **Who works on the project, and on what hours?** A small senior core plus specialists per project, always with a named owner. We work from Bogotá, on UTC-5 all year, Monday to Friday, 9:00 to 18:00, in Spanish, English or French. If your company is in the US, we work as your nearshore partner, on your hours. **Where do you host the application and the data?** Mainly on Supabase (database, authentication and storage), OVHcloud and Microsoft Azure. We pick the provider and region based on the data residency your case requires and what your company already uses, and we handle personal data under Colombia's Law 1581 of 2012 and the GDPR in Europe. --- # Web application development | Slash Digital Lab, Bogotá URL: https://slash-digital.io/en/web-applications/ Digital product & software Platforms that operate: e-commerce, portals, event platforms and owned SaaS. Performance, accessibility and security by default. Performance Core Web VitalsAccessibility WCAG 2.2 AACases Chanel · Paperless Start a project See real work In short Slash Digital Lab builds production web applications from Bogotá, and a web app is measured on three things: speed on mobile, stability under load and traceability of what happens. We build platforms to that standard, from e-commerce to internal systems. Evidence: the Paperless platform at JCDecaux Mexico, the payments work with Chanel EMEA, the B2B and B2C wine e-commerce of UVAS Trading in Madrid and the official app of Colombia Tech Week 2026. When the web is part of a full product, we run it as a digital product, with a mobile app when needed. What we deliver ## What platforms we build We pick the stack by the case, not the fashion; the decision is documented. 01 ### E-commerce & payments Catalogue, checkout, local payment methods (Klarna, Adyen) and reconciliation. 02 ### Internal platforms SSO, approvals, dashboards and audit (see Paperless). 03 ### Event platforms Agenda, networking, chat and maps that survive event-scale peaks. 04 ### Owned SaaS Full-stack product with multi-tenancy, roles and billing. How we work ## How we ship production web - ### Architecture Stack, performance and security decisions documented before coding. - ### Build Sprints with demos; per-page performance budgets. - ### Testing Accessibility, load and edge cases on low-end phones. - ### Launch Deployment, monitoring and an incident-response plan. Scope and deliverables ## What the project delivers - **Performance:** Per-page budgets and Core Web Vitals targets. - **Security:** OWASP, headers, dependency reviews. - **Accessibility:** WCAG 2.2 AA verified. - **CMS:** Frictionless self-service content management. - **Integrations:** Payments, CRM, ERP and own APIs. - **Observability:** Logs, metrics and alerts from day one. Proof in production ## Platforms in production Read more Chanel · EMEA ### Local payment methods: a method to launch them in days With Chanel EMEA we opened the path to integrate local payment methods (such as Klarna) on Chanel.com without depending on the worldwide roadmap: an autonomy diagnosis, Hybris and Adyen analysis, and a secure proxy architecture on a shared Adyen merchant account. Learn more → Read more JCDecaux Mexico ### From paper to a secure internal platform Paperless centralizes purchasing, inventory, work orders, advances and refunds for JCDecaux Mexico in an internal platform with Single Sign-On, approvals and audit-ready traceability. Learn more → Read more UVAS Trading · Madrid ### Premium e-commerce for wines and champagnes B2B and B2C e-commerce for UVAS Trading (Madrid): wines and champagnes from 25 regions, tasting-note product pages, secure checkout and a CMS so the team can run the catalogue without friction. Learn more → Frequently asked questions ## The questions we hear often **Modernize legacy or start fresh?** Both; discovery shows where rebuilding pays and where modernization is enough. **Headless or SAP Commerce?** Yes; the proxy architecture we designed with Chanel EMEA is an example. **Can our team edit content?** Yes: a CMS for content changes without developer dependency. **How long is a portal?** Discovery to production typically runs 8–16 weeks by scope. **What about post-launch growth?** Yes: two months of support included after delivery; then a version plan, monitoring and a priority-driven backlog on the support plan you choose. **What technology do you build with?** With the stack the case calls for, and the decision is documented before coding. What doesn't change is the standard: per-page performance budgets with Core Web Vitals targets, verified WCAG 2.2 AA accessibility and OWASP security with dependency reviews. We integrate payments such as Klarna or Adyen, CRM, ERP and your own APIs. **How do you protect the platform and user data?** With security by default: OWASP, headers and dependency reviews; SSO, approvals and audit on internal platforms; logs, metrics and alerts from day one, and an incident-response plan at launch. When you need it, our own pentest practice tests the platform. **Where do you host the application and the data?** Mainly on Supabase (database, authentication and storage), OVHcloud and Microsoft Azure. We pick the provider and region based on the data residency your case requires and what your company already uses, and we handle personal data under Colombia's Law 1581 of 2012 and the GDPR in Europe. --- # Mobile app development in Colombia | Slash Digital Lab URL: https://slash-digital.io/en/mobile-applications/ Digital product & software Production apps for iOS and Android: native or cross-platform, embedded AI, store publishing and operation after launch. Stack Native · FlutterStores App Store · Google PlayCases Tech Week · Mazette Start a project See real work In short Slash Digital Lab develops iOS and Android apps from Bogotá, and a production app is more than a companion: onboarding, notifications, offline, analytics and a release cycle. We publish to the App Store and Google Play and keep the app alive with your team. Evidence: the official app of Colombia Tech Fest and the full Flutter app for Réseau Mazette (iOS and Android). For AI on the phone itself: small models on device. The rest of the product: digital products and web applications. What we deliver ## What a Slash app includes We choose the stack by the case: Flutter when it sustains the product, native when the device demands it. 01 ### Cross-platform Flutter One codebase, two stores; native-grade performance for community products. 02 ### Native when it matters Native iOS/Android for advanced capabilities: camera, sensors, payments. 03 ### Embedded AI Copilots, recommendation and moderation inside the app. 04 ### Stores & releases Publishing, versioning, review cycles and an update plan. How we work ## From idea to the stores - ### Discovery Users, use cases and store constraints, defined in writing. - ### Prototype The flow tested with real users before building everything. - ### Build & beta Sprints with TestFlight and Play Console internal testing. - ### Launch Publication, analytics, notifications and a release plan. Scope and deliverables ## What the project delivers - **Platforms:** iOS · Android · web app. - **Stack:** Flutter or native, by case. - **AI:** Copilots and embedded models. - **Offline:** Maps and content without network. - **Stores:** Publishing, listings and releases. - **Analytics:** Events and KPIs from day one. Proof in production ## Apps in the stores Read more Colombia Tech Fest ### The official tech festival app Official Colombia Tech Fest attendee app: invite-only onboarding, live agenda, directory, 1:1 meetings, chat, offline maps and notifications, built for event-scale concurrency. Learn more → Read more Réseau Mazette · Flutter app ### Full Flutter app for iOS and Android Réseau Mazette connects French-speaking professionals and entrepreneurs in Spain. We shipped a full native Flutter app for iOS and Android: directory, events, messaging, memberships and AI copilot Mia, plus the branded web experience. Learn more → Frequently asked questions ## The questions we hear often **Native or cross-platform?** By case: Flutter for community products with sustainable budgets, native for specific capabilities. **Do you publish to the stores for us?** Yes: listings, screenshots and the review process included. **PWA instead of an app?** When the case allows it (see Peoplesnap); a web app skips stores entirely. **Post-launch support?** Yes: two months of support included after delivery; after that, releases, hotfixes and a priority-driven roadmap on the support plan you choose. **Notifications and chat?** Yes: segmented push and messaging, as in the Tech Week app. **How long does store publishing take?** Leave weeks, not days. Apple says it reviews most submissions in under 48 hours, but on Google Play a personal developer account created after November 13, 2023 first needs a closed test with at least 12 testers for at least 14 days. **How do users get in, and what happens to their data?** Through controlled access. In the Colombia Tech Fest app, every account starts from an invitation code and the guest accepts the privacy policy; in Réseau Mazette, members sign in with their website credentials. If the app lets people create accounts, it must let them delete them in the app: Apple and Google Play require it. **How long does it take to build an app?** It starts with a two-week discovery that sets the scope; then we work in two-week sprints with a demo, so you see the app grow on your phone. Add store review time and, after delivery, two months of included support. --- # Data analytics & engineering in Colombia | Slash Digital Lab URL: https://slash-digital.io/en/data/ Data & automation Pipelines that survive production, solid governance and clear dashboards: the pieces that turn data into decisions, with cost in plain sight. Goal Decisions in secondsScope Pipelines → dashboardsCase 3,000+ live screens Start a project See real work In short Slash Digital Lab builds data and analytics solutions for companies, because imported data nobody queries isn't an asset: it's a cost. We build the full route (ingestion, quality, governance, dashboards) so any team decides with the right data in seconds. Observability, data contracts and cost control from the first sprint. Case in production: the urban information platform for JCDecaux and Transmilenio distributes GPS, weather and air quality across 3,000+ screens in Bogotá. The same platform feeds internal systems and owned products; the repetitive work goes to process automation. What we deliver ## What the solution includes First the decision, then the dashboard. Never the other way around. 01 ### Governance Catalogue, quality and access: who sees what and why it's trustworthy. 02 ### Integration Secure pipelines and APIs to and from your systems. 03 ### Visualisation Dashboards leadership actually opens, on mobile and desktop. 04 ### AI on data Natural-language questions with citations to the data that backs them. How we work ## How the solution is built - ### Decision map What gets decided daily or weekly, and with which data today. - ### Ingestion & quality Connected sources, data contracts and quality alerts. - ### Model & metrics Single definitions agreed across teams. - ### Adoption Mobile dashboards, training and an improvement cycle. Scope and deliverables ## What the project delivers - **Governance:** Catalogue, quality and access. - **Integration:** APIs, pipelines and batch. - **Storage:** Warehouse or lakehouse. - **Observability:** Freshness, volume and cost per source. - **Visualisation:** Mobile and desktop dashboards. - **Security:** Least privilege and auditability. - **Languages:** ES · EN · FR. Proof in production ## Data in production Read more JCDecaux × Transmilenio ### Live data across 3,000+ urban screens Urban information platform for JCDecaux and Transmilenio: GPS bus arrivals, weather and air quality across 3,000+ digital totems in Bogotá, with QR continuity to mobile. Learn more → Read more Planet Fitness Mexico ### Out with ServiceDesk. Owned software. Fewer licenses. SaaS-exit program with Planet Fitness Mexico: we replaced ServiceDesk with owned multi-area ticketing, shipped PF Rooms (door tablets + Exchange) across CDMX, and we're building the next product to turn off another vertical license, with no paid seats. Learn more → Read more JCDecaux Mexico ### From paper to a secure internal platform Paperless centralizes purchasing, inventory, work orders, advances and refunds for JCDecaux Mexico in an internal platform with Single Sign-On, approvals and audit-ready traceability. Learn more → Engineering ## Data platform and engineering A data platform is a factory: nobody watching the line means defects ship. Documented architecture, alerts that wake someone up, and cost nobody discovers on the invoice. 01 ### Ingestion & contracts Connected sources with schemas and breakage alerts. 02 ### Transformation Layered modeling with tests and documented lineage. 03 ### Observability Freshness, volume and cost visible per source and per team. 04 ### Scale An architecture built to grow without rebuilding everything. Frequently asked questions ## The questions we hear often **Do we start with a dashboard?** We start with the decision you need to make; the dashboard comes after. **Can you connect AI to our data?** Yes, with access control and source traceability. **Specific tools?** We choose per case: standard or open-source warehouse, ELT and BI; the decision is documented. **Do you migrate off Excel?** Yes: structure first, automate second, retire the critical spreadsheet last. **How long?** A first domain typically reaches production in 4–8 weeks. **How much does it cost?** It depends on how many sources need connecting and their quality, the volume and frequency of the data, and how many dashboards and users rely on it. We quote after a free first read. **Big data or standard architecture?** Most cases resolve with well-operated standard architecture; big data comes later. **Do you work with our data team?** Yes: we run it together, or you hand us the platform. **What about data compliance?** Retention, access and minimisation defined by design. **What do you need from us to start?** The decisions your team makes every day or week, the data it uses for them today and access to those sources. With each area we agree single definitions for every metric. For the free first read, two lines about your case are enough. **Who works on the project, and on what hours?** A small senior core plus specialists per project, always with a named owner, alongside your data team or in charge of the whole platform. We work from Bogotá, on UTC-5 all year, Monday to Friday, 9:00 to 18:00, in Spanish, English or French. --- # AI process automation for business | Slash Digital Lab URL: https://slash-digital.io/en/automation/ Data & automation We automate processes with clear rules and traceability: approvals, reconciliations, documentation, alerts and reports. AI where it adds judgement; none where it's noise. Rule Traceability firstAI Where it adds judgementEvidence PF Mexico · Paperless Start a project See real work In short Slash Digital Lab automates business processes with AI and software. Automation isn't headcount-cutting: it's freeing your team from the repetitive. We design flows with auditable evidence: who approved what, when, with which data. Evidence: the Planet Fitness Mexico program (owned ticketing, PF Rooms) and JCDecaux Mexico's Paperless, which digitized purchasing and inventory with SSO. With AI: Auto Fiche for Les Gentlemen du Jeu generates product sheets in PrestaShop from automatic web research. When a process needs judgement, we use AI agents; when it needs clean data, data solutions. What we deliver ## What we automate First what hurts, not what shines. Every step leaves a record. 01 ### Internal flows Approvals, purchasing, advances and work orders with explicit rules. 02 ### Documents Generation, review and filing with version control. 03 ### Process agents Classification, extraction and actions under human supervision. 04 ### Reconciliation Payments, invoices and reports that used to live in spreadsheets. How we work ## From process to flow - ### Mapping The process as it actually runs, with times and rework measured. - ### Flow design Rules, exceptions and decision points made explicit. - ### Implementation A pilot on real data with a rollback plan agreed up front. - ### Operation Savings metrics and periodic flow supervision. Scope and deliverables ## What the project delivers - **Scope:** One critical process or a multi-area program. - **Traceability:** Every action lands in an auditable record. - **AI:** Classification and extraction with human review. - **Integration:** ERP, email, SSO and APIs. - **Rollback:** Defined before the pilot starts. - **Languages:** ES · EN · FR. Proof in production ## Processes automated Read more Planet Fitness Mexico ### Out with ServiceDesk. Owned software. Fewer licenses. SaaS-exit program with Planet Fitness Mexico: we replaced ServiceDesk with owned multi-area ticketing, shipped PF Rooms (door tablets + Exchange) across CDMX, and we're building the next product to turn off another vertical license, with no paid seats. Learn more → Read more JCDecaux Mexico ### From paper to a secure internal platform Paperless centralizes purchasing, inventory, work orders, advances and refunds for JCDecaux Mexico in an internal platform with Single Sign-On, approvals and audit-ready traceability. Learn more → Read more Les Gentlemen du Jeu · Auto Fiche ### AI product sheets for board games Auto Fiche PrestaShop module for Les Gentlemen du Jeu: generates board-game product sheets with AI (SEO, players, duration, age, FAQ and images) from automatic web research. Learn more → Frequently asked questions ## The questions we hear often **Do you use n8n or Make?** Where it fits; when the process is critical we build an owned service with audit logs. **What if the process changes monthly?** We stabilize it first: automating chaos only accelerates it. **RPA or AI agents?** Rules for stable steps; AI where there's variation, always supervised. **How long is a pilot?** Weeks; mapping fixes the exact scope before coding. **Do you automate finance?** Yes: approvals, reconciliations and recurring reports are typical cases. **How much does it cost to automate a process?** It depends on scope, one critical process or a multi-area program, and on integrations with your ERP, email, SSO and APIs; mapping fixes that scope before coding. We quote after a free first read and measure savings against a baseline taken before the pilot. **What do you need to start?** A stable process we can map as it actually runs, with its times and rework measured. If it changes every month, we stabilize it first, because automating chaos only accelerates it. For the free first read, two lines on the process that hurts your team most are enough. **How do you keep control of what the automation does?** Every action lands in an auditable record: who approved what, when and with which data. AI classifies and extracts under human review, rollback is defined before the pilot, and we supervise the flow periodically with savings metrics. --- # DOOH solutions & urban digital screens | Slash Digital Lab URL: https://slash-digital.io/en/dooh/ DOOH & experiences The digital layer of street screens: real-time urban information, wayfinding, interaction and physical-digital campaigns for brands and cities. 3,000+ urban screens in production10+ countries deployedCase JCDecaux × Transmilenio Start a project See real work In short Slash Digital Lab builds DOOH (digital out-of-home) solutions for brands and cities. A street screen that only loops ads is an expensive poster. We turn it into useful information (bus arrivals, weather, air quality, wayfinding) and into a brand channel people actually use. Evidence: the information platform for JCDecaux and Transmilenio in Bogotá, the 3D wayfinding at Cardales de Cayalá (Guatemala) and activations like Cartier Trinity across three countries. To verify that content reached every screen: computer vision. What we deliver ## What the DOOH layer includes Data, content and interaction on one platform, with the operation an urban network demands. 01 ### Live data Urban APIs (transit, weather, air) served to hundreds of screens at low latency. 02 ### Wayfinding 3D maps, step-by-step routes and QR continuity to mobile. 03 ### Physical-digital campaigns From QR to jumbotron: UGC, mosaics and measurement. 04 ### Network operation Content CMS, scheduling and per-totem monitoring. How we work ## How a DOOH network gets built - ### Inventory Screens, locations, connectivity and owners. - ### Platform Data service, player and CMS with per-location SLAs. - ### Content Dynamic, brand-safe templates powered by live data. - ### Operation Monitoring, reporting and network evolution. Scope and deliverables ## What the project delivers - **Network:** Documented inventory and connectivity. - **Platform:** Data services, player and CMS. - **Content:** Dynamic templates powered by live data. - **Interaction:** QR, touch and camera with consent. - **Measurement:** Monitoring and per-location reporting. - **Languages:** ES · EN · FR. Proof in production ## DOOH in production Read more JCDecaux × Transmilenio ### Live data across 3,000+ urban screens Urban information platform for JCDecaux and Transmilenio: GPS bus arrivals, weather and air quality across 3,000+ digital totems in Bogotá, with QR continuity to mobile. Learn more → Read more JCDecaux Guatemala · Cardales de Cayalá ### Smart wayfinding for premium retail Interactive wayfinding for Cardales de Cayalá: 3D mall map, on-screen step-by-step routes, itinerary handoff to mobile and an updatable venue directory. Learn more → Read more Cartier · Trinity 100 ### A Trinity century as a physical-digital experience Physical-digital activation for Cartier's Trinity ring 100th anniversary: branded photo booth, three symbolic captures, on-site print and first-party data into the CRM. The Trinity photobooth was such a success that we reproduced the experience in Bogotá, Panama and São Paulo. Learn more → Read more Peoplesnap ### From smartphone to DOOH mosaic, without a single booth Peoplesnap is Slash Experience's photobooth product: a white-label web app that turns each guest's phone into the activation. QR, selfie, AI editing (100+ parameters), social share, print and mosaics on event screens. Already live for Mouton Cadet, La French Tech Bordeaux, CEA Marcoule, Denver Pub Crawl, With Agency, We Love Beer and CES Las Vegas 2024. Learn more → Frequently asked questions ## The questions we hear often **Are you a screen vendor?** No: we build the digital layer that makes screens useful for brands and cities. **Does it work for events?** Yes: live mosaics, participation and moderation (see Peoplesnap). **Do you manage ad content?** We integrate the CMS with your scheduling and media provider. **How many screens can the platform take?** The reference network exceeds 3,000 and is built to grow. **How is impact measured?** Availability, interaction and reach per location, with regular reporting. **What do you need to start a DOOH network?** Your network inventory: screens, locations, connectivity and owners, which we document at the start. We're not a screen vendor; on that network we set up the data service, player and CMS, with per-location SLAs, then the content templates powered by live data. **What technology runs behind the screens?** A platform with a data service, player and CMS: urban transit, weather and air-quality APIs served at low latency, dynamic templates, 3D maps with QR handoff to mobile and per-totem monitoring. The reference network in Bogotá exceeds 3,000 screens. **Do you use cameras in front of the screens?** Only with consent: QR, touch and camera interaction are designed that way. When computer vision is involved, we count without recognizing anyone and apply Law 1581 and the SIC's video-surveillance guide in Colombia, with a visible notice, a defined purpose and short retention. --- # AI photobooths & brand activations | Slash Digital Lab URL: https://slash-digital.io/en/interactive-experiences/ Phygital We fuse physical and digital spaces: boothless web photobooth, brand mosaics on giant screens, navigable WebGL and activations that leave useful data. 0 hardware to transportGenAI Edits in secondsCases Cartier · L'Oréal · Peoplesnap Start a project See real work In short Slash Digital Lab creates brand activations and interactive experiences with AI (web photobooth, mosaics, WebGL). A great interactive experience does two things: moves people in the moment and leaves useful traces (a shared photo, first-party data, the event's big screen). We design and run both sides. Evidence: Cartier Trinity in Bogotá, Panama and São Paulo; the 3D mosaic of Guardians for L'Oréal; and Peoplesnap, our own product presented at CES Las Vegas 2024. How we run generative AI live: generative AI in brand activations. What we deliver ## What experiences we build From QR to mosaic: participation with moderation, measurement and first-party data. 01 ### Boothless photobooth QR → selfie → AI edit → share or print. No logistics, no booth. 02 ### Mosaics & murals Hundreds of photos forming the brand on screen or in 3D web. 03 ### Retail activations Virtual try-on, souvenirs and brand play in store. 04 ### Events & stadiums Live participation with moderation, screens and measurement. How we work ## How an activation comes together - ### Concept Moment, brand and call to action in one measurable sentence. - ### Product A white-label web app live in hours, not weeks. - ### Event On-site operation, moderation and the on-screen mosaic. - ### Data First-party export to the CRM and a reach report. Scope and deliverables ## What the activation includes - **Format:** White-label web app. - **AI editing:** Backgrounds, relighting, avatars, face swap, GIF. - **Outputs:** Social, print, DOOH mosaic. - **Data:** Forms, challenges, sweepstakes and CSV export. - **Moderation:** Live, before any screen or feed. - **Languages:** ES · EN · FR. Proof in production ## Activations in production Read more Cartier · Trinity 100 ### A Trinity century as a physical-digital experience Physical-digital activation for Cartier's Trinity ring 100th anniversary: branded photo booth, three symbolic captures, on-site print and first-party data into the CRM. The Trinity photobooth was such a success that we reproduced the experience in Bogotá, Panama and São Paulo. Learn more → Read more L'Oréal · Guardians ### From GenAI selfie to a 3D brand mosaic Guardians program for L'Oréal: GenAI face-swap in Celebrate Consumer Care (pick a Guardian, selfie → brand avatar) and, as the centerpiece, a gigapixel Three.js mosaic where hundreds of avatars form the logo, navigable in 3D for demos and lobbies. Learn more → Read more Peoplesnap ### From smartphone to DOOH mosaic, without a single booth Peoplesnap is Slash Experience's photobooth product: a white-label web app that turns each guest's phone into the activation. QR, selfie, AI editing (100+ parameters), social share, print and mosaics on event screens. Already live for Mouton Cadet, La French Tech Bordeaux, CEA Marcoule, Denver Pub Crawl, With Agency, We Love Beer and CES Las Vegas 2024. Learn more → Frequently asked questions ## The questions we hear often **Do we need our own screens?** No: the web app runs on guests' phones and the venue's screens. **How fast is it ready?** Setup takes hours: design, copy, QR and operation. **Do you handle moderation?** Yes: live review before content reaches a screen or a feed. **Is it GDPR compliant?** Yes, end to end, with explicit consent and controlled export. **Can we print on site?** Yes: on-site print where the venue allows it (see Cartier Trinity). --- # Cybersecurity services in Colombia | Slash Digital Lab URL: https://slash-digital.io/en/cybersecurity/ Pentest and cybersecurity We assess, prioritize and help fix. Manual and automated pentest of web apps and APIs with a launch offer: no high or critical findings, COP $0. Guarantee COP $0 with no high findingsScope Web · APIsRetest 1 in paid assessments Start a project See the launch offer In short Slash Digital Lab provides cybersecurity services for companies from Bogotá, and honest security starts with actually looking: manual and automated testing on your real surface, with validated impact and clear priorities. If we find no high or critical vulnerabilities, the assessment costs nothing. The full launch offer (COP pricing, scope, retest) lives on the Pentest Colombia page. For the new risks: deepfake fraud, prompt injection in AI agents and your AI supply chain. What we deliver ## What the practice includes From test to report, from patch to retest: security that ends in a fix. 01 ### Web & API pentest Manual plus automated, with reproducible evidence for every finding. 02 ### Hardening Configuration, dependencies, secrets and exposed surface. 03 ### Incident response Containment, analysis and reporting when something blows up. 04 ### Human factor Controlled phishing and social engineering, from COP $5M. How we work ## How an assessment runs - ### Scope Domains, apps and APIs in a signed document; nothing outside contract. - ### Assessment Manual plus automated; every finding validated and reproducible. - ### Report Executive and technical: evidence, impact, priority. - ### Retest Verification after patching. Included in paid assessments. 471.5Mattack attempts per year in Colombia (CRC/FortiGuard) 1.3Mattempts per day COP $0if no high findings 1retest included in paid assessments Scope and deliverables ## The assessment, in writing - **Scope:** Web apps and APIs, in a signed document. - **Method:** Manual plus automated, hand-validated. - **Severity:** CVSS and business context. - **Report:** Executive + technical with evidence. - **Retest:** 1 included in paid assessments. - **Extras:** Controlled phishing from COP $5M. Frequently asked questions ## The questions we hear often **What if you find nothing?** You keep the full report with medium, low and informational findings, at no cost. **How is the price calculated?** COP $0 with no high findings; COP $5M for one; capped at COP $9M. Details on the Pentest page. **Do you test APIs?** Yes, together with the web: that's where the expensive doors open today. **Do you run simulated phishing?** Yes, from COP $5M, with a report and team training. **Do you help remediate?** Yes: we can prepare an independent remediation proposal. **How do you decide a finding is high or critical?** With CVSS adjusted to your business context, and only if the finding is validated by hand and reproducible within the signed scope. Every high or critical finding ships with the steps to reproduce it, so your team can check it independently before it counts toward the price. **Who do we talk to about a pentest, and in which language?** Esteban Zuleta, co-founder: he takes your request, does the first read and confirms scope and pricing in COP. The service is delivered in Spanish, English and French, and the output is an executive and technical report with evidence, impact and priority. **Do you assess mobile apps or other systems?** The launch offer covers web apps and APIs; other scopes are quoted separately. What's in goes into a signed document listing domains, apps and APIs, and nothing outside the contract gets tested. For the human factor, controlled phishing and social-engineering tests start at COP $5 million. --- # Services: AI, digital product, data, DOOH and pentest URL: https://slash-digital.io/en/services/ Services Five service families and one standard: strategy, design and engineering until it works in production. 5 service families18 servicesES · EN · FR working languages Start a project See real work 01 ## Artificial intelligence AI ### Artificial intelligence Applied AI for real operations: agents, RAG, vision, voice and automation. Learn more→ AI Lab ### Slash AI Lab Our lab where we evaluate models and publish the method. Learn more→ Agents ### AI agents Agents with tools, memory and human oversight: support, ops and back office. Learn more→ Chatbots ### AI chatbots Trilingual assistants (ES/EN/FR) connected to private data, CRM and processes. Learn more→ Consulting ### AI consulting Diagnosis and architecture before building. Actionable roadmaps. Learn more→ Training ### AI training We train your teams to use AI with judgement, not fear. Learn more→ Fine-tuning ### Fine-tuning & RAG When to tune the model and when to give it knowledge. Judgement, not hype. Learn more→ Vision ### Computer vision Computer vision for retail, DOOH, CCTV analytics and physical-digital experiences. Learn more→ 02 ## Digital product & software Product ### Digital products From idea to production: product, design and engineering in one team. Learn more→ Software ### Custom software We replace SaaS licenses with owned products your team controls. Learn more→ Web ### Web applications Production web platforms: fast, secure and built to scale. Learn more→ Mobile ### Mobile apps Production iOS/Android apps: native or cross-platform, AI included. Learn more→ Spotlight ### Networking apps For events and communities that run on connecting people: agenda, directory, 1:1 meetings, chat and memberships, on iOS, Android and web. - Agenda - Directory - 1:1 meetings - Chat - Memberships - AI copilot Let's talk about your app→ - Official app · Bogotá Colombia Tech Week 2026 Agenda, directory, 1:1 meetings and chat for thousands of attendees. See the case → - Flutter app · iOS & Android Réseau Mazette Directory, events, messaging, memberships and the Mia AI copilot for French-speaking professionals in Spain. 03 ## Data & automation Data ### Data solutions From raw data to decisions in seconds: governance, integration, visualisation. Learn more→ Auto ### Automation Repetitive processes, out. Robust flows with traceability, in. Learn more→ 04 ## DOOH & experiences DOOH ### DOOH & urban screens The digital layer of street screens: live data, DOOH and interaction. Learn more→ Phygital ### Interactive experiences Physical + digital: boothless photobooth, brand mosaics and WebGL moments. Learn more→ 05 ## Cybersecurity Cyber ### Cybersecurity Offensive assessment, hardening and response: real security, no smoke. Learn more→ Pentest ### Pentest Colombia Web & API pentest with guaranteed results: no high findings, COP $0. Learn more→ --- # Nearshore software development in Colombia for US companies URL: https://slash-digital.io/en/nearshore-software-development/ For US companies A Bogotá team that works your business day: custom software, AI and data, shipped to production for brands like Cartier, Chanel and JCDecaux. Hire a project, a dedicated team or individual engineers. UTC-5 all yearCartier USA US clientUSD invoicing · US law Start a project See real work In short Slash Digital Lab is a software and AI company in Bogotá, Colombia, that US companies hire as a nearshore partner. Bogotá runs on UTC-5 all year: the same clock as New York in winter and one hour behind it in summer, so calls, reviews and releases happen in your business day. US clients include Cartier USA. We build custom software, AI and data platforms in English, invoice in US dollars and sign under US law when you ask. What we deliver ## Four ways to work with us The same standard in each: English, your business day and code you own. 01 ### Fixed-scope project A defined build with a written quote after a free first read: scope, milestones and acceptance criteria up front. 02 ### Dedicated team A team that works only on your product, billed monthly, on your priorities and your backlog. 03 ### Individual engineers Engineers who join your own team, tools and rituals, for as long as you need them. 04 ### Support after launch Maintenance, operation and improvements on a retainer once the product is live. How we work ## How an engagement starts - ### Free first read A call about your goal, constraints and timeline, with an honest view of whether we're the right team. - ### Proposal in USD The engagement model, the scope or team profile, the timeline and the price, in US dollars. - ### Contract and NDA Under US law if you prefer, with an NDA when the project requires it. The code and IP are yours. - ### Delivery on your hours Short iterations with demos in your business day, in English. Scope and deliverables ## The practical details - **Time zone:** Bogotá, UTC-5 all year: New York time in winter, one hour behind it in summer. - **Languages:** English, Spanish and French. - **Engagement:** Fixed-scope project, dedicated team, individual engineers or support after launch. - **Invoicing:** In US dollars. - **Contracts:** Under US law on request; NDA when the project requires it. - **Ownership:** The code and IP of everything we build for you belong to you. - **Security:** Testing by our own pentest practice when you need it. - **Company:** Slash Digital Lab SAS, Carrera 8A 97-07, Bogotá, Colombia. Founded in 2022. Proof in production ## Work for international brands Read more Cartier · Trinity 100 ### A Trinity century as a physical-digital experience Physical-digital activation for Cartier's Trinity ring 100th anniversary: branded photo booth, three symbolic captures, on-site print and first-party data into the CRM. The Trinity photobooth was such a success that we reproduced the experience in Bogotá, Panama and São Paulo. Learn more → Read more Rose Legal · Smokeball exit ### Out with Smokeball: owned practice management Complete practice management for Rose Legal (Australia) replacing Smokeball: clients, matters, time & billing, LPUGR-aligned trust accounting, documents, conflicts, reporting and role-based access. Learn more → Read more Chanel · EMEA ### Local payment methods: a method to launch them in days With Chanel EMEA we opened the path to integrate local payment methods (such as Klarna) on Chanel.com without depending on the worldwide roadmap: an autonomy diagnosis, Hybris and Adyen analysis, and a secure proxy architecture on a shared Adyen merchant account. Learn more → Frequently asked questions ## The questions we hear often **Do you work with US companies?** Yes. US clients include Cartier USA, and our production work spans brands like Chanel, Cartier, JCDecaux and L'Oréal. **What hours do you work?** Bogotá is UTC-5 all year: New York time in winter and one hour behind it in summer. Our 9:00 to 18:00 day overlaps at least seven hours with the East Coast and six with the West Coast. **How can we hire you?** Four ways: a fixed-scope project, a dedicated team billed monthly, individual engineers who join your team, or support after launch on a retainer. **Do you invoice in US dollars?** Yes. US clients are invoiced in USD. **Can the contract be under US law?** Yes, when you ask for it. We also sign an NDA when the project requires it. **Who owns the code?** You do. The code and IP of everything we build for you belong to you. **How much does it cost?** It depends on the engagement model, the scope or team size, and how long you need us. We quote in USD after a free first read. **Why Colombia?** Your hours, about four hours' flight from Miami, and a tech scene we help build: we were a Gold Sponsor of Colombia Tech Week 2026 and built its official app. --- # Web & API pentesting in Colombia | Slash Digital Lab URL: https://slash-digital.io/en/pentest/ Launch offer · Colombia At Slash Digital Lab (Bogotá, Colombia), the price of your pentest depends only on the high or critical vulnerabilities we find and validate within the agreed scope. COP $0 no high findingsCOP $5M 1 high findingCOP $9M cap: 2 or more Start a project Read the pentest analysis ## Launch offer: price by findings No high findingsCOP $0 If we find no confirmed high or critical vulnerabilities. 1 high findingCOP $5 M If we identify 1 confirmed high or critical vulnerability. 2+ high findingsCOP $9 M Maximum, however many high or critical vulnerabilities we find. The assessment runs in full regardless of the number of findings. ### The best moment to find a vulnerability is before it becomes an incident. We run offensive assessments of web apps and APIs to find exploitable flaws, validate their real impact and prioritize what to fix first. We don't just hand over scanner output. ### What the assessment includes A real security assessment of web apps and APIs, with manual validation of the relevant findings. - Web & API pentestManual plus automated testing to find what scanners miss. - Executive & technical reportEvidence, impact, priority and remediation guidance. - 1 retest included in paid assessmentsOnce fixes land, we verify again. - Retest holds regardless of who fixesValid whether your team, another vendor or Slash ships the fix. ### No high vulnerability found? You still get the report. At no cost: medium, low and informational findings, with evidence and remediation guidance. ### We can also help you fix it. On request, Slash prepares an independent proposal to implement the recommended fixes. ### And the human factor? Controlled phishing and social-engineering tests from COP $5 million. ## Cyberattacks in Colombia: the context Colombia · 2025 471.5 M attack attempts detected in the country during the year 1.3 Mper day ### Did you know Colombia saw nearly 1.3 million attack attempts a day? That equals 471.5 million attack attempts over the year. A vulnerability can stay invisible until someone exploits it. Source: CRC / FortiGuard Labs, prospective cybersecurity analysis, 2026. “Attack attempts” does not equal successful attacks. Severity ## How “high” and “critical” are decided The price depends on severity, so the rule has to be verifiable. Every finding is manually validated and must be reproducible; severity uses CVSS plus business context. - ### Reproducible evidence Every high or critical finding ships with the steps to reproduce it, so your team can check it independently. - ### CVSS + business context The industry-standard score, adjusted to the real impact on your operations and data. - ### Within the authorized scope Anything that can't be reproduced within the scope agreed in the contract isn't reported as high or critical. Frequently asked questions ## The questions we hear often **What if you find nothing?** You keep the full report with medium, low and informational findings, at no cost. **How is the price calculated?** COP $0 with no high findings; COP $5M for one; capped at COP $9M. **Do you test APIs?** Yes, together with the web: that's where the expensive doors open today. **Do you run simulated phishing?** Yes, from COP $5M, with a report and team training. **Who fixes the findings?** Whoever you choose: your team, another vendor or Slash; the retest still applies. **How do we start?** Send us the domain or platform; we do a first read and confirm scope in the contract. **Do you sign an NDA?** Yes: scope and confidentiality are signed before testing starts. We only assess the domains, apps and APIs in the signed document, and the offer is subject to that scope definition and the contract. Every finding comes with reproducible evidence in an executive and technical report. **What should we hand over for the test?** For the first read, the domain or platform. For the test, unless you have a reason to do otherwise, we recommend grey box: two accounts per role (and per customer if the platform is multi-tenant), the API documentation and the business context of each critical flow, such as payments, approvals or personal data. --- # Cartier Trinity 100: a phygital photobooth in 3 cities URL: https://slash-digital.io/en/case-studies/cartier/ Cartier · Trinity 100 A branded photo booth in premium malls: three photos (three Trinity bands), on-site print and social sharing, all feeding Cartier's CRM. Start a project Watch the video At the Trinity 100 booth in a mall, guests tap the screen, scan a QR code and pose. In short For the 100th anniversary of Cartier's Trinity ring, we built a physical-digital activation: a branded photo booth in premium malls. Guests posed for three photos, one for each Trinity band, received a print on the spot and shared the result on social media, while every capture fed first-party data into Cartier's CRM. It worked so well that we reproduced it in Bogotá, Panama and São Paulo. Client Cartier · Trinity 100 Sector Luxury retail Where Bogotá · Panama · São Paulo Year 2025 Deliverables - Branded photo booth - Three-photo Trinity flow - On-site printing - One-click social sharing - First-party data to CRM Services - Interactive experiences - DOOH & urban screens 3countries: Bogotá · Panama · São Paulo 3captures per guest (three bands) CRMfirst-party data on every shot 01 The context ## A ring's centenary, celebrated in person Cartier's Trinity ring turned 100, and Cartier wanted to celebrate it in Latin America. The brief had two sides. The experience had to honor the Trinity legacy with real emotion, bringing the physical and the digital together. It also had to feed Cartier's CRM with first-party data. The venue was the premium mall, so every visit to the booth had to leave three things behind: a memory, a photo to share and useful data. 02 Design and build ## From the touchscreen to the print We designed the experience around a branded Trinity 100 booth, with a curtain and a host to welcome guests. Each session started on the touchscreen and continued inside the booth, which took three photos, one for each Trinity band, on three gold tones: the Trinity aesthetic applied to every picture. Before leaving, guests saw the result on screen and could retake it. The result took two forms. On the spot, the booth printed a tall strip with the three photos and the Trinity 100 signature. On screen, guests could scan a QR code with their phone or share the photo on social media in one click. Every capture also fed first-party data into Cartier's CRM, so the celebration and the data came from the same moment. - 01 ### Three photos, three bands Each session produced three photos, one for each band of the Trinity ring, on three gold tones. The ring's design became the structure of the experience. - 02 ### A print on the spot The booth printed the photos on site, with the Trinity 100 signature, so guests left with a physical keepsake of the anniversary. - 03 ### Share in one click From the booth screen, guests shared their photo on social media in a single click, which brought organic UGC to the anniversary. - 04 ### First-party data, every shot Every capture fed first-party data into Cartier's CRM, so the celebration also left useful information for the brand. 03 Gallery ## The project in pictures 3 images · Open any image to see it full size. Booth screen: the guest's photos in a gold frame, the Trinity 100 print, share and retake buttons. The printed strip: three photos on three gold tones, one for each Trinity band, signed Trinity 100. Three screens side by side: one-click social sharing, the three-photo strip and the final print. 04 Outcomes ## One booth, three markets After the first activation's success, we reproduced the experience in Bogotá, Panama and São Paulo. Across these three key markets, it brought high engagement, organic UGC and useful data for Cartier's CRM. Cartier came away with a format proven in three countries, from the booth and the three-photo flow to the link with its CRM. - Experience reproduced in Bogotá, Panama and São Paulo - High engagement at the booth across three key markets - Organic UGC from one-click social sharing - Useful first-party data flowing into Cartier's CRM --- # Chanel EMEA: a method to launch local payments in days URL: https://slash-digital.io/en/case-studies/chanel/ Chanel · EMEA Europe was fully deployed; countries needed local adaptation and global teams faced 3.5× more demand than they could deliver. Start a project See the gallery Chanel.com with a fine jewelry campaign: the storefront where EMEA's local payment methods had to fit. In short With Chanel EMEA, we delivered Autonomy & Agility, a program that opened the way to add local payment methods such as Klarna to Chanel.com without depending on the worldwide roadmap. Over 6 sprints and 18 weeks, we diagnosed EMEA's autonomy, analyzed Hybris and Adyen, and designed a secure proxy on a shared Adyen merchant account. The result: a clear path to launch a payment method in days. Client Chanel · EMEA Sector Luxury e-commerce Where EMEA Year 2025 Deliverables - EMEA autonomy diagnosis - Hybris and Adyen analysis - Agile integration guide - Payment proxy architecture - Klarna technical case Services - Web applications - Digital products 01 · Sprints6 sprints · 18 weeks 02 · Team12 people · 800 hours 03 · GoalA new method in days 01 The context ## Local payments stuck in a global queue The European deployment was complete, and each country now asked for local adaptation: one payment method per market, such as Klarna. The platform ran on a technical framework under global control, Hybris (SAP Commerce) with Adyen for payments, and the global teams faced 3.5× more demand than they could deliver. The central roadmap was built for the lowest common denominator. A method that mattered to one country rarely made it to the top of the list, so each Klarna or country-specific method waited years. Chanel EMEA needed a way to move on its own, without touching sensitive configuration and with risk kept under control. 02 Design and build ## A proxy ahead of any replatform Autonomy & Agility ran in 6 three-week sprints over 18 weeks, with 12 people, 800 hours and a flash report every week. We started with the diagnosis: we mapped the technical framework under global control and showed why country-by-country payment methods fell off the central roadmap. Then we analyzed Hybris and Adyen, ran a market study, prepared the technical case for connecting Klarna or any enabled method, and wrote an agile guide for connecting new payment methods. At the center sits a proxy that mirrors a Klarna-style flow on a shared Adyen merchant account, so a new method follows the pattern Chanel.com already uses. The shopper clicks pay, the merchant shows the payment form and the shopper submits the details. The merchant server starts the authorization, the proxy sends a new authorization request to Adyen, the response returns to the merchant server and the result page appears. The shopper never sees the proxy. Risk profiles, reconciliation and support are covered, with no access to Chanel's sensitive data. - 01 Delivered ### EMEA autonomy diagnosis Mapping of the technical framework under global control, operational gaps (15% of potential versus 80% complexity) and why country-by-country payment methods fell off the central roadmap. - 02 Delivered ### Klarna and beyond, with Adyen Market study, ideation and technical case: how to connect a new method (Klarna or any enabled one) by mirroring Chanel.com's pattern on a shared Adyen merchant account, with risk profiles, reconciliation and support, without touching sensitive configuration. - 03 Delivered ### Transparent proxy and agile method A proxy architecture invisible to the shopper: connects payments (and shipping) ahead of major platform shifts (headless / SAP Commerce Cloud). Work in 3-week sprints with weekly flash reports. 03 Gallery ## The project in pictures 5 images · Open any image to see it full size. Authorization sequence: the merchant server starts the payment, the proxy requests a new authorization from Adyen. Context: the Chanel.com checkout in French, at the payment step where methods are chosen. Context: the Chanel.com payment methods block, with Visa, Amex, Mastercard, CB, PayPal and Apple Pay. Context: the Chanel.com home page, with the Métiers d'Art show and a fragrance campaign. Context: a fragrance and beauty gifting campaign on Chanel.com, where the purchase journey begins. 04 Outcomes ## A path measured in days Chanel EMEA now has a way to add a local payment method in days rather than years, outside the queue of the worldwide roadmap. It keeps Adyen's security and follows a pattern ready to replicate country by country. The same logic supports post-purchase autonomy for shipping and carriers, ahead of major platform shifts such as headless or SAP Commerce Cloud. - A repeatable way to add a local method in days - Adyen's security, with no access to Chanel's sensitive data - A pattern ready to replicate country by country - The same logic reusable for shipping and carriers --- # Club América × Coleccioname: the official digital album URL: https://slash-digital.io/en/case-studies/club-america/ Coleccioname · Club América × Fanki Coleccioname is Club América's (Mexico) Panini-style digital sticker album with Fanki: buy packs, place holographic stickers, trade duplicates with other fans, track progress and complete the official collection from any device. Start a project Watch the video Coleccioname promo: the album, pack opening, checkout from any device and the web album's four tabs. In short Coleccioname is Club América's official digital sticker album, a web platform we built with Fanki. Fans buy packs, place holographic stickers, trade duplicates with other fans and track their progress from any device, in Spanish or English. The sticker-album ritual now lives online, with no paper and no queues. Client Coleccioname · Club América Sector Sports and collectibles Where Mexico Deliverables - Paginated digital album - Packs, store and checkout - Fan-to-fan trade marketplace - Progress bar and leaderboard - Bilingual ES/EN interface Services - Web applications - Digital products 01 · AlbumOfficial stickers and pages 02 · MarketPacks, store and trades 03 · FansLive progress and ranking 01 The context ## The sticker-album ritual, moved online The goal was to move the excitement of a sticker album online without losing its rituals. Fans still had to open packs, fill pages and trade with the community, and Club América's brand had to stay present at every step, on mobile and on desktop. In digital form, the album also had to hold the store, the trades between fans and each collector's progress in one place. 02 Design and build ## Four tabs, one collection With Fanki, we built a full web platform organized around four tabs: My album, My packs, Trades and Offers. Fans buy packs in the store and pay at checkout, open them to reveal holographic stickers, and place each sticker in the paginated album. Duplicates go to a fan-to-fan marketplace, where fans propose trades and make offers to each other. To motivate collectors, a progress bar shows how much of the album is already complete, and a leaderboard adds the social push to finish the official collection. The whole experience works in Spanish and in English and wears the club's blue and yellow, on mobile as well as on desktop. - 01 Live ### Official digital album Paginated Club América album: holographic stickers, slots to fill, smooth navigation and brand look (blue / yellow) in Spanish and English. - 02 Live ### Packs, store and trades Buy packs, open stickers, fan-to-fan trade marketplace and offers, plus checkout to keep collecting. - 03 Live ### Progress and ranking Album progress bar, leaderboard and social motivation to complete América's official collection. 03 Gallery ## The project in pictures 4 images · Open any image to see it full size. Promo frame: the new official Club América album and its cover, in the club's blue and yellow. Propose trades: a holographic player sticker over the album pages, ready to offer to another fan. From any device: the sticker checkout on a phone, set against a stadium in the club's colors. The web album: two pages of sticker slots, page arrows, the four tabs and the progress bar. 04 Outcomes ## A live collecting community around América Club América now has an official album that lives online, with a live collecting community around it. Fans open packs, propose trades and complete the official album. Album, market and ranking run on one platform, on mobile and desktop, in the club's colors and in two languages. - Official Club América album live on mobile and desktop - Fans trade duplicates and make offers to each other - Packs, store and checkout inside the same platform - Collecting with no paper and no queues --- # Colombia Tech Week 2026: Gold Sponsor and official app URL: https://slash-digital.io/en/case-studies/colombia-tech-week-2026/ Colombia Tech Week 2026 We bet on the talent redefining tech in Latin America: we were a 2026 Gold Sponsor and built the official app (agenda, directory, 1:1 meetings and chat). Start a project All case studies Colombia Tech Week 2026 logo: the Bogotá week we backed as a Gold Sponsor. In short We were a Gold Sponsor of Colombia Tech Week 2026 in Bogotá and built the official app of the week. It gave thousands of attendees an agenda, a directory, 1:1 meetings and chat in one place. Our team spent the whole week on site, meeting founders, companies and the ecosystem. Client Colombia Tech Week Sector Tech events Where Bogotá Year 2026 Deliverables - Gold sponsorship - Official event app - Agenda and directory - 1:1 meetings and chat Services - Mobile apps - Digital products Gold Sponsor2026 Aug 10–16Bogotá 2026 Official appnetworking included 01 The context ## A week to back Latin American talent Colombia Tech Week 2026 took place in Bogotá from 10 to 16 August and gathered thousands of attendees. We chose to back it as a 2026 Gold Sponsor, a bet on the talent redefining tech in Latin America. The week also needed one place where attendees could follow the agenda, find each other and set up meetings. Building it became our job. 02 Design and build ## From a search to a 1:1 meeting We built the official app of the week around one tab bar: home, map, agenda, networking, chat and a personal calendar. In networking, attendees searched the directory and opened profiles showing role, industry, company and meeting availability in 30-minute blocks. Buttons let them request a meeting or block a user, and the 1:1 chat let two people agree to meet during Colombia Tech Week. The app in use: search an attendee, open their profile and 30-minute slots, then chat 1:1. - 01 ### Networking by profile Attendees found each other by search and saw role, industry, company and availability in 30-minute blocks before asking for a meeting. - 02 ### 1:1 meetings and chat A meeting request button and a 1:1 chat helped attendees agree on a time to meet during the week, and a block button kept them in control. - 03 ### The week in one app Home, map, agenda, networking, chat and a personal calendar shared one tab bar, so attendees followed the week from their phones. - 04 ### Gold Sponsor 2026 We backed the week as a 2026 Gold Sponsor, alongside the other sponsors on the official list, public at colombiatechweek.co/patrocinadores. 03 Outcomes ## A week spent with the ecosystem The official app accompanied attendees for the whole week, with the agenda, the directory, 1:1 meetings and chat in one place. Meanwhile, our team was in Bogotá from the first day to the last, meeting founders, companies and the ecosystem in person. Our Gold sponsorship is on the public 2026 sponsor list. - Official app of the week for thousands of attendees - Gold Sponsor on the public 2026 sponsor list - Networking built on 30-minute meeting slots - A full week in Bogotá with founders and companies See the official sponsors ↗ --- # Tilal Al Ghaf at MIPIM Cannes: stand landing and contact URL: https://slash-digital.io/en/case-studies/tilal-al-ghaf/ Tilal Al Ghaf · MIPIM Cannes 2025 Landing for Tilal Al Ghaf's (Majid Al Futtaim) stand at MIPIM Cannes 2025: present the destination's latest Dubai projects and make it easy to contact people visiting the stand. Start a project See the gallery The Tilal Al Ghaf logo and the MIPIM Cannes 2025 line over villas facing the lagoon. In short For MIPIM Cannes 2025, we built the landing page for the Tilal Al Ghaf (Majid Al Futtaim) stand. It presented the destination's latest projects in Dubai, from BO MONDE to Alaya Beach, and gave investors and partners a direct way to reach the team, on tablet or mobile. The team could walk visitors through the portfolio at the stand and turn visits into qualified conversations. Client Tilal Al Ghaf · MIPIM Sector Real estate · Events Where Cannes · Dubai Year 2025 Deliverables - Stand landing page - Project showcase - Contact form and buttons - Tablet and mobile layout Services - Web applications - Digital products 01 · EventMIPIM Cannes 2025 02 · DestinationTilal Al Ghaf · Dubai 03 · DeliveryLanding + contact 01 The context ## Dubai, presented from a stand in Cannes At MIPIM Cannes 2025, the Tilal Al Ghaf stand in the Palais des Festivals received investors and partners in a high-intensity flow. The team had the destination's recent Dubai launches to present, including BO MONDE, Ghaf Woods, Lanai Islands, Amara and Alaya Beach, without breaking the face-to-face conversation. The stand needed one clear digital piece that did two jobs: present the portfolio on the spot, and capture each visitor's interest without friction, on tablet or mobile. 02 Design and build ## A landing page paced for the stand We designed a custom landing page that reads in three steps. First, an event header sets the scene: Tilal Al Ghaf, Dubai, MIPIM Cannes 2025. Next, a showcase gathers the recent launches and neighborhoods so a visitor grasps the portfolio in seconds. Finally, contact buttons and a form, designed for the pace of the stand, record an investor's or partner's interest without interrupting the conversation. Everything works on tablet and mobile. - 01 Delivered ### MIPIM stand landing Event page for the stand at the Palais des Festivals: brand message, Dubai / Tilal Al Ghaf context and a call to speak with the team in Cannes. - 02 Delivered ### Project showcase A selection of launches and neighbourhoods (including BO MONDE, Ghaf Woods, Lanai Islands, Amara, Alaya Beach) so visitors grasp the portfolio in seconds. - 03 Delivered ### Contact in the moment Form and CTAs designed for the stand: capture investor and partner interest instantly, without breaking the pace of face-to-face conversation. 03 Gallery ## The project in pictures 9 images · Open any image to see it full size. Cannes harbor and the Palais des Festivals, home of the Tilal Al Ghaf stand at MIPIM 2025. Four projects side by side: island villas from above, two pool villas and a modern façade. The BO MONDE scale model, one of the recent launches in the project showcase. Ghaf Woods: a double-height living room that opens onto the forest. Lanai Islands from above: villas set on islands in the lagoon. Amara: a contemporary villa exterior, another of the projects presented in Cannes. The lagoon beach at Tilal Al Ghaf, lined with villas and palm trees. Two people doing yoga in the lagoon at sunset. The mood of Alaya Beach: a dancer silhouetted against a sunset over the water. 04 Outcomes ## From stand visits to qualified conversations At the Palais des Festivals, the team had one tool to present Tilal Al Ghaf and its launches, on tablet or mobile. Visitors saw the projects on screen and left their details without stepping out of the conversation. Every visit had an immediate next step, with investor and partner interest already on record. - Tilal Al Ghaf's Dubai launches presented at the stand - Investor and partner interest captured in the moment - A clear digital journey on tablet and mobile - An immediate next step after every visit --- # JCDecaux × Transmilenio: live data on 3,000+ urban screens URL: https://slash-digital.io/en/case-studies/jcdecaux-transmilenio/ JCDecaux × Transmilenio An urban information platform for Bogotá: GPS bus arrivals, weather and air quality on digital totems, with QR continuity to mobile. Start a project See the gallery A Bogotá bus stop with a JCDecaux digital totem, the kind of screen the platform feeds. In short For JCDecaux and Transmilenio, we built the urban information platform that feeds more than 3,000 digital totems in Bogotá. It consolidates live APIs (GPS bus arrivals, weather, air quality) and pushes them to the screens with low latency. A QR code at the stop lets riders continue on their phone. The result: geo-referenced information across the network and an architecture ready to scale. Client JCDecaux × Transmilenio Sector Urban mobility · DOOH Where Bogotá Deliverables - Live data platform - API consolidation - Feeds for 3,000+ totems - QR continuity to mobile Services - DOOH & urban screens - Data solutions 3,000+urban screens fed live QRphysical → mobile continuity APIsGPS, weather and air quality live 01 The context ## Real-time information at city scale Anyone waiting at a Bogotá bus stop needs to know which bus is coming and when. That information depends on several sources: GPS positions, schedules, weather and air quality. The challenge was to inform an entire city in real time, at more than 3,000 stops, without lag or downtime. An arrival time only helps if it is current, so low latency and uptime were conditions of the project from the start. 02 Design and build ## From live APIs to thousands of totems We designed the platform data-first. It connects to the live APIs (GPS for bus arrivals, weather and air quality), consolidates the data and pushes it to thousands of totems with low latency. On screen, riders see the stop, the routes by color, the next buses in minutes with accessibility icons, the time, the temperature and an air-quality indicator; when no bus is scheduled, the screen says so. QR codes on the totem then carry the stop's routes to the rider's phone. - 01 ### Live GPS arrivals Arrival times come from the buses' GPS positions and appear in minutes, with the route number and accessibility icons, so riders see what is coming next. - 02 ### Weather and air quality Next to the arrivals, the totem shows the time, the temperature and an air-quality indicator, which turns the stop into a small urban information point. - 03 ### QR continuity to mobile QR codes on the totem send the stop's routes to the rider's phone, so the information stays with them after they walk away from the totem. - 04 ### Low latency at scale One data-first platform consolidates the live APIs and pushes updates to more than 3,000 totems, on an architecture that is ready to scale. 03 Gallery ## The project in pictures 2 images · Open any image to see it full size. Portrait totem: stop code, color-coded routes, air quality, temperature and a QR code to check the stop's routes. Arrival screen design: route 14 at Parque Metropolitano Simón Bolívar, next buses in minutes, weather and air quality. 04 Outcomes ## Geo-referenced information across the network At Bogotá's stops, riders read live arrivals, weather and air quality on the totem, and can take the stop's routes to their phone with a QR code. The information is geo-referenced across the network, which supports better trip planning. For JCDecaux and Transmilenio, the platform is an architecture ready to scale. - Live data on more than 3,000 digital totems in Bogotá - GPS arrivals, weather and air quality on one screen - QR continuity from the stop to the rider's phone - A data-first architecture ready to scale --- # L'Oréal × Guardians: from GenAI selfie to a 3D brand mosaic URL: https://slash-digital.io/en/case-studies/loreal-guardians/ L'Oréal · Guardians A GenAI face-swap for Celebrate Consumer Care plus a gigapixel Three.js mosaic where hundreds of avatars form the logo, navigable in 3D. Start a project See the gallery Inside the Guardians Mosaic: hundreds of face-swap avatars in orange, silver and purple Guardian suits. In short For L'Oréal's Guardians program, we built two connected pieces. Celebrate Consumer Care, a Peoplesnap experience in English and French, renders each participant as one of four Guardians through a GenAI face-swap. The Guardians Mosaic then gathers hundreds of those avatars into a gigapixel mosaic of the L'Oréal logo, navigable in 3D: individual selfies become one collective brand wall. Client L'Oréal · Guardians Sector Beauty · Consumer Care Deliverables - Peoplesnap face-swap experience - GenAI avatar pipeline - Three.js / WebGL mosaic - Protected 3D demo access Services - Interactive experiences - Artificial intelligence - Computer vision 01 · LiveConsumer Care face-swap 02 · LiveGuardians 3D mosaic 01 The context ## Two scales: the phone and the lobby L'Oréal's Guardians program called for two pieces at two scales. On the phone, a memorable Consumer Care activation in English and French, ready for internal events. In the room, a piece sized for event lobbies and private demos that would turn the Guardian community into a brand object. Both share the same cast: four Guardian archetypes in sci-fi suits. 02 Design and build ## From one selfie to hundreds of avatars Celebrate Consumer Care runs on Peoplesnap, our photobooth product. The landing sets out three steps: select a Guardian, smile, share. The participant picks one of the four Guardians and takes a selfie, and our GenAI pipeline swaps their face onto the chosen character. The result is a brand avatar they can download and share, in English or French. The Guardians Mosaic brings those avatars into a Three.js / WebGL app. Hundreds of them form a gigapixel mosaic of the L'Oréal logo, navigable in 3D from the whole logo down to a single face. Under the hood, a texture atlas packs the avatars together and shaders draw them on the GPU. Access is protected, for private demos and event lobbies. - 01 Live ### Celebrate Consumer Care PeopleSnap experience: 4-Guardian picker, selfie capture and GenAI face-swap pipeline: downloadable, shareable brand avatar (EN/FR), ready for internal events. - 02 Live ### Guardians Mosaic Three.js piece: hundreds of face-swap avatars form a gigapixel L'Oréal logo mosaic, navigable in 3D, for private demos and event lobbies. 03 Gallery ## The project in pictures 5 images · Open any image to see it full size. The four Guardian archetypes in sci-fi suits and headsets: the characters every participant can become. The Celebrate Consumer Care picker: each participant chooses which of the four Guardians to become. How to take part, in English or French: select a Guardian, smile, share. A generated avatar card: a participant rendered as a Guardian, with the CARE badge. One avatar from the mosaic, up close: every tile in the wall is a participant's face-swap portrait. 04 Outcomes ## Every selfie, a tile in the wall Both pieces are live. L'Oréal now has a bilingual face-swap experience ready for internal events, and a 3D mosaic ready to project in a lobby or open in a private demo. Participation is measurable, the user-generated content is of high quality, and the Guardian community has become a brand object. - Consumer Care face-swap live in English and French - Guardians Mosaic live for private demos and event lobbies - Measurable participation and high-quality user-generated content - A projection-ready piece built from participants' own avatars --- # Nigloland × L'Oréal: a Belle Époque AI souvenir in the park URL: https://slash-digital.io/en/case-studies/nigloland/ Nigloland 2026 × Supersonic / L'Oréal Photo souvenir at Nigloland for Supersonic / L'Oréal: QR funnel, attraction map, in-scene selfie and Belle Époque GenAI compose in French and English. Start a project See the gallery A Belle Époque AI souvenir: a couple in 1887 costume before an exhibition pavilion, with Supersonic badges. In short For Nigloland, a theme park in France, we built a mobile photo souvenir set in the Belle Époque world of Supersonic 1887. A QR code opens the journey, the park map guides visitors along a photo trail, and a GenAI pipeline turns their selfie into a period portrait, digital or printable. It is a pilot for physical-digital GenAI experiences in leisure. Client Nigloland × L'Oréal Sector Leisure · theme park Where France Year 2026 Deliverables - Mobile park app (FR/EN) - QR journey and park map - Belle Époque GenAI pipeline - Email capture - Digital and printable souvenir Services - Interactive experiences - Artificial intelligence - Web applications GenAIBelle Époque composition QRfrom the park to the phone FR · ENbilingual experience 01 The context ## A Belle Époque souvenir, no booth required Nigloland is a theme park in France. The goal was to turn a visit into a souvenir worth sharing, styled after the Belle Époque: the Supersonic 1887 universe, with the look of an Exposition Universelle. The constraints were practical. No heavy booth to install in the park, and no friction on the phone between scanning a code and getting the picture. 02 Design and build ## From a QR code to a period portrait We built a mobile park app in French and English, around four parts. A QR journey: codes placed in the park open the app, and the attraction map tracks each visitor's photo trail. In-scene experiences: the visitor picks a period costume and takes a selfie. Email capture, built into the flow. And a GenAI pipeline that composes the Belle Époque portrait as a digital souvenir, also printable. - 01 ### QR-triggered journey QR codes placed in the park open the experience on the visitor's phone. The home screen, with the park mascot pointing the way, leads to the attraction map. - 02 ### Map and photo trail The park map shows the zones and tracks the visitor's progress on the photo trail, with a clear instruction: scan the QR code on site. - 03 ### Costume first, then selfie Visitors choose their period costume, Voyageur or Voyageuse, then take an in-scene selfie. The GenAI pipeline composes a Belle Époque portrait in the Supersonic 1887 style. - 04 ### A souvenir to keep The portrait comes back in an Art Nouveau frame with the Supersonic 1887 badge, as a digital souvenir that can also be printed. Email capture is part of the flow. 03 Gallery ## The project in pictures 4 images · Open any image to see it full size. The mobile home screen: the park mascot points the way, and a QR code starts the adventure. The park map by zone, with photo-trail progress and a prompt to scan the QR on site. Before the camera opens, visitors pick a period costume: Voyageur or Voyageuse. The finished souvenir: a GenAI Belle Époque portrait in an Art Nouveau frame, with the Supersonic 1887 badge. 04 Outcomes ## A pilot for GenAI in leisure Nigloland now has a photo souvenir that fits in the visitor's pocket: a QR scan, a costume, a selfie, and a Belle Époque portrait to share, keep or print. Nothing heavy to install in the park. The project also serves as a pilot for physical-digital experiences with GenAI in leisure, a bridge between the brand and the park visit. - A shareable Belle Époque souvenir, straight from the visitor's phone - No heavy booth to install in the park - One bilingual journey, in French and English, with email capture - A physical-digital GenAI pilot for the leisure sector --- # Peoplesnap: from smartphone to DOOH mosaic, no booth URL: https://slash-digital.io/en/case-studies/peoplesnap/ Peoplesnap · Slash Experience A white-label web app that turns each guest's phone into the activation: QR, selfie, AI editing with 100+ parameters, social, print and on-screen mosaics. Start a project Watch the video Concept demo: a “Scan me, go live” arena screen and the phone flow, from welcome to selfie. In short Peoplesnap is our own photobooth product, from Slash Experience: a white-label web app that turns each guest's phone into the activation. Guests scan a QR code, take a selfie and get an AI-edited photo for social media, print or a DOOH mosaic, while the brand collects first-party data. We presented it at CES Las Vegas 2024, and it has run campaigns from Mouton Cadet to We Love Beer. Client Peoplesnap Sector Events · own product Deliverables - White-label photobooth web app - AI editing, 100+ parameters - DOOH mosaics and live camera - First-party data to CRM - Back office and analytics Services - Interactive experiences - DOOH & urban screens - Artificial intelligence - Digital products 0 hardwareOnly the guest's smartphone 7+ campaignsBrands and events in the live carousel +100 AI paramsCreative edit and live moderation Native DOOHMosaics and live camera on screen 01 The context ## Where the classic photo booth falls short Brands want user-generated content and engagement at their events. Traditional photo booths make that hard to scale. They bring heavy logistics and hardware costs, capture little useful data, and produce photos that are awkward to share on social media. We created Peoplesnap in 2023–2024 to take the booth out of the equation and put the experience on the phone every guest already carries. 02 Design and build ## A photobooth with no booth Every campaign runs on the same core, dressed in the client's brand. A guest scans a QR code, allows the camera, takes a selfie and completes the form. Peoplesnap edits the photo in seconds, live moderation checks it, and it goes out to social media, to print or into a mosaic on the event screens. Each entry also lands in the CRM, and the back office tracks participation. For CES Las Vegas 2024, the flow invited visitors to join a giant mosaic and try to win tickets for the Sphere. After the selfie, guests left their email to receive the mosaic as a digital memento. The same structure, restyled each time, has served Mouton Cadet, La French Tech Bordeaux, CEA Marcoule, Denver Pub Crawl, With Agency and We Love Beer. CES 2024 flow: join the giant mosaic, allow the camera, confirm the selfie, get the mosaic by email. - 01 Product ### White-label web app, not a booth Custom design (logo, copy, colours, backgrounds). Guests scan a QR, allow the camera, take the selfie and complete the form. Nothing to install, nothing to ship, nothing to build on site. - 02 AI ### Creative edit in seconds 100+ parameters: custom background, upscaling, relighting, avatars and cartoons, face swap, animated GIF, lip sync and live moderation so content is checked before it hits a screen or social. - 03 DOOH ### From phone to jumbo screen Photos feed digital mosaics and live camera in stadiums, trade shows, festivals and brand parties. The same flow that creates social UGC powers the event's screens: Peoplesnap as the participation layer for DOOH. - 04 Data ### First-party data and back office Built-in forms, challenges and sweepstakes. CSV export, participation analytics and a back office to tune the experience on the fly. GDPR end to end. 03 Gallery ## The project in pictures 14 images · Open any image to see it full size. A Peoplesnap photo mosaic on a giant outdoor screen by the sea, built from participants' photos. Mockup: an arena screen with a “Scan me, go live” QR, a fan photo and a sponsor slot. The outdoor mosaic next to phones running campaign flows: one product, each event with its own branding. The digital mosaic of CEA Marcoule's Journée des familles, made of participants' photos. A CES 2024 branded Peoplesnap photo as it looks when shared in an Instagram post. Client flows on peoplesnap.io, with the product promise: 1 content, 1,000 possibilities. The selfie capture step in the phone's browser: one button, no app to install. CES 2024 welcome screen: join the giant mosaic and try to win Sphere tickets, in three steps. Mouton Cadet event: guests pick their Mouton Cadet to generate a custom experience. La French Tech Bordeaux night: the flow in French, under the event's own branding. CEA Marcoule family day: the phone flow that feeds the day's digital mosaic. Denver Pub Crawl with a Mardi Gras theme: the same three steps, restyled for the event. With Agency summer party: the flow in the agency's black-and-white identity. We Love Beer festival: each selfie joins the digital mosaic, in the festival's gold branding. 04 Outcomes ## Zero hardware, measurable return A Peoplesnap activation goes live in hours, not weeks, scales to the size of the crowd and needs no hardware beyond the guest's smartphone. Its return is measurable in user-generated content, organic reach and data. For brands, it is a direct bridge from a fan's phone to the screens of a stadium, a trade show or a festival, GDPR-aligned from end to end. - Nothing to install, ship or build on site - 7+ campaigns, from Mouton Cadet to CES Las Vegas 2024 - First-party data into the CRM, with CSV export and analytics - The same product, restyled for every brand Visit peoplesnap.io ↗ --- # Alain Afflelou Magic Sunny: sunglasses try-on with selfies URL: https://slash-digital.io/en/case-studies/afflelou/ Alain Afflelou · Magic Sunny Alain Afflelou Magic Sunny activation: try on sunglasses on mobile with four selfies, clear camera-permission guidance and share or save in French and English. Start a project See the gallery A smiling model in round tinted sunglasses, on the orange that runs through the Magic Sunny app. In short For Alain Afflelou's Magic Sunny activation, we built a bilingual web app for trying on sunglasses in four selfies. The flow moves through the terms, a guided camera permission and the four shots, then ends with a choice: share them or keep them. The result is a retail activation ready for the store floor, with one clear funnel from capture to share. Client Alain Afflelou · Magic Sunny Sector Eyewear retail Deliverables - Bilingual store web app - Terms (CGU) onboarding - Guided camera permission - Four-selfie try-on flow - Share and download screen Services - Interactive experiences - Web applications 4selfies per guest Cameraguided permissions FR · ENbilingual activation 01 The context ## Selfies with frames, without friction Alain Afflelou wanted a memorable activation for Magic Sunny, for use in store and at events. The idea was simple: take selfies wearing different frames. The hard parts sat around it. Camera permissions had to be clear, since a web app cannot capture anything until the guest allows the camera, and sharing to social media had to take no effort. 02 Design and build ## Four steps, from terms to share We built a web app for the store and for events, in French and English. It runs in four steps. The guest accepts the terms (CGU), reads a short screen on camera access, takes four selfies wearing different sunglasses, then shares them or saves them to the phone. Each step does one job, which keeps the funnel easy to follow from capture to share. - 01 ### Terms first The intro screen shows the Magic Afflelou logo and a terms (CGU) checkbox. Guests accept the terms of use before the camera ever opens. - 02 ### A guided camera permission A dedicated screen explains the browser's camera prompt before it appears, so guests know what to tap when it does. - 03 ### 1, 2, 3… 4 selfies Guests take four selfies with different sunglasses. The app lays them out in a 2×2 grid, so the pairs can be compared side by side. - 04 ### Share or keep The last screen offers two exits: share the selfies on social media, or download them for guests who would rather keep them to themselves. 03 Gallery ## The project in pictures 4 images · Open any image to see it full size. The intro screen: Magic Afflelou logo and the terms (CGU) checkbox, accepted before the camera opens. The camera access screen explains the browser's permission prompt before it appears. 1, 2, 3… 4 selfies: a 2×2 grid to compare different sunglasses side by side. The last step: share the four selfies on social media, or download them and keep them. 04 Outcomes ## Ready for the store floor Afflelou now has an activation ready for both the store floor and events, opened in the phone's browser in French or English. The funnel is clear, from capture to share, and each guest ends up with four selfies in different sunglasses, posted to social media or kept on the phone. - Ready for the store floor and for events - Terms, camera and selfies in one guided flow - Share to social media or download from one screen - Available in French and English --- # Planet Fitness Mexico: out with ServiceDesk, owned software URL: https://slash-digital.io/en/case-studies/planet-fitness-mexico/ Planet Fitness Mexico A SaaS-exit program: ServiceDesk replaced by owned multi-area ticketing, PF Rooms shipped across CDMX, and the next product underway to turn off another license. Start a project See the gallery The ticketing dashboard: ticket totals, SLA, recent tickets, clubs and categories on one screen. In short For Planet Fitness Mexico we run a SaaS-exit program, one owned product at a time. We replaced ServiceDesk with multi-area ticketing and put PF Rooms on meeting-room door tablets across CDMX sites, with Microsoft Exchange / Graph as the source of truth. Both are live with no per-seat license, and the third chapter, the next vertical SaaS to replace, is in progress. Client Planet Fitness Mexico Sector Fitness Where Mexico City (CDMX) Deliverables - Owned multi-area ticketing - QR-linked assets and SLAs - PF Rooms door tablets - Exchange / Graph integration - Dashboards and reports Services - Custom software - Automation - Data solutions 01 · LiveServiceDesk → owned ticketing 02 · LivePF Rooms across CDMX sites 03 · In progressNext SaaS to turn off 01 The context ## Paying per seat for someone else's roadmap Planet Fitness Mexico was paying recurring SaaS licenses for tools it did not control. The service desk ran on ServiceDesk, operations were fragmented across teams, and room management would have meant one more generic workplace subscription. The company had little control over the roadmap, and cost pressure in Mexico was real. The goal was to turn those licenses off one by one and replace each with software Planet Fitness owns. 02 Design and build ## One chapter, one license turned off We design, build and run each chapter together with Planet Fitness. The first replaced ServiceDesk with a ticketing platform for IT, maintenance, HR, marketing and purchasing. Staff sign in with their corporate email and open a ticket: they find the asset by name, serial number or QR code, then set the area, priority, club and zone. Each ticket carries its location, assignee, related asset and SLA time. The asset module tracks what is operational, in maintenance or out of service. Reports cover SLA compliance, average MTTR and volume by area and club, alongside surveys and dashboards. The second chapter put PF Rooms on the doors of meeting rooms across CDMX sites. Each tablet reads Microsoft Exchange / Graph as the source of truth, so it shows live availability and the day's agenda, and anyone can book the room on the spot. Calendars stay where they already live, with no extra workplace subscription. The third chapter, now in progress, follows the same pattern for the next licensed vertical SaaS. - 01 Live ### Ticketing without ServiceDesk Multi-area platform (IT, Maintenance, HR, Marketing, Purchasing): tickets, QR-linked assets, SLAs, surveys and dashboards. Zero per-seat license, roadmap under Planet Fitness control. - 02 Live ### PF Rooms at the door Door tablets across CDMX: live availability, day agenda and quick booking, with Microsoft Exchange / Graph as source of truth, without another generic workplace subscription. - 03 In progress ### The next SaaS we turn off Third chapter: replace the next licensed vertical SaaS with an owned product, same pattern as ticketing and PF Rooms, no paid seats. 03 Gallery ## The project in pictures 9 images · Open any image to see it full size. Ticket Management System sign-in: every team enters with its corporate email. The ticket inbox, with quick views and filters by club and department. A ticket in detail: location, assignee, related asset, SLA time, quick status actions, reassignment and PDF. New ticket: find the asset by name, serial number or QR, then set area, priority, club, zone. Asset management: operational, in maintenance or out of service, with scheduled maintenance and QR scanning. Reports and analytics: SLA compliance, average MTTR, closed tickets and volume by area and club. System configuration: the admin screen where ticket categories are defined and kept up to date. PF Rooms: the room shows as available, with a button to book it and the day's agenda. PF Rooms on a portrait door tablet, with availability read live from Microsoft Exchange. 04 Outcomes ## Two license savings live, a third underway ServiceDesk is gone and PF Rooms is live on the floor. Planet Fitness now runs its service desk and its meeting rooms on software it owns, with no paid seats, and the roadmap belongs to the company. The same pattern now targets a third license: the next vertical SaaS to turn off is in progress. - ServiceDesk replaced by owned multi-area ticketing, live in production - PF Rooms live on door tablets across CDMX sites - Zero per-seat licenses, roadmap owned by Planet Fitness - Third chapter in progress: the next licensed vertical SaaS --- # JCDecaux Mexico × Paperless: paper to internal platform URL: https://slash-digital.io/en/case-studies/jcdecaux-paperless/ JCDecaux Mexico · Paperless Purchasing, inventory, work orders, advances and refunds centralized in an internal platform with SSO, approvals and audit-ready traceability. Start a project See the gallery The Paperless dashboard: pending documents and an expense summary in MXN, next to the module menu. In short For JCDecaux Mexico, we built Paperless, an internal platform that took critical processes off paper. Purchasing, inventory, work orders, advances and refunds now live in one place, behind corporate Single Sign-On, with custom approval flows, dashboards and audit-ready records. The result: key processes digitized, with stronger security and transparency. Client JCDecaux Mexico Sector Enterprise operations Where Mexico Deliverables - Modular internal platform - Single Sign-On access - Custom approval flows - Dashboards and audit-ready records Services - Custom software - Web applications - Automation SSOcorporate access from day one 0critical processes on paper 100%approvals with traceability 01 The context ## Critical processes, still on paper At JCDecaux Mexico, critical processes still ran on paper. That meant errors, delays and zero traceability, in an environment that demands control. A request could not be followed from submission to approval, and there was no record an audit could rely on. The company needed a secure internal tool, open through corporate accounts, where each process would follow its approval path and leave a trace. 02 Design and build ## One login, one module per process We built Paperless as a modular internal platform. Employees sign in with their corporate account through Single Sign-On; administrators have their own login. Each process has its module: purchasing, advances, refunds, catalogs and work orders, with inventory in the same platform. A request moves through a custom approval flow, and a dashboard shows what is pending: documents awaiting action and an expense summary in Mexican pesos. Every step is recorded, so the history is ready for an audit. - 01 ### SSO from day one Employees sign in with their corporate account from the first day, with no extra password to remember; administrators have their own access to the portal. - 02 ### Custom approval flows Each process follows its own approval path, defined for JCDecaux Mexico, and every approval stays traceable from start to finish. - 03 ### One module per process Purchasing, advances, refunds, catalogs and work orders each have their own module in the same interface, alongside inventory. - 04 ### Dashboards and audit-ready records The dashboard shows pending documents and the expense summary, and the records of every process are ready for an audit. 03 Gallery ## The project in pictures 2 images · Open any image to see it full size. The Paperless sign-in screen: Single Sign-On for employees and a separate administrator login. The JCDecaux Paperless home: module menu, plus a user menu with sub-authorizations and the administrator portal. 04 Outcomes ## Off paper, on the record No critical process depends on paper anymore. Purchasing, inventory, work orders, advances and refunds are digitized in one platform, every approval is traceable and records are audit-ready. Security and transparency improve with SSO access and a record of every step. The platform is also a flexible base for the rest of the ops stack; its operational counterpart, the Asset Manager, covers corporate assets. - Zero critical processes left on paper - Every approval traceable, every record audit-ready - Corporate SSO access from day one - A flexible base for the rest of the ops stack --- # JCDecaux Mexico Asset Manager: inventory under control URL: https://slash-digital.io/en/case-studies/jcdecaux-asset-manager/ JCDecaux Mexico · Asset Manager Corporate asset manager for JCDecaux Mexico: QR-tagged inventory, signed assignment letters, contracts, PDF/CSV reports and Admin IT / Auditor roles. Paperless’s operational counterpart. Start a project See the gallery Inventory management: asset totals, CSV import and export, and the filterable list of equipment. In short For JCDecaux Mexico, we built the Asset Manager, the operational counterpart of Paperless. It brings corporate assets into one platform: a QR-tagged inventory, assignments to employees with signed assignment letters, contracts, and PDF and CSV reports, with Admin IT and Auditor roles. Asset operations are now digitized, with fewer assignment errors and exports ready for audit. Client JCDecaux Mexico · Asset Manager Sector Enterprise asset management Where Mexico Deliverables - QR-tagged inventory - Signed assignment letters - Contracts and employee records - PDF, CSV and Excel reports - Admin IT and Auditor roles Services - Custom software - Web applications - Automation QRtagged inventory PDF/CSVaudit-ready reports RolesIT Admin · Auditor 01 The context ## Who has what, and since when Equipment, assignments and responsibility letters (cartas responsivas) lived in scattered documents, with no audit-ready traceability. Without a single source, it was hard to know which asset each employee had, since when, and whether the letter was signed, and nothing was ready to show an auditor. JCDecaux Mexico needed one platform for its corporate assets, designed as the operational counterpart of Paperless. 02 Design and build ## From corporate login to audit-ready export We built the Asset Manager as a web platform with a corporate login. The flow follows an asset's life: it enters the inventory with its QR identifier (or arrives through a CSV import), is assigned to an employee with a signed assignment letter, is tracked over time with delivery and return dates, and shows up in the reports an audit asks for. Contracts, asset types, employees and system users each have their own section, and two roles, Admin IT and Auditor, organize access. - 01 ### QR-tagged inventory Assets carry QR identifiers, and the inventory shows totals, available and assigned items and total value, with search plus quick and advanced filters. - 02 ### Traceable assignments Assignments record the employee, the company, the delivery and return dates and the status, with active and finished assignments in plain view. - 03 ### Reports built for audits Reports show assets by category and by company, flag the ones with problems, and export to PDF, CSV or Excel when an audit needs them. - 04 ### Admin IT and Auditor roles The corporate login separates two roles, Admin IT and Auditor: the IT team and auditors work in the same platform, each with their own role. 03 Gallery ## The project in pictures 4 images · Open any image to see it full size. Corporate sign-in to the JCDecaux Mexico Asset Management System. The inventory: totals for all, available and assigned assets, total value, search, quick and advanced filters. Active and finished assignments: employee, company, delivery and return dates, and the status of each. Reports and statistics: assets by category and by company, assets with problems, and CSV, Excel and PDF export. 04 Outcomes ## Asset operations, digitized and audit-ready JCDecaux Mexico now manages its corporate assets in one platform instead of scattered documents. Assets, assignments, contracts and employees sit in the same place, which reduces assignment errors. When an audit comes, the team exports the reports it needs in PDF, CSV or Excel. Paperless covers the processes and the Asset Manager covers the equipment: two sides of the same internal operations. - Asset operations digitized in one platform - Fewer errors when assigning equipment to employees - Audit-ready exports in PDF, CSV and Excel - Separate roles for the IT team and auditors --- # Colombia Tech Fest: the official festival app URL: https://slash-digital.io/en/case-studies/colombia-tech-fest/ Colombia Tech Fest The attendee app: invite-only onboarding, live agenda, directory, 1:1 meetings, chat, offline maps and notifications, built for event-scale peaks. Start a project See the gallery Step 1 of 4: a guest enters an invitation code, the only way to create an account. In short We built the official attendee app of Colombia Tech Fest, a product the festival owns. Guests joined by invitation only, then found a live agenda, a directory, 1:1 meetings, chat, offline maps and notifications in one place, while the organizers ran it with admin tooling. Built for the event's peak concurrency, the app is ready for future editions. Client Colombia Tech Fest Sector Tech events Where Colombia Deliverables - Invite-only onboarding - Live agenda and directory - 1:1 meetings and chat - Offline maps and notifications - Admin tooling Services - Mobile apps - Digital products - Web applications 1:1meetings and chat between guests Offlinemaps without connection Inviteonly: no open signup 01 The context ## Thousands of guests, no open door Colombia Tech Fest needed one app to bring together the agenda, networking and messaging for thousands of invited guests. The guest list was closed, so there could be no open public signup: every account had to start from an invitation. And the app had to hold up at the event's peak, when most guests connect at the same time. 02 Design and build ## Four steps in, then the whole festival Access starts with an invitation. A first-time guest enters the invitation code, accepts the privacy policy, creates a password and completes a profile: four steps, visible on screen from the start. After that, guests sign in with the email and password of their Fest account. The app has no public signup form. Inside, a tab bar leads to home, map, agenda, networking, chat and a personal calendar. The agenda updates live, maps work offline, and guests receive notifications and save favorites. In networking, a guest searches the directory, opens a profile with role, industry, company and 30-minute meeting slots, requests a meeting and continues in a 1:1 chat. Behind the scenes, admin tooling lets the festival run its own product. Networking in the app: find an attendee, check their 30-minute availability, then continue in a 1:1 chat. - 01 ### Invite-only by design Every account starts from an invitation code and four clear steps. The festival knows who gets in, and there is no open signup to moderate. - 02 ### Meetings in 30-minute slots Profiles show role, industry, company and availability in 30-minute blocks, so guests see when someone is free before they request a meeting and move to chat. - 03 ### Offline maps, live agenda Maps work without a connection and the agenda updates live, so the app stays useful on site even when the network drops. - 04 ### A product the festival owns Invitations, favorites, meetings, chat and admin tooling form one product that belongs to the festival, ready for its future editions. 03 Gallery ## The project in pictures 2 images · Open any image to see it full size. Onboarding in four steps: invitation, privacy, password and profile. It starts by validating the code. Sign-in with a Fest account email and password: no public signup, and first-timers use their invitation. 04 Outcomes ## Ready for the next editions Guests had a complete, live attendee experience in a single app, from the invitation to the last 1:1 chat, with no open signup. Because the product belongs to the festival and comes with its own admin tooling, it is ready for future editions. - A complete live attendee experience in one app - Invite-only access, with no open public signup - Offline maps and a live agenda for every guest - An owned product, ready for future editions --- # Paperslate: E-Ink tablet with AI retranscription URL: https://slash-digital.io/en/case-studies/paperslate/ Paperslate · Hardware + software We delivered the product and e-commerce stack for Paperslate, an E-Ink tablet with a real writing feel: site, AI, usage and FAQ pages, Pro and Paperslate catalogue, and a pipeline that retranscribes to Word or PDF. Start a project See the gallery A hand writes meeting notes and a diagram on the Paperslate E-Ink tablet. In short Paperslate is a digital paper (E-Ink) tablet with a real writing feel, in Pro and standard models, plus a GenAI layer in beta that retranscribes handwriting, formulas and diagrams to Word or PDF. We delivered the product and e-commerce stack: site, AI, usage and FAQ pages, catalogue and the retranscription pipeline. The result: a hardware and software brand in market, with packs on sale. Client Paperslate · E-Ink + AI Sector E-Ink digital paper Deliverables - Site and e-commerce - AI, usage and FAQ pages - Pro and Paperslate catalogue - AI retranscription pipeline Services - Digital products - Artificial intelligence - Web applications 2models: Pro and standard Word / PDFAI retranscription in beta E-Inkreal writing feel 01 The context ## From notebook to a credible digital flow Paperslate's challenge was to move handwritten notes into a credible digital flow. That meant holding three promises at once: the feel of paper, long battery life and AI transcription without friction. The brand also had to explain a product that is hardware and software at the same time, in two models: Pro and standard. 02 Design and build ## From storefront to retranscription pipeline We delivered the product and e-commerce stack. The site presents the tablet and sells the packs online. Three pages explain the product: AI, usage and FAQ. The catalogue brings together the Pro and Paperslate models. And the retranscription pipeline, in beta, turns handwriting, formulas and diagrams into Word or PDF files. - 01 ### Packs on sale The e-commerce sells the tablet in packs, with a product page for each model, Pro and standard, and direct purchase from the site. - 02 ### Pages that explain The AI, usage and FAQ pages answer the questions of anyone moving from paper to an E-Ink tablet. - 03 ### Retranscription in beta The GenAI pipeline retranscribes handwriting, formulas and diagrams to Word or PDF. It is in beta, and the site says so plainly. 03 Gallery ## The project in pictures 4 images · Open any image to see it full size. The home page hero opens on the Paperslate packs, the offer sold in the online store. The AI retranscription page, flagged as beta, presents the GenAI layer behind the tablet. The Paperslate Pro product page in the store: the tablet with its stylus. Hands take notes with the stylus on the Paperslate Pro, as they would on paper. 04 Outcomes ## A digital paper brand in market Paperslate is now in market as a hardware and software brand, with packs on sale. Its story is clear and backed by AI: the tablet keeps the feel of paper, and the GenAI layer, in beta, takes those notes to Word or PDF. - A hardware and software brand in market - Packs on sale in the online store - A clear product story, backed by AI - An AI retranscription pipeline in beta --- # Enviable AI: aspirational portraits generated by AI URL: https://slash-digital.io/en/case-studies/enviable-ai/ Enviable AI · GenAI product We built Enviable AI, a GenAI product that turns a portrait into a professional-quality image: a fast mobile flow, scene templates, credit-based monetisation and a privacy-first architecture aligned with GDPR. Start a project All case studies Before and after: a studio portrait, then the same person placed in a generated scene. In short Enviable AI turns a portrait into a professional-quality aspirational image. We built the web product: a GenAI engine, a fast mobile flow, scene templates and credit-based monetisation, on a privacy-first architecture aligned with GDPR. In its first weeks it generated thousands of images with high engagement: a proprietary AI product, validated in the market. Client Enviable AI · AI portraits Sector Generative AI Deliverables - GenAI engine - Fast mobile flow - Scene templates - Credit-based monetisation - Privacy-first architecture Services - Artificial intelligence - Digital products Thousandsof images in the first weeks Creditsmonetisation model GDPRprivacy-first architecture 01 The context ## Studio quality, no photographer Enviable AI starts from a promise: studio-quality images of yourself, without a photographer, on your phone, in seconds. That set two requirements. The flow had to stay quick on a small screen. And it could not compromise personal data: a portrait identifies a person, so privacy had to be built into the architecture from the start. 02 Design and build ## From portrait to scene, on a phone We built Enviable AI as a privacy-first web product designed for the phone. A person uploads a portrait, picks a scene template and the GenAI engine generates the aspirational image. The interface is responsive and the flow stays fast from upload to result. Monetisation runs on credits, and the architecture is aligned with GDPR. - 01 ### Scene templates Scene templates guide the generation: a person picks a setting and the GenAI engine places their portrait in it, at professional quality. - 02 ### A fast mobile flow Everything happens on the phone, in a responsive interface: upload the portrait, choose the scene and get the image, with no photographer or studio. - 03 ### Credit-based monetisation People pay with credits for the images they generate, which ties revenue directly to how much the product is used. - 04 ### Privacy-first architecture Portraits are personal data. The architecture was designed privacy-first and aligned with GDPR, so creating images does not compromise that data. 03 Gallery ## The project in pictures 1 images · Open any image to see it full size. The landing page explains the product at a glance: headline, then a before and after carousel. 04 Outcomes ## Thousands of images in the first weeks The product generated thousands of images in its first weeks, with high engagement. For Enviable AI, that meant a proprietary AI product validated in the market: a GenAI engine, a fast mobile flow, a credit-based revenue model and a privacy-first foundation aligned with GDPR. - Thousands of images generated in the first weeks - High engagement from users of the product - A proprietary AI product, validated in the market - Privacy built into the architecture, aligned with GDPR --- # UVAS Trading: premium e-commerce for wines and champagnes URL: https://slash-digital.io/en/case-studies/uvas-trading/ UVAS Trading · Madrid A custom online store that sells wines and champagnes from 25 regions to trade buyers and consumers, with filters by country, grape and style, tasting notes, secure payment and a CMS the team runs itself. Start a project All case studies The store's hero: three champagne bottles on stone under a free-shipping banner, setting the brand's tone. In short For UVAS Trading in Madrid, we built a B2B and B2C e-commerce for wines and champagnes from 25 regions: filters by country, grape and style, product pages with tasting notes and secure payment. A CMS lets the team run the catalogue without calling an agency. The result: stronger credibility online, more orders from trade and consumers, and lighter internal operations. Client UVAS Trading · Madrid Sector Wines and champagnes Where Madrid Deliverables - B2B and B2C e-commerce - Filterable catalogue - Tasting-note product pages - Secure payment - Self-managed CMS Services - Web applications - Digital products 25wine and champagne regions B2B + B2Cone store, two audiences CMScatalogue run without an agency 01 The context ## One store for trade and consumers UVAS Trading sells wines and champagnes from 25 regions to two audiences: trade buyers and consumers. Its e-commerce had to carry the elegance of the brand and still sell without friction to both. There was a third condition: the team had to run the catalogue on its own, without calling an agency for every change. 02 Design and build ## From grape filter to secure payment We built a custom e-commerce around the buying path. A customer filters the catalogue by country, grape and style, opens a product page with tasting notes and pays through a secure checkout. Behind the store, a CMS lets the UVAS Trading team manage the range without calling an agency. The same store takes B2B and B2C orders. - 01 ### Country, grape, style The catalogue of 25 regions filters by country, grape and style, the criteria people already use to choose a wine. - 02 ### Tasting-note product pages Product pages carry tasting notes, so trade buyers and consumers can decide with the right information before they order. - 03 ### Secure payment A secure checkout closes the buying path: the basic condition for turning trust in the brand into an order. - 04 ### A CMS for the team The team manages the catalogue from a CMS and adjusts the range when it needs to, without waiting for an agency. 03 Gallery ## The project in pictures 1 images · Open any image to see it full size. The catalogue: champagne product cards with filters by country, grape and style to find the right bottle. 04 Outcomes ## A more credible store, more orders UVAS Trading gained credibility online and won more orders, from trade buyers and consumers alike. Wines and champagnes from 25 regions sit in one catalogue that the team runs on its own. That lightens internal operations: range changes do not depend on an agency. - Stronger credibility online for the brand - More B2B and B2C orders - A catalogue the team runs without an agency - Lighter internal operations for the team --- # Nigloween 2025: AI costume selfies at Nigloland URL: https://slash-digital.io/en/case-studies/nigloween/ Nigloland · Nigloween 2025 During Nigloween, Nigloland's Halloween season, visitors scanned a QR code, picked pumpkin, vampire, witch or skeleton, took a selfie and got a GenAI edit in seconds. No booth, no native app. Start a project See the gallery The Nigloween photo spot: an arch with the Nigloween sign, two costumed mascot statues and a pumpkin throne. In short For Nigloween 2025, the Halloween season at Nigloland, we put an AI costume photo experience on visitors' phones. Built on Peoplesnap, our photobooth product, the web app ran from a QR code to a GenAI-edited souvenir, with no booth and no native app. It was Nigloween's first AI photo tool, and the trade press covered it. Client Nigloland · Nigloween 2025 Sector Leisure · theme park Where France Year 2025 Deliverables - Costume selfie web app - QR entry in the park - GenAI photo edit - Shareable Halloween souvenir Services - Interactive experiences - Artificial intelligence 4Halloween costumes Secondsfrom selfie to GenAI edit 1stAI photo tool at Nigloween 01 The context ## Halloween on the phone, not in a booth Nigloween is the Halloween season at Nigloland, a theme park in France. For 2025, the brief called for a memorable AI photo that visitors could take on their phones and share. Two constraints came with it: no booth to install in the park, and no native app to download before taking part. 02 Design and build ## From QR code to costume portrait We built the experience on Peoplesnap, our photobooth product. A QR code in the park opened a web app in French. Visitors chose a costume (pumpkin, vampire, witch or skeleton), took a selfie, and our GenAI edit returned the costumed portrait in seconds. The result was a souvenir made for sharing, created on the visitor's own phone, with nothing to install. - 01 ### Four seasonal costumes Pumpkin, vampire, witch or skeleton: visitors picked their costume on screen, in a French-language experience featuring the park's mascots. - 02 ### From QR to souvenir A QR code in the park opened the web app. After the costume pick and the selfie, the GenAI edit came back in seconds, ready to share. - 03 ### No booth, no download Everything ran in the phone's browser: no booth to set up in the park, and no native app for visitors to install. 03 Gallery ## The project in pictures 2 images · Open any image to see it full size. The costume choice on a phone, in front of a Nigloween arch: everything runs in the browser. Nigloween key visual: a girl in clown costume, a big top, a roller coaster and a full moon. 04 Outcomes ## Nigloween's first AI photo tool The app gave each visitor a costumed portrait made on their own phone, ready to post, and every shared photo added to the season's user-generated content. The experience bridged the physical park and the phone, with no booth and no app to install. As Nigloween's first AI photo tool, it drew coverage in the trade press. - Nigloween's first AI photo tool, covered by the trade press - Costume selfies edited by GenAI in seconds - Seasonal user-generated content from visitors' own phones - No booth in the park, no native app to install --- # Cardales de Cayalá: 3D wayfinding for premium retail URL: https://slash-digital.io/en/case-studies/jcdecaux-cardales/ JCDecaux Guatemala · Cardales de Cayalá Interactive wayfinding for Cardales de Cayalá in Guatemala: a 3D map of the mall, step-by-step routes on screen, itineraries sent to mobile by QR and an updatable venue directory. Start a project All case studies The Cardales de Cayalá map with a route to a venue and its card: directions and call buttons. In short For JCDecaux Guatemala, we built an interactive wayfinding system for Cardales de Cayalá. Visitors find a venue on a 3D map of the mall, follow the route step by step on screen, and take the itinerary with them on their phone through a QR code. The result: navigation without friction, more foot traffic to key venues and a stronger premium image. Client JCDecaux Guatemala · Cardales de Cayalá Sector Premium retail · Wayfinding Where Guatemala Deliverables - 3D mall map - Step-by-step routes - QR itinerary handoff - Updatable venue directory Services - DOOH & urban screens - Interactive experiences - Web applications 3Dmap of the mall Step by steproutes on screen QRitinerary to mobile 01 The context ## Finding the way in a large, complex site Cardales de Cayalá is a premium retail destination in Guatemala. The site is large and architecturally complex, so finding a specific venue is not always obvious for a visitor. The goal had two parts: guide people through it without friction, and reinforce the destination's premium positioning. The answer had to go further than a map: draw the route to the chosen venue, then stay with the visitor away from the screen. 02 Design and build ## From the screen to the visitor's phone We built the wayfinding platform on a 3D map of the mall. A visitor picks a venue on the map, and its card opens with buttons to call or get directions. The screen then draws the route step by step. A QR code hands the itinerary over to the visitor's phone, so they can follow it while they walk. Behind the scenes, an administrator login and an updatable venue directory keep the information current. - 01 ### 3D map of the mall A 3D map with venue pins lets visitors read a large, architecturally complex site at a glance, before they start walking. - 02 ### Step-by-step routes Once a venue is chosen, the screen draws the route to it step by step, next to a card with call and directions buttons. - 03 ### QR handoff to mobile A QR code transfers the itinerary to the visitor's phone, so the route stays with them once they leave the screen behind. - 04 ### Updatable venue directory The venue directory can be updated, so the map keeps reflecting the venues on site as they change. 03 Gallery ## The project in pictures 1 images · Open any image to see it full size. The wayfinding interface: administrator sign-in next to the map with its venue pins. 04 Outcomes ## Frictionless visits, busier key venues Visitors move through Cardales de Cayalá without friction: they find the venue, see the route and carry it on their phone. Key venues get more foot traffic, and the wayfinding itself reinforces the destination's premium image. For JCDecaux Guatemala, the platform is also a base for smart retail. - Frictionless navigation across a large, complex site - More foot traffic to key venues - A stronger premium image for the destination - A base for smart retail --- # Rose Legal: owned practice management after Smokeball URL: https://slash-digital.io/en/case-studies/rose-legal/ Rose Legal · Smokeball exit We built Rose Legal (Australia) a complete practice management platform to replace Smokeball: clients, matters, time and billing, LPUGR-aligned trust accounting, documents and role-based access, with no paid seats. Start a project See the gallery Rose Legal's platform on laptop, tablet and phone: one owned system in place of Smokeball. In short We built Rose Legal, an Australian law firm, a complete practice management platform to replace Smokeball. Clients, matters, time and billing, LPUGR-aligned trust accounting, documents, conflict search, reporting and role-based access now live in one system the firm owns. The firm no longer pays per seat and sets its own roadmap. Client Rose Legal · Smokeball exit Sector Legal Where Australia Deliverables - Practice management platform - LPUGR trust accounting - Time and billing - Conflict search and templates - Audit log and RBAC Services - Custom software - Web applications 0paid seats LPUGRaligned trust accounting RBACroles and audit log 01 The context ## Per seat, locked in, hard to adapt Rose Legal, a law firm in Australia, ran its practice on Smokeball. Each user meant another per-seat license, the firm was locked into one vendor, and there was little room to adapt the product to Australian operations. Trust accounting was the most sensitive part: in Australia it has to be aligned with the LPUGR rules, and a third-party product left little space to adjust it. 02 Design and build ## One owned platform for the whole practice We built an owned full-stack platform around the firm's daily work. After sign-in, a dashboard gives the overview. From there, the team opens a client and its matters, records time and prepares invoices in the same place. On the trust side, receipts, transfers and reconciliation run through a module aligned with LPUGR and produce their PDFs. Documents, templates, conflict search and reporting live in the same system, role-based access (RBAC) sets who does what, and the audit log keeps the record. - 01 ### LPUGR trust accounting Trust receipts, transfers and reconciliation follow the Australian rules (LPUGR), with PDFs, so the firm is ready for trust compliance. - 02 ### Clients, matters and billing Clients, matters, time and invoices share one system the firm owns, so a new user no longer means a new paid seat. - 03 ### Conflict search and templates Conflict search, documents and templates sit next to the matters they serve, in the same workspace as the rest of the practice. - 04 ### Audit log and RBAC Role-based access decides who can see and change what, and the audit log keeps a trace of activity across the platform. 03 Gallery ## The project in pictures 2 images · Open any image to see it full size. Sign-in to the Legal Practice Management System: client management, trust compliance, matter tracking, billing and document automation. Password reset in the Rose Legal interface: enter an email, get a reset link, return to sign-in. 04 Outcomes ## No paid seats, trust compliance ready Rose Legal now runs on owned software in place of Smokeball. The firm pays no per-seat licenses, its trust accounting is ready for compliance with the Australian rules, and every future change goes on a roadmap it controls. When the practice needs an adjustment, the decision stays with the firm. - Smokeball replaced by owned practice management software - No per-seat licenses left to pay - Trust accounting aligned with LPUGR, compliance ready - Roadmap under the firm's control --- # Réseau Mazette: native Flutter app for iOS and Android URL: https://slash-digital.io/en/case-studies/mazette/ Réseau Mazette · Flutter app Réseau Mazette connects French-speaking professionals and entrepreneurs in Spain. We shipped its native Flutter app on the App Store and Google Play: directory, events, messaging, memberships and the Mia AI copilot. Start a project See the gallery The app's events screen: a networking drink in Málaga, with city filters and the tab bar. In short Réseau Mazette connects French-speaking professionals and entrepreneurs in Spain. We shipped its full native app in Flutter, published on the App Store and Google Play, with the directory, events, messaging, memberships and Mia, an AI copilot. We also delivered the branded web experience, so members find the same network on the phone and on the web. Client Réseau Mazette · Flutter app Sector Professional network Where Spain Deliverables - Native Flutter app - App Store and Google Play - Mia AI copilot - Directory, events and messaging - Branded web experience Services - Mobile apps - AI chatbots - Web applications Flutternative app for iOS and Android 2stores: App Store and Google Play MiaAI copilot in the app 01 The context ## A production app with the web's identity Réseau Mazette brings together French-speaking professionals and entrepreneurs across Spain, with events in cities such as Málaga, Barcelona, Valencia and Madrid. The network needed a production app for iOS and Android, a full product rather than a light companion to the website. It had to carry the same brand as the web and keep networking, events and Mia, the AI copilot, always within reach. 02 Design and build ## One Flutter app, two stores, Mia inside We built the app in Flutter and published it on the App Store and Google Play. The home screen greets the member and shows upcoming events and the newest members. The directory helps members find each other, events can be filtered by city, and messaging keeps the conversation going in the app. Memberships are handled in the app too, and a dedicated tab opens a chat with Mia, the AI copilot. The app is aligned with the web rebuild, and we also delivered the branded web experience around it. - 01 ### Native on iOS and Android One Flutter app ships to both stores, so every member gets the same full product whatever phone they carry. - 02 ### Mia, the AI copilot Mia has its own tab next to the directory and events: members open the chat and ask their questions there. - 03 ### Events by city Events filter by city, from Málaga to Barcelona, and the home screen brings the next ones to the top. - 04 ### Directory and messaging Member cards with photo and role make the right contact easy to find, and messaging carries the conversation on from there. 03 Gallery ## The project in pictures 6 images · Open any image to see it full size. App home: a personal greeting, upcoming events and the newest members of the network. The events list, filtered by city from Málaga to Barcelona, so members see what is near them. Mia, the AI copilot, greets members in its own tab and waits for their questions. The web directory: member cards with photo and role, plus filters to find the right contact. Events on the web: talk and networking in Valencia, AI business training in Madrid, a pitch contest. The Notre réseau web page: the network's manifesto, less ego and more business. 04 Outcomes ## The network, in every member's pocket Réseau Mazette now has a native app on iOS and Android, live on both stores, with AI built into the member experience and a website that supports it. Members carry the directory, events, their messages and Mia in one app, and the network keeps one brand from the web to the phone. - Native Flutter app published on the App Store and Google Play - Mia, the AI copilot, built into the app - Directory, events, messaging and memberships in one place - App and branded web experience under one identity --- # Rotary RM: music publishing platform with SACEM tools URL: https://slash-digital.io/en/case-studies/rotary-rm/ Rotary RM · Music publishing Rotary RM publishes music from the electronic scene in Paris. We built its platform: catalogue, contract wizards, e-signatures, artist and label portals and a SACEM deposit assistant, behind secure access. Start a project All case studies A vinyl record with a neon label: Rotary RM, music publishing for the electronic scene in Paris. In short Rotary RM is a music publisher for the electronic scene in Paris. We built its publishing platform: catalogue, contract wizards, e-signatures, artist and label portals and a SACEM deposit assistant, behind secure access. Contracts and SACEM répartitions left spreadsheets and email for one space shared by admins and artists. Client Rotary RM · Music publishing Sector Music publishing Where Paris Deliverables - Catalogue and contract wizards - Contract e-signatures - Artist and label portals - SACEM deposit assistant - Secure access Services - Custom software - Web applications - Automation 1shared admin / artist space SACEMdeposit assistant Portalsfor artists and labels 01 The context ## Contracts in spreadsheets, répartitions in inboxes Rotary RM publishes music from the electronic scene in Paris. Its contracts and SACEM répartitions lived in spreadsheets and email, and there was no single space where the publishing team and the artists could see the same information. The publisher needed a product built for its own trade, able to follow the catalogue as it grows. 02 Design and build ## From contract wizard to SACEM deposit We built a web platform with two sides behind secure access: an admin space for the publisher and portals for artists and labels. The catalogue is the base. A new agreement starts in a contract wizard that walks the team through it step by step, then gets signed electronically without leaving the platform. SACEM deposits have their own assistant, and répartitions are kept in the same place instead of spreadsheets. - 01 ### Contracts without email chains Wizards structure each contract step by step and e-signatures close it in the platform, so agreements stop living in inboxes. - 02 ### SACEM in the workflow Deposits are prepared with a dedicated assistant and répartitions are kept in the platform, where they no longer depend on spreadsheets. - 03 ### Portals for artists and labels Artists and labels sign in to their own portal, inside the same secure space the publishing team uses for administration. - 04 ### A catalogue at the core Contracts, deposits and portals all build on the same catalogue, shared by the admin and artist sides behind secure access. 03 Gallery ## The project in pictures 1 images · Open any image to see it full size. Entry screen: a dark visual on the electronic scene, next to secure sign-in for admins and artists. 04 Outcomes ## Digital publishing ops, ready to scale Rotary RM's publishing operations are now digital. Contracts, signatures, SACEM deposits and répartitions sit in one space instead of spreadsheets and email, and artists reach the same platform through their portal. The result is a vertical product built for music publishing, ready to scale with the catalogue. - Contracts and SACEM répartitions moved out of spreadsheets and email - One secure space for admins, artists and labels - Guided contract wizards with e-signatures - A vertical product ready to scale the catalogue --- # Veto: end-to-end messaging without a phone number URL: https://slash-digital.io/en/case-studies/veto-messenger/ Veto · Private messenger Veto is an extreme-privacy messenger: local identity, keys on the device, end-to-end encryption and relays that only see ciphertext. We built the trilingual site and the /app PWA, released as an honest public alpha. Start a project All case studies Veto's landing message (no phone number, keys on your device) beside the app's chat interface. In short Veto is an extreme-privacy messenger: identity is local, with no phone number, and messages are end-to-end encrypted. We built its trilingual marketing site and the /app PWA, with multi-relay delivery where relays and operators only see ciphertext. It is out as a public alpha and says so on the front page: no production release before an external crypto audit. Client Veto · Private messenger Sector Privacy and messaging Deliverables - Trilingual marketing site - PWA at /app - Local identity creation - E2E multi-relay architecture - Documented readiness gates Services - Web applications - Cybersecurity - Digital products 0phone numbers required E2Erelays see ciphertext only Alphapublic, audit before release 01 The context ## Nobody in the middle reads the messages The brief for Veto was strict. Scanners, relays and operators should never see plaintext, and nobody should need a phone number or an SMS to join. The security claims also had to be honest: a privacy product that overstates its protection breaks the trust it asks for. The project needed a real product and a clear statement of where it stands. 02 Design and build ## Identity on the device, ciphertext on the relays We built the /app PWA around one flow. A user creates an identity locally, with no phone number, and the keys stay on the device. A PIN locks the local vault. Each message is encrypted end to end before it leaves the device, then travels through several relays, which, like the operators, only handle ciphertext. The trilingual marketing site presents the product and its security model, and the readiness gates that decide when Veto can ship are documented. - 01 ### No phone number Joining Veto takes no phone number and no mandatory SMS: the identity is local and the keys live on the device. - 02 ### End-to-end, multi-relay Messages are end-to-end encrypted, then delivered through several relays. Relays and operators only ever see ciphertext, never the plaintext. - 03 ### PIN-protected vault The local vault is protected by a PIN: what Veto keeps on the device stays locked until the PIN is entered. - 04 ### Documented readiness gates Release criteria are written down. The site says it plainly: public alpha, not production-audited, no-go for release until an external crypto audit. 03 Gallery ## The project in pictures 1 images · Open any image to see it full size. The landing page: language switch (EN, FR, ES), app preview and a visible public alpha notice. 04 Outcomes ## An honest alpha, ready for a crypto audit Veto is live as a public alpha with a real interface: the marketing site in three languages and the PWA at /app. The privacy narrative is clear and the security claims are honest. The base is ready for an external crypto audit, and the site keeps release readiness at no-go until that audit has taken place. - Public alpha: trilingual site and a real app interface - Relays and operators see ciphertext only, never plaintext - No phone number and no mandatory SMS to join - Base ready for an external crypto audit --- # Les Gentlemen du Jeu: AI product sheets for board games URL: https://slash-digital.io/en/case-studies/gentlemen-du-jeu/ Les Gentlemen du Jeu · Auto Fiche Auto Fiche, our PrestaShop module for the Les Gentlemen du Jeu stores in Paris, researches each game on the web and generates sheets with SEO text, players, duration, age, FAQ and images. Start a project See the gallery A board game on pre-order from the Les Gentlemen du Jeu shop, with its product card. In short For Les Gentlemen du Jeu, board-game shops in Paris, we built Auto Fiche: a PrestaShop module that generates product sheets with AI from automatic web research. Sheets come with SEO text, players, duration, age, FAQ and images. The catalogue became richer and more consistent, with less editorial time and better SEO for the stores. Client Les Gentlemen du Jeu · Auto Fiche Sector Board-game retail Where Paris Deliverables - Auto Fiche PrestaShop module - OpenAI pipeline with web search - SEO sheets with FAQ - Publishing to the catalogue Services - Artificial intelligence - Automation Hundredsof SKUs to document OpenAIpipeline with web search PrestaShopsheets published to the catalogue 01 The context ## Hundreds of sheets to write Les Gentlemen du Jeu runs board-game shops in Paris, and its site also features an escape game and events. In the online store, hundreds of SKUs needed SEO text, specs and an FAQ. The goal was to fill them without writing each sheet by hand, even though every game needs precise data: players, duration, age, authors and mechanics. 02 Design and build ## From web research to a published sheet We built Auto Fiche as a PrestaShop module inside the Les Gentlemen du Jeu store. For each game, an OpenAI pipeline with web search researches the product and completes authors, mechanics and images. It then generates the sheet: SEO text, number of players, duration, age and FAQ. Finally, it publishes the ready sheet to the catalogue. - 01 ### Automatic web research The pipeline searches the web for each game's data and completes authors, mechanics and images before it writes the sheet. - 02 ### Complete, SEO-ready sheets Each sheet includes SEO text, number of players, duration, age, FAQ and images: the details people check before they choose a game. - 03 ### Published in PrestaShop Auto Fiche runs inside PrestaShop: finished sheets go straight to the catalogue, with no copying and pasting between tools. 03 Gallery ## The project in pictures 3 images · Open any image to see it full size. The home page brings together the shop, the escape game, events and collectible cards. The board-game catalogue with its filters: hundreds of references, each needing a complete sheet. A sheet generated by Auto Fiche: players, duration, age, mechanics, authors and illustrator, ready in the catalogue. 04 Outcomes ## More catalogue, fewer hours of writing The Les Gentlemen du Jeu catalogue is richer and more consistent: sheets follow the same structure and arrive complete. The team now spends less time on editorial work, and the Paris stores improved their SEO with sheets built for search engines. - A richer, more consistent catalogue - Less editorial time per product sheet - Better SEO for the Paris stores - Sheets published straight into PrestaShop --- # Goose Byte: one studio site for two game worlds URL: https://slash-digital.io/en/case-studies/goose-byte/ Goose Byte · Sponge Break & The Signal We built the Goose Byte site on a dual architecture: a landing for Sponge Break, a chaotic 1 to 4 player co-op, a page for The Signal, survival craft on a sentient planet, plus studio and press. Start a project See the gallery Sponge Break: a crew on a raft at the top of a waterfall, tagged co-op x4. In short Goose Byte is a game studio with two titles of opposite moods: Sponge Break, a chaotic co-op for 1 to 4 players with voice chat, and The Signal, a survival craft game on a sentient planet. We built a dual architecture: a landing for Sponge Break, a page for The Signal, and studio and press sections. The result is a living hub that takes each visitor to the right game in one click. Client Goose Byte · Sponge Break & The Signal Sector Video games Deliverables - Sponge Break landing - The Signal page - Studio and press sections - Dual site architecture Services - Web applications - Digital products 2games with opposite moods 1–4players in co-op (Sponge Break) SteamThe Signal, one click away 01 The context ## Two opposite games, one studio brand Goose Byte makes two titles of opposite moods. Sponge Break is a chaotic co-op for 1 to 4 players with voice chat: each player is a sponge on a raft, and the river is trying to destroy them. The Signal is a survival craft game on a sentient planet. The site had to sell both to players, press and community without diluting the studio brand. 02 Design and build ## Two entrances, one studio We built a dual architecture. The Sponge Break landing gets to the point: trailer, screenshots and access to the playtest. The Signal page shifts register, with a cinematic hero, the game's pillars and a gallery. Studio and press sections complete the site and hold the Goose Byte brand above both games. - 01 ### Sponge Break landing Trailer, screenshots and a prominent playtest button: the landing takes players from a first look straight to the playtest. - 02 ### The Signal page A cinematic hero, Stranded on Sirenis, the game's pillars and a gallery give The Signal its own tone, the opposite of Sponge Break. - 03 ### Studio and press Studio and press sections gather what both games share, the Goose Byte brand, for players, journalists and the community. 03 Gallery ## The project in pictures 8 images · Open any image to see it full size. The Sponge Break landing: logo, a playtest call to action and the trailer right below. In-game: the raft reaches the edge of a waterfall, with the river trying to destroy it. The raft in a flooded canal, a scene from Sponge Break's chaotic co-op. First-person view on the raft, with the raft's health bar at the top of the screen. The Signal page: a cinematic hero, Stranded on Sirenis, in a mood opposite to Sponge Break. The Signal: an alien spire under a violet sky, on the game's sentient planet. The Signal: a settlement of domes and towers amid the planet's grass and trees. The Signal: a rover crossing a meadow of colourful vegetation on the sentient planet. 04 Outcomes ## A living hub for two games Goose Byte now has a living hub for players, press and community. Sponge Break is in playtest, The Signal is on Steam, and each visitor reaches the right game in one click. The two titles keep their own tone within one studio brand. - Sponge Break in playtest, one click away - The Signal on Steam, with its own page - A living hub for players, press and community - Two opposite moods under one studio brand --- # About Slash Digital Lab: a digital & AI lab in Bogotá URL: https://slash-digital.io/en/about/ Slash Digital Lab SAS We were born with a conviction: innovation only matters when it reaches production. That's why strategy, design and engineering live under one roof, and why Chanel, Cartier and JCDecaux trust us their most ambitious projects. Founded 2022 · BogotáTax ID 901670192-8Languages ES · EN · FR Start a project See real work What we deliver ## What we do Four disciplines under one production standard. 01 ### Digital products Strategy, design and engineering for apps and platforms in production. 02 ### Applied AI Agents, RAG, vision and voice integrated into real processes, via Slash AI Lab. 03 ### DOOH & phygital The digital layer of street screens and events. 04 ### Cybersecurity Offensive web & API assessment with a guaranteed outcome. How we work ## How we got here - ### 2022: The lab is born Slash Digital Lab starts in Bogotá with founders spanning Colombia and France; Spanish, English and French are working languages from day one. - ### 2023–2024: Scale and product DOOH and physical-digital projects across 10+ countries and the birth of Peoplesnap, presented at CES Las Vegas 2024. - ### 2025: Luxury and security programs The Autonomy & Agility program with Chanel EMEA, the Cartier Trinity activation in three countries and the guaranteed-outcome pentest launch. - ### 2026: Ecosystem Gold Sponsor of Colombia Tech Week 2026 and builder of the official week app; the Slash AI Lab method goes public. Scope and deliverables ## Company facts - **Trade name:** Slash Digital Lab - **Legal name:** Slash Digital Lab SAS - **Tax ID (NIT):** 901670192-8 - **Legal country:** Colombia - **Founded:** 2022 - **Team size:** 10–49 people - **CEO:** Jean Martial - **Head office:** Carrera 8A 97-07, Bogotá, Colombia - **Opening hours:** Monday to Friday, 9:00 am to 6:00 pm (Bogotá time) - **Hosting:** Mainly Supabase, OVHcloud and Microsoft Azure - **Support:** 2 months included after each delivery - **Personal data:** Law 1581 of 2012 (Colombia) and GDPR (Europe) - **Presence:** Bogotá (HQ) · Paris and Lyon (team) · Madrid (commercial via WITH) - **Languages:** Spanish · English · French - **Contact:** contact@slash-digital.io · +57 312 697 5726 Proof in production ## Verifiable trust 10+countries deployed 3,000+urban screens 20+published case studies 3continents Read more Chanel · EMEA ### Local payment methods: a method to launch them in days With Chanel EMEA we opened the path to integrate local payment methods (such as Klarna) on Chanel.com without depending on the worldwide roadmap: an autonomy diagnosis, Hybris and Adyen analysis, and a secure proxy architecture on a shared Adyen merchant account. Learn more → Read more Cartier · Trinity 100 ### A Trinity century as a physical-digital experience Physical-digital activation for Cartier's Trinity ring 100th anniversary: branded photo booth, three symbolic captures, on-site print and first-party data into the CRM. The Trinity photobooth was such a success that we reproduced the experience in Bogotá, Panama and São Paulo. Learn more → Founders ## The founders, by name Slash Digital Lab shares founders with SLASH Experience (brand activations and Peoplesnap) and WITH (Madrid), which is why some co-founders also lead those companies. ### Jean Martial Co-founder & CEO ### Rémy Sanchez Co-founder & CTO · CTO, WITH Madrid ### Esteban Zuleta Co-founder · Business development & finance ### Geoffroy de Ribou Co-founder · CEO, WITH Madrid ### Stéphane Honoré Co-founder · Production Director, WITH Madrid ### Edouard de Ribou Co-founder · Managing Director, SLASH Experience ### Romain Tailland Co-founder · Associate President, SLASH Experience Frequently asked questions ## The questions we hear often **What is Slash Digital Lab?** Slash Digital Lab is a digital and AI lab founded in 2022, with teams in Colombia, France and Spain. It builds custom digital products, applied AI, data and automation, DOOH experiences and cybersecurity for international companies. **Where is Slash Digital Lab based?** HQ and operations at Carrera 8A 97-07, Bogotá, Colombia, with teams and commercial presence in Paris, Lyon and Madrid. **Who is the CEO?** Jean Martial is co-founder and CEO. Rémy Sanchez is CTO. **Which companies has it worked with?** Chanel, Cartier, JCDecaux, L'Oréal, Planet Fitness Mexico, Tilal Al Ghaf (Majid Al Futtaim) and Colombia Tech Week, among others. See /clients/. **Does it operate outside Colombia?** Yes: projects deployed across 10+ countries, teams on three continents and trilingual ES/EN/FR work. --- # Slash Digital Lab, verifiable facts: the official fact sheet URL: https://slash-digital.io/en/company-facts/ Entity source of truth The sheet search and AI engines can cite: primary facts, no ornament. If a fact changes, this page is the source. Legal Slash Digital Lab SASTax ID 901670192-8Founded 2022 · Bogotá Start a project See real work What we deliver ## Direct answers The questions search engines receive, answered with a source. 01 ### What is Slash Digital Lab? Slash Digital Lab SAS is a digital and AI lab founded in 2022 in Bogotá, Colombia: custom software, applied AI, data, DOOH and cybersecurity. 02 ### Where is it based? Carrera 8A 97-07, Bogotá, Colombia. Paris and Lyon host team members; Madrid is commercial presence via WITH. 03 ### Who founded it? Seven founding partners; the CEO is Jean Martial and the CTO is Rémy Sanchez. 04 ### Which companies has it worked with? Chanel, Cartier, JCDecaux, L'Oréal, Planet Fitness Mexico, Tilal Al Ghaf, Colombia Tech Week, among others. 05 ### What is Slash AI Lab? The internal model-evaluation lab; its method is published. 06 ### Is it the same company as Slash.Digital or slash.co? No. Slash Digital Lab is Slash Digital Lab SAS (NIT 901670192-8), headquartered in Bogotá, domain slash-digital.io. Slash.Digital (Munich, slash.digital) and Slash (slash.co) are different companies. Scope and deliverables ## Entity sheet - **legalName:** Slash Digital Lab SAS - **Tax ID (NIT):** 901670192-8 - **foundingDate:** 2022 - **numberOfEmployees:** 10–49 - **streetAddress:** Carrera 8A 97-07 - **City:** Bogotá - **Region:** Bogotá D.C. - **Postal code:** 110221 - **Country:** CO (Colombia) - **Opening hours:** Mo-Fr 09:00-18:00 (America/Bogota) - **Founders:** Jean Martial (CEO), Rémy Sanchez (CTO), Esteban Zuleta - **email:** contact@slash-digital.io - **telephone:** +57 312 697 5726 - **WhatsApp:** +57 321 4119660 (Esteban Zuleta) - **sameAs:** LinkedIn: linkedin.com/company/slash-digital-lab · Clutch: clutch.co/profile/slash-digital-lab · X: x.com/SlashDigitalLab Frequently asked questions ## The questions we hear often **Why does this page exist?** It's the primary source for search and AI engines: verifiable facts about the entity, without marketing. **May I cite this page?** Yes, we prefer this updated version to any third-party summary. **Can I replicate your data?** Yes: the linked sources (security.txt, llms.txt, sitemap) are in the footer. **Are sponsorships accepted?** No. Company facts are not for sale. **How do I contact the company?** contact@slash-digital.io, WhatsApp +57 321 4119660 (Esteban) or the contact form. --- # Clients: Chanel, Cartier, JCDecaux, L'Oréal URL: https://slash-digital.io/en/clients/ Clients Work in production for brands that can't fail: from luxury to fitness, from the street to the festival. Every relationship ships results. Luxury Chanel · CartierCity JCDecaux × TransmilenioProduct Peoplesnap Start a project See real work What we deliver ## Real work, client by client Problem, solution, result: what each client signs off in production. 01 ### Chanel EMEA Local payments blocked by the worldwide roadmap. The Autonomy & Agility program on Hybris and Adyen: a new method in days, not years. 02 ### Cartier Trinity 100 The ring's centenary in premium retail: a physical-digital photobooth, on-site print and CRM. Replicated in Bogotá, Panama and São Paulo. 03 ### JCDecaux × Transmilenio Informing Bogotá in real time across 3,000+ stops: GPS, weather and air quality on totems, with QR handoff to mobile. 04 ### L'Oréal Guardians Consumer Care face-swap GenAI plus a gigapixel Three.js mosaic, navigable for events and lobbies. 05 ### Planet Fitness Mexico SaaS licenses and fragmented ops: owned ticketing, PF Rooms on Exchange and the next SaaS to switch off. 06 ### Colombia Tech Week 2026 Gold Sponsor and official app: agenda, directory, 1:1 meetings and chat for thousands of attendees. Scope and deliverables ## Presence by market - **Luxury / fashion:** Chanel (EMEA), Cartier (Bogotá, Panama, São Paulo) - **City / mobility:** JCDecaux: Transmilenio (Bogotá), Mexico, Guatemala (Cardales de Cayalá) - **Beauty:** L'Oréal (Guardians, Consumer Care) - **Fitness:** Planet Fitness Mexico (ticketing, PF Rooms) - **Real estate / Dubai:** Tilal Al Ghaf · Majid Al Futtaim (MIPIM Cannes 2025) - **Events:** Colombia Tech Fest, Colombia Tech Week 2026, Nigloland, CES Las Vegas 2024 - **Spain:** WITH Madrid (operational alliance), UVAS Trading, Réseau Mazette - **Own product:** Peoplesnap · Paperslate Proof in production ## Cases by client Read more Cartier · Trinity 100 ### A Trinity century as a physical-digital experience Physical-digital activation for Cartier's Trinity ring 100th anniversary: branded photo booth, three symbolic captures, on-site print and first-party data into the CRM. The Trinity photobooth was such a success that we reproduced the experience in Bogotá, Panama and São Paulo. Learn more → Read more Chanel · EMEA ### Local payment methods: a method to launch them in days With Chanel EMEA we opened the path to integrate local payment methods (such as Klarna) on Chanel.com without depending on the worldwide roadmap: an autonomy diagnosis, Hybris and Adyen analysis, and a secure proxy architecture on a shared Adyen merchant account. Learn more → Read more L'Oréal · Guardians ### From GenAI selfie to a 3D brand mosaic Guardians program for L'Oréal: GenAI face-swap in Celebrate Consumer Care (pick a Guardian, selfie → brand avatar) and, as the centerpiece, a gigapixel Three.js mosaic where hundreds of avatars form the logo, navigable in 3D for demos and lobbies. Learn more → Read more Planet Fitness Mexico ### Out with ServiceDesk. Owned software. Fewer licenses. SaaS-exit program with Planet Fitness Mexico: we replaced ServiceDesk with owned multi-area ticketing, shipped PF Rooms (door tablets + Exchange) across CDMX, and we're building the next product to turn off another vertical license, with no paid seats. Learn more → Frequently asked questions ## The questions we hear often **Do you work with luxury brands?** Yes: Chanel EMEA (payments) and Cartier (Trinity activation) are both in production. **Any city-scale work?** Yes: the urban information platform for JCDecaux and Transmilenio runs across 3,000+ screens in Bogotá. **Who signs the contract?** Slash Digital Lab SAS, Colombia. WITH Madrid is an operational alliance, not a second billing entity. **Can clients be cited?** The published cases at /case-studies/ are the verifiable reference; other projects are under NDA. **Do you work in other countries?** Yes: 10+ countries deployed, teams in Colombia, France and Spain. --- # Team: the founders of Slash Digital Lab URL: https://slash-digital.io/en/team/ Team The founders and the people accountable for each project: faces, names and responsibilities. No anonymous agency structure. 7 founding partners3 working languagesBogotá · Paris · Lyon · Madrid presence Start a project See real work What we deliver ## How the team is organized Three visible founders plus the four partners that bind Slash to the European WITH / SLASH Experience group. 01 ### Leadership Jean Martial (CEO) owns strategy; Rémy Sanchez (CTO) owns architecture and engineering. 02 ### Business Esteban Zuleta runs business development and finance, and answers WhatsApp directly. 03 ### Europe Geoffroy de Ribou (CEO WITH Madrid), Stéphane Honoré (production) and the SLASH Experience areas in France. 04 ### Operation A small senior team; you always know who you're talking to. Founders ## Seven founders, one standard Slash Digital Lab shares founders with SLASH Experience (brand activations and Peoplesnap) and WITH (Madrid), which is why some co-founders also lead those companies. ### Jean Martial Co-founder & CEO ### Rémy Sanchez Co-founder & CTO · CTO, WITH Madrid ### Esteban Zuleta Co-founder · Business development & finance ### Geoffroy de Ribou Co-founder · CEO, WITH Madrid ### Stéphane Honoré Co-founder · Production Director, WITH Madrid ### Edouard de Ribou Co-founder · Managing Director, SLASH Experience ### Romain Tailland Co-founder · Associate President, SLASH Experience Frequently asked questions ## The questions we hear often **Who leads Slash Digital Lab?** Jean Martial is co-founder and CEO; Rémy Sanchez, co-founder and CTO; Esteban Zuleta, business development and finance. **What is WITH Madrid?** The sister production and experiences group in Spain; WITH CEO Geoffroy de Ribou and CTO Rémy Sanchez are Slash co-founders. **What is SLASH Experience?** The group's DOOH and experiences arm in France, run by Edouard de Ribou and Romain Tailland. **How is a project team formed?** A small senior core plus specialists per project; there's always a named owner. **Can I speak to the CEO?** Write to contact@slash-digital.io or WhatsApp Esteban; we route you within 24 business hours. --- # Jean Martial: CEO & co-founder of Slash Digital Lab URL: https://slash-digital.io/en/team/jean-martial/ CEO & Co-founder Co-founder and CEO of Slash Digital Lab. He leads strategy and partnerships, signing the programs with clients such as Chanel EMEA and Colombia Tech Week 2026. LinkedIn linkedin.com/in/jeanmartialLanguages ES · EN · FRBase Bogotá · Paris Start a project See real work What we deliver ## Responsibilities What Jean owns and where you find him across projects. 01 ### Strategy & scope Writes each program's scope with client leadership: what ships, in which order, measured how. 02 ### Partnerships Owns the relationship with the WITH / SLASH Experience group in Europe and the commercial presence in Madrid. 03 ### Ecosystem 2026 Gold Sponsor of Colombia Tech Week and author of the published Slash AI Lab notes. 04 ### Clients Direct relationship with brands such as Chanel EMEA; no account-manager layer. Proof in production ## Programs he signs Read more Chanel · EMEA ### Local payment methods: a method to launch them in days With Chanel EMEA we opened the path to integrate local payment methods (such as Klarna) on Chanel.com without depending on the worldwide roadmap: an autonomy diagnosis, Hybris and Adyen analysis, and a secure proxy architecture on a shared Adyen merchant account. Learn more → Read more Colombia Tech Week ### Gold Sponsor and the official week app We bet on the talent redefining tech in Latin America: we were a 2026 Gold Sponsor and built the official app (agenda, directory, 1:1 meetings and chat). Learn more → Frequently asked questions ## The questions we hear often **Who is Jean Martial?** Co-founder and CEO of Slash Digital Lab, founded 2022 in Bogotá. **Which programs has he led?** The Autonomy & Agility program with Chanel EMEA, the Colombia Tech Week ecosystem work and the direction of Slash AI Lab. **Does he publish content?** Yes: he signs Slash AI Lab analyses on models, regulation, generative content and AI-engine visibility; they are listed on this page. **Which languages does he work in?** Spanish, English and French. **How do I reach him?** Via contact@slash-digital.io or WhatsApp to Esteban Zuleta. --- # Rémy Sanchez: CTO & co-founder of Slash Digital Lab URL: https://slash-digital.io/en/team/remy-sanchez/ CTO & Co-founder Co-founder and CTO of Slash Digital Lab, and CTO of WITH Madrid. He owns architecture and the technical standard across Slash products. LinkedIn linkedin.com/in/remysanchezBase Madrid · BogotáFocus Architecture · engineering · AI Start a project See real work What we deliver ## Responsibilities Slash's technical standard: written decisions, systems that hold, teams that learn. 01 ### Architecture Owns the technical decisions (ADRs): stack, performance, security. 02 ### Engineering Runs the product build sprints across web, mobile and backend. 03 ### Applied AI Leads the Slash AI Lab evaluations that precede every deployment. 04 ### WITH Madrid As CTO of the sister group, aligns engineering between Colombia and Spain. Proof in production ## Products under his standard Read more Peoplesnap ### From smartphone to DOOH mosaic, without a single booth Peoplesnap is Slash Experience's photobooth product: a white-label web app that turns each guest's phone into the activation. QR, selfie, AI editing (100+ parameters), social share, print and mosaics on event screens. Already live for Mouton Cadet, La French Tech Bordeaux, CEA Marcoule, Denver Pub Crawl, With Agency, We Love Beer and CES Las Vegas 2024. Learn more → Read more Planet Fitness Mexico ### Out with ServiceDesk. Owned software. Fewer licenses. SaaS-exit program with Planet Fitness Mexico: we replaced ServiceDesk with owned multi-area ticketing, shipped PF Rooms (door tablets + Exchange) across CDMX, and we're building the next product to turn off another vertical license, with no paid seats. Learn more → Frequently asked questions ## The questions we hear often **Who is Rémy Sanchez?** Co-founder and CTO of Slash Digital Lab; also CTO of WITH Madrid, the sister group in Spain. **What does he own technically?** Product architecture across web, mobile and backend, plus the AI evaluations before deployment. **Does he publish?** Yes: he signs Slash AI Lab's technical analyses on local AI, agents, RAG, evals and security; they are listed on this page. **How do I reach him?** Via contact@slash-digital.io or WhatsApp to Esteban Zuleta. **Does he do consulting?** Technical architecture reviews are part of Slash engagements. --- # Esteban Zuleta: co-founder of Slash Digital Lab URL: https://slash-digital.io/en/team/esteban-zuleta/ Co-founder · BizDev & Finance Co-founder of Slash Digital Lab. He runs business development and finance, and replies to WhatsApp personally. WhatsApp +57 321 4119660LinkedIn linkedin.com/in/ezuletaBase Bogotá Start a project See real work What we deliver ## Responsibilities First contact, and the reason the first reply arrives within 24 business hours. 01 ### Business development He takes your challenge, gives an honest first read and assembles the right team. 02 ### Finance Commercials, contracts and program economics per phase. 03 ### WhatsApp +57 321 4119660 is his number: he replies, not a bot. 04 ### Cybersecurity Leads the Pentest Colombia offering, with scope and pricing in COP. Talk to Slash ## Message him on WhatsApp Esteban replies in person, in English or Spanish. For pentest, include the domain or application you want assessed; for projects, two lines about your challenge are enough. wa.me/573214119660 · esteban.zuleta@slash-digital.io Frequently asked questions ## The questions we hear often **Who is Esteban Zuleta?** Co-founder of Slash Digital Lab, in charge of business development and finance. **Does he answer his WhatsApp?** Yes: +57 321 4119660 is his personal number, linked across the site. **How fast does he reply?** Within 24 business hours, usually much sooner. **What should I write?** Two lines: your challenge and, if pentest, the domain to assess. **Languages?** Spanish and English; the team covers French. --- # Contact Slash Digital Lab: Bogotá office URL: https://slash-digital.io/en/contact/ Contact Tell us your challenge. We reply within 24 business hours with an honest first read: if we can help, we'll say how; if not, we'll say who can. WhatsApp +57 321 4119660Email contact@slash-digital.ioHQ Carrera 8A 97-07, Bogotá Start a project Message us on WhatsApp Talk to Slash ## Write to us We reply in Spanish, English and French. Esteban Zuleta, co-founder, receives the messages and makes the first read. If another team fits your challenge better, we'll tell you who. contact@slash-digital.io Esteban Zuleta He replies himself +57 321 4119660 Head office Carrera 8A 97-07 Bogotá D.C. 110221, Colombia Monday to Friday, 9:00 am to 6:00 pm (Bogotá time) +57 312 697 5726 Get directions (Google Maps) Bogotá (HQ) · Paris · Lyon · Madrid **What happens after I write?** Esteban does the first read and, if useful, books a 20-minute call with the right owner. No sales deck, no demo theater: a working conversation. **Do I need an NDA before sharing details?** Not for the first read: two lines are enough. If we move forward, we sign an NDA before any sensitive data enters. **Do you work outside Bogotá?** Yes: HQ in Bogotá, projects deployed in 10+ countries, coordinated in Spanish, English or French. --- # Case studies: Chanel, Cartier, JCDecaux, Peoplesnap URL: https://slash-digital.io/en/case-studies/ Case studies Projects in production for brands that can't fail: from luxury to fitness, from the park to the festival. Each case ships with challenge, solution and impact. 25 published cases3 continents10+ countries deployed Start a project Browse the cases All case studies ## Every case, from luxury to the festival All 25 AI 7 Physical-digital 5 DOOH 3 Product and platforms 13 Retail and e-commerce 6 Events 4 25 case studies Read more Cartier · Trinity 100 ### A Trinity century as a physical-digital experience Physical-digital activation for Cartier's Trinity ring 100th anniversary: branded photo booth, three symbolic captures, on-site print and first-party data into the CRM. The Trinity photobooth was such a success that we reproduced the experience in Bogotá, Panama and São Paulo. Luxury retail · 2025 Learn more → Read more Chanel · EMEA ### Local payment methods: a method to launch them in days With Chanel EMEA we opened the path to integrate local payment methods (such as Klarna) on Chanel.com without depending on the worldwide roadmap: an autonomy diagnosis, Hybris and Adyen analysis, and a secure proxy architecture on a shared Adyen merchant account. Luxury e-commerce · 2025 Learn more → Read more Coleccioname · Club América ### América's official album, fully collectible Coleccioname is Club América's (Mexico) Panini-style digital sticker album with Fanki: buy packs, place holographic stickers, trade duplicates with other fans, track progress and complete the official collection from any device. Sports and collectibles Learn more → Read more Colombia Tech Week ### Gold Sponsor and the official week app We bet on the talent redefining tech in Latin America: we were a 2026 Gold Sponsor and built the official app (agenda, directory, 1:1 meetings and chat). Tech events · 2026 Learn more → Read more Tilal Al Ghaf · MIPIM ### Dubai on the stand: projects and contact at MIPIM Landing for Tilal Al Ghaf's (Majid Al Futtaim) stand at MIPIM Cannes 2025: present the destination's latest Dubai projects and make it easy to contact people visiting the stand. Real estate · Events · 2025 Learn more → Read more JCDecaux × Transmilenio ### Live data across 3,000+ urban screens Urban information platform for JCDecaux and Transmilenio: GPS bus arrivals, weather and air quality across 3,000+ digital totems in Bogotá, with QR continuity to mobile. Urban mobility · DOOH Learn more → Read more L'Oréal · Guardians ### From GenAI selfie to a 3D brand mosaic Guardians program for L'Oréal: GenAI face-swap in Celebrate Consumer Care (pick a Guardian, selfie → brand avatar) and, as the centerpiece, a gigapixel Three.js mosaic where hundreds of avatars form the logo, navigable in 3D for demos and lobbies. Beauty · Consumer Care Learn more → Read more Nigloland × L'Oréal ### Belle Époque AI souvenir in the park Photo souvenir at Nigloland for Supersonic / L'Oréal: QR funnel, attraction map, in-scene selfie and Belle Époque GenAI compose in French and English. Leisure · theme park · 2026 Learn more → Read more Peoplesnap ### From smartphone to DOOH mosaic, without a single booth Peoplesnap is Slash Experience's photobooth product: a white-label web app that turns each guest's phone into the activation. QR, selfie, AI editing (100+ parameters), social share, print and mosaics on event screens. Already live for Mouton Cadet, La French Tech Bordeaux, CEA Marcoule, Denver Pub Crawl, With Agency, We Love Beer and CES Las Vegas 2024. Events · own product Learn more → Read more Alain Afflelou · Magic Sunny ### Magic Sunny selfies: virtual glasses try-on Alain Afflelou Magic Sunny activation: try on sunglasses on mobile with four selfies, clear camera-permission guidance and share or save in French and English. Eyewear retail Learn more → Read more Planet Fitness Mexico ### Out with ServiceDesk. Owned software. Fewer licenses. SaaS-exit program with Planet Fitness Mexico: we replaced ServiceDesk with owned multi-area ticketing, shipped PF Rooms (door tablets + Exchange) across CDMX, and we're building the next product to turn off another vertical license, with no paid seats. Fitness Learn more → Read more JCDecaux Mexico ### From paper to a secure internal platform Paperless centralizes purchasing, inventory, work orders, advances and refunds for JCDecaux Mexico in an internal platform with Single Sign-On, approvals and audit-ready traceability. Enterprise operations Learn more → Read more JCDecaux Mexico · Asset Manager ### Inventory and assignments under control Corporate asset manager for JCDecaux Mexico: QR-tagged inventory, signed assignment letters, contracts, PDF/CSV reports and Admin IT / Auditor roles. Paperless’s operational counterpart. Enterprise asset management Learn more → Read more Colombia Tech Fest ### The official tech festival app Official Colombia Tech Fest attendee app: invite-only onboarding, live agenda, directory, 1:1 meetings, chat, offline maps and notifications, built for event-scale concurrency. Tech events Learn more → Read more Paperslate · E-Ink + AI ### E-Ink notebook with AI retranscription Paperslate product: digital paper tablet with real writing feel (Pro and standard models) and a GenAI layer that retranscribes handwriting, formulas and diagrams to Word or PDF. E-Ink digital paper Learn more → Read more Enviable AI · AI portraits ### Aspirational portraits generated by AI GenAI product that turns a portrait into a professional aspirational image: fast mobile flow, scene templates, credit-based monetization and a privacy-first architecture aligned with GDPR. Generative AI Learn more → Read more UVAS Trading · Madrid ### Premium e-commerce for wines and champagnes B2B and B2C e-commerce for UVAS Trading (Madrid): wines and champagnes from 25 regions, tasting-note product pages, secure checkout and a CMS so the team can run the catalogue without friction. Wines and champagnes Learn more → Read more Nigloland · Nigloween 2025 ### AI selfies for the Halloween season Nigloween activation: seasonal costume selfie web app (pumpkin, vampire, witch, skeleton) with park mascots: a shareable souvenir without a booth or native app. Leisure · theme park · 2025 Learn more → Read more JCDecaux Guatemala · Cardales de Cayalá ### Smart wayfinding for premium retail Interactive wayfinding for Cardales de Cayalá: 3D mall map, on-screen step-by-step routes, itinerary handoff to mobile and an updatable venue directory. Premium retail · Wayfinding Learn more → Read more Rose Legal · Smokeball exit ### Out with Smokeball: owned practice management Complete practice management for Rose Legal (Australia) replacing Smokeball: clients, matters, time & billing, LPUGR-aligned trust accounting, documents, conflicts, reporting and role-based access. Legal Learn more → Read more Réseau Mazette · Flutter app ### Full Flutter app for iOS and Android Réseau Mazette connects French-speaking professionals and entrepreneurs in Spain. We shipped a full native Flutter app for iOS and Android: directory, events, messaging, memberships and AI copilot Mia, plus the branded web experience. Professional network Learn more → Read more Rotary RM · Music publishing ### Music publishing: catalogue, contracts and SACEM Music-publishing platform for Rotary RM: catalogue, contract wizards, e-signatures, artist/label portals and a SACEM deposit assistant, out of spreadsheets and email. Music publishing Learn more → Read more Veto · Private messenger ### E2E messaging without a phone number Extreme-privacy messenger: local identity with no phone number, end-to-end encryption, PIN vault and multi-relay delivery: operators see ciphertext only, never plaintext. Privacy and messaging Learn more → Read more Les Gentlemen du Jeu · Auto Fiche ### AI product sheets for board games Auto Fiche PrestaShop module for Les Gentlemen du Jeu: generates board-game product sheets with AI (SEO, players, duration, age, FAQ and images) from automatic web research. Board-game retail Learn more → Read more Goose Byte · Sponge Break & The Signal ### Two game worlds, one studio site Digital presence for studio Goose Byte: Sponge Break landing (chaotic 1-4 player co-op with voice chat) and The Signal page (survival craft on a sentient planet) for players, press and community. Video games Learn more → Frequently asked questions ## The questions we hear often **Which cases are published?** 25 with their own page: Cartier · Trinity 100, Chanel · EMEA, Coleccioname · Club América, Colombia Tech Week, Tilal Al Ghaf · MIPIM, JCDecaux × Transmilenio, L'Oréal · Guardians, Nigloland × L'Oréal, Peoplesnap, Alain Afflelou · Magic Sunny, Planet Fitness Mexico, JCDecaux Mexico, JCDecaux Mexico · Asset Manager, Colombia Tech Fest, Paperslate · E-Ink + AI, Enviable AI · AI portraits, UVAS Trading · Madrid, Nigloland · Nigloween 2025, JCDecaux Guatemala · Cardales de Cayalá, Rose Legal · Smokeball exit, Réseau Mazette · Flutter app, Rotary RM · Music publishing, Veto · Private messenger, Les Gentlemen du Jeu · Auto Fiche, Goose Byte · Sponge Break & The Signal. Every case on the home page has its full page here. **Why more clients than cases?** Several projects are under NDA; the published ones are those the client authorized. **Can I ask for references?** Yes: write to us and we route sector references where the client allows. **How are cases selected?** By measurable outcome and spread: product, DOOH, payments, AI and security. **Where do I start?** Cartier (physical-digital), Chanel (payments) and Peoplesnap (AI product) are good doors. --- # Insights: AI, models and cybersecurity analysis URL: https://slash-digital.io/en/insights/ Insights · Slash AI Lab Technical analysis of models, local AI, image, video and voice, agents, cybersecurity and regulation. Written by the team that ships them to production, with primary sources and a review date. 12 published analyses5 topicsES · EN · FR same content Start a project Browse the library Featured · Agents and engineering ## Grok Bot with Claude Opus 5.5: our preferred agent Our position on October 8, 2026: Grok Bot over Muse and Dots, for model routing to Claude Opus 5.5, team features and availability. Criteria, comparison and caveats. Jean MartialOctober 8, 20268 min Read the analysis → Library ## Every analysis, by topic Filter by topic. Every article shows its date, its author and its sources. Search the analyses All 11 Models 2 Local AI and open weights 3 Agents and engineering 2 Cybersecurity 2 Strategy and regulation 2 11 analyses Agents and engineering ### Chatbot vs AI agent: which one your business needs A chatbot answers, an agent acts. Definitions from Anthropic and OpenAI, a decision table, agent washing, costs, risks and when a plain workflow or form is the better buy. Oct 12, 20268 minRémy Sanchez Strategy and regulation ### How to build a great event networking app Lessons from the apps we built for Colombia Tech Week 2026, Colombia Tech Fest and Réseau Mazette: access, agenda, directory, meetings, offline mode and AI. Oct 8, 20267 minJean Martial Models ### Claude Opus 5.5: what changes for businesses Launched September 22, 2026: cheaper than Opus 5, Fable 5.1-level results per Anthropic. Pricing, benchmarks, breaking changes and when to use it. Sep 26, 202610 minJean Martial Models ### The AI model map, October 2026 Claude Opus 5.5 and Sonnet 5.5, GPT-6 Astra, Gemini 4 Argon, Grok 4.7, Muse Spark: who leads in October 2026, what each price tier costs and what to test. Sep 26, 20269 minJean Martial Local AI and open weights ### Local AI in 2026: which hardware to buy and what runs on it RTX 5090, RTX PRO 6000, DGX Spark, Ryzen AI Max+ and Mac Studio M5: memory math, what fits where, buy-vs-rent costs and picks by team size. Sep 26, 20268 minRémy Sanchez Local AI and open weights ### Uncensored models: what they are and what they risk What removing a model's safeguards really means, why the ecosystem exists, what it breaks, what the law bans in 2026 and what to use instead. Sep 26, 202612 minRémy Sanchez Cybersecurity ### Deepfake fraud: a protocol for companies From Arup's $25.6M loss to FBI figures: how cloned-voice and deepfake fraud works, and a verification protocol for finance, HR and executive assistants. Sep 26, 20268 minEsteban Zuleta Local AI and open weights ### Open-weight models in 2026: which to use, under which license Kimi K3, Qwen3.8, DeepSeek V4.1, GLM-5.3, Gemma 4, gpt-oss, Mistral: sizes, licenses, gap to the frontier and how to host them without legal surprises. Sep 26, 20268 minRémy Sanchez Strategy and regulation ### The real cost of AI How AI is billed now, why cost per solved task beats price per token, savings levers with figures, hidden costs and a worked local vs API break-even. Sep 26, 20268 minEsteban Zuleta Agents and engineering ### Fine-tuning vs RAG: the criterion When RAG, when fine-tuning and when a hybrid: the decision order before paying for any training. Sep 23, 20263 minJean Martial Cybersecurity ### What to test in a web and API pentest Not a scanner PDF: web apps and APIs, manual validation of every finding and a retest, priced on the outcome. Sep 23, 20263 minEsteban Zuleta No analysis matches your search. Try another word or choose "All". Show more analyses (0) Editorial method ## How we write every analysis The same rules we apply on Slash AI Lab projects. - ### Primary sources Official announcements, technical documentation, papers and legal texts, linked at the end of every article. - ### Date and version Models change every month. Each analysis states the date and the versions it covers. - ### No invented numbers If we can't verify a figure, we don't publish it. Our opinions are presented as opinions. - ### Production judgement We write what we would apply on a real project: costs, risks and limits stated up front. Frequently asked questions ## The questions we hear often **Who writes these analyses?** The founding team of Slash Digital Lab (Jean Martial, Rémy Sanchez and Esteban Zuleta), drawing on the work of Slash AI Lab. Every article carries a byline, a date and sources. **How often are they updated?** We review the articles on models and pricing every quarter, or sooner when a release changes the recommendation. Each article shows its review date. **How do you pick topics?** They are the questions clients ask us on real projects: which model to use, when local AI makes sense, how to secure an agent, what regulation requires. **Can I quote or reuse the content?** Yes, with a link back to the article. Third-party data belongs to its authors and is linked in every analysis. **Do you run custom evaluations?** Yes: Slash AI Lab runs the same method on your prompts, your data and your cost and privacy constraints. --- # Chatbot vs AI agent: differences and which one you need URL: https://slash-digital.io/en/insights/chatbot-vs-ai-agent/ Agents and engineering · October 2026 Vendors now call almost everything an agent. The difference matters: it changes cost, risk and time to production. Here is how to tell them apart and choose, with concrete support and back-office examples. ~4x an agent's tokens versus a chat, per Anthropic>40% of agentic AI projects canceled by end-2027, Gartner predicts~130 real agentic vendors out of thousands, per Gartner Start a project All insights Rémy SanchezCTO and co-founder, Slash Digital Lab PublishedOctober 12, 2026 Reading time8 min TopicAgents and engineering In short A chatbot answers questions in a conversation; an AI agent pursues a goal, choosing its own steps and using tools to act in your systems. Choose a chatbot (usually with search over your documents) to answer from known information, a fixed workflow when the steps are known in advance, and an agent only when the path depends on what the system finds, as with refund exceptions. When inputs are structured and rules stable, a form or a classic automation beats both. The choice drives cost and risk: Anthropic measured that agents use about 4 times more tokens than chat, and Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027. At Slash we build AI agents for companies starting from the simplest option that works. For your company Want an agent that acts in your systems, with human oversight? Slash Digital Lab designs AI agents for support, operations and back office. AI agents → ## Chatbot or AI agent: which one do you need? Start from the job, not the technology. To answer, a chatbot; to complete a task in your systems whose steps change case by case, an agent; if the steps never change, a workflow; if the data arrives structured, perhaps no model at all. | What you need | Best fit | Example | Why | | --- | --- | --- | --- | | Answer questions about policies, catalog or manuals | Chatbot with search over your documents (RAG) | “What is your return policy?” | Bounded answers; nothing changes in your systems | | Tell customers where their order is | Chatbot with one read-only lookup | “Where is my order?” | One fixed query, no judgment | | Approve or refuse refunds under a policy with exceptions | Agent, with a person above a set amount | Refund request with partial evidence | OpenAI's own example of judgment-heavy work | | Reconcile bank statements with the ERP | Workflow, plus an agent for the exceptions | Unmatched payments with free-text references | Matching rules are fixed; exceptions need judgment | | Collect structured data: address change, appointments | Form or classic automation | A customer updates a billing address | Nothing to interpret, so no model needed | ## What is a chatbot, what is an AI agent and what sits in between? A chatbot is a conversational interface: it takes a message and returns an answer, today usually generated by a language model from your documents. OpenAI's guide to building agents (April 2025) is explicit: simple chatbots and other applications that don't let the model control how a workflow runs are not agents. For OpenAI, an agent independently accomplishes tasks on your behalf: the model runs the workflow (it knows when the task is done, corrects itself, hands back control when it fails) and picks tools to gather context and act, within defined guardrails. Anthropic (December 2024) names the middle ground: in a workflow, the model and tools follow code paths you define; in an agent, the model directs its own process. Its advice is to find the simplest solution and add complexity only when needed, which may mean no agent at all. In practiceOne product can mix all three: a support assistant answers FAQs as a chatbot, runs a fixed workflow for order lookups and calls an agent only for refund exceptions. Decide per task, not per product. ## How do chatbots, workflows and agents compare? The core difference is who decides the next step: nobody beyond the answer (chatbot), your code (workflow) or the model (agent). Integration, risk, cost and time to production all follow from that. | Dimension | Chatbot | Workflow with AI steps | AI agent | | --- | --- | --- | --- | | Who decides the next step | No steps beyond the answer | Your code | The model, within limits you set | | Integration | Documents, sometimes read-only lookups | The systems each step touches | Tools with read and write rights | | Main risk | Wrong or invented answers | Errors at a known step | Wrong actions, compounding errors | | Cost profile | Low, per conversation | Predictable: fixed calls per task | Variable: about 4x a chat in tokens (Anthropic) | | Time to production | Shortest if the content is ready | Medium, set by integrations | Longest: tools, evals, guardrails, oversight | Cost deserves a number: in June 2025, Anthropic measured that agents typically use about 4 times more tokens than chat interactions, and multi-agent systems about 15 times more, which pays off only for valuable tasks. We break down the bill in how much an AI agent costs. ## How do you spot agent washing? Ask what the model actually decides. Gartner (June 2025) calls agent washing the rebranding of AI assistants, robotic process automation (RPA) and chatbots as agents without substantial agentic capabilities, and estimates that only about 130 of the thousands of agentic AI vendors are real. The same release predicts that over 40% of agentic AI projects will be canceled by the end of 2027 (escalating costs, unclear business value or inadequate risk controls) and notes that many use cases sold as agentic don't need an agentic implementation. Five questions for any vendor: - Which decisions does the model make at run time, and which are scripted? - Which tools can it call, with which credentials, and what can each one change? - When does it hand over to a person? - What are its pass rate and cost per resolved task on your cases, in your language? - Are its actions logged where the agent cannot edit them? A scripted flow behind a chat window is often the right product. The problem is paying agent prices, and taking agent risks, for it. ## When is an agent worth it? Four concrete cases When the path depends on what the system finds and each task is worth the extra tokens and controls. OpenAI's guide gives three signals: decisions full of exceptions, rules too costly to maintain and heavy reliance on unstructured data; otherwise, a deterministic solution may suffice. - FAQ support. Policies, hours, products: a chatbot with search over your documents. Tools would add risk, not value. - Order status. One lookup, one answer, the same steps every time: a chatbot with a read-only query, or a fixed workflow. - Refunds. Policy, history, evidence and exceptions are weighed case by case: here an agent earns its place. OpenAI cites refund approval as complex decision-making and lists large refunds among actions that need human oversight. Typical design: the agent gathers facts and proposes; a person approves above a set amount. - Back-office reconciliation. Matching bank lines to ERP invoices follows rules, so a workflow does most of it; for the leftovers (partial payments, free-text references), an agent can propose matches for a person to confirm. Anthropic sees customer support as a natural fit for agents: a chat interface plus tools to read customer records and orders, actions such as refunds and success measured in resolutions. Without actions to take, an agent is an expensive chatbot. ## When is a form or a plain workflow better than both? When the input is structured, the rules are stable and errors are expensive. Changing a billing address, booking an appointment or approving an expense under a fixed limit needs validation, not language understanding: a form, an API call or a classic automation is faster, cheaper and easier to audit. - Structured inputs: if data arrives in fields, a model only adds cost and variance. - Hard rules: rates, limits and eligibility belong in code, where they can be tested. - Low volume: a task done a few times a month rarely repays design, evaluation and monitoring. Models still help at the edges: routing incoming emails to the right form, or extracting fields from a PDF before a rule-based process. More in what to automate first with AI. Rule of thumbIf the process fits on a one-page flowchart with no “it depends” box, you probably don't need an agent. ## What risks does each option carry, and how do you control them? A chatbot mostly risks bad answers; an agent, bad actions. So controls sit on the text for a chatbot and on tools and permissions for an agent. | Risk | Chatbot | AI agent | Control | | --- | --- | --- | --- | | Wrong or invented answer | Damages trust | Also feeds wrong actions | Search with citations, evals on real cases | | Prompt injection (LLM01) | Leaked system prompt, improper answers | Hidden instructions trigger actions | Treat everything it reads as untrusted; least privilege | | Excessive agency (LLM03) | Rare without tools | The main risk | Few tools, scoped credentials, confirmation of irreversible steps | | Compounding errors | Limited to one answer | Carried from step to step | Caps on steps and spend, checks between stages | | Personal data | Conversation logs | Logs plus whatever the tools reach | Law 1581 or GDPR, retention rules | Simon Willison's “lethal trifecta” sums up the agent case: private data, untrusted content and a way to send data out, combined in one agent, let an attacker steal data through what it reads. Remove one leg before launch; the full rules are in AI agents in production. Europe: since August 2, 2026, Article 50(1) of the AI Act requires providers to make sure people know they are talking to an AI system, unless it is obvious. Colombia: Law 1581 of 2012 governs the personal data a chatbot or agent processes, including data sent to models hosted abroad. And Gartner expects rules guaranteeing access to a human to raise assisted service volume by 30% by 2028: design the handoff from day one. ## How do you start without overbuilding? - Pick one task with an owner and a metric: resolution rate, handling time or error rate. - Build the simplest version that could work: one model call with search over your documents, then a workflow. - Write 20 to 50 test cases from real conversations, in your customers' languages; Anthropic suggests that size for a first evaluation set. - Connect tools read-only first, then add write actions one at a time, with confirmation and logging. - Define when a person takes over: after how many failures, and for which actions. - Measure cost per resolved task on real traffic before widening the scope. An assistant inside your own product is a sound first step. In the app we built for Réseau Mazette, a network of French-speaking professionals and entrepreneurs in Spain, Mia, the AI copilot, has its own tab next to the directory and events, where members open the chat to ask their questions (see the case). For assistants on your website or WhatsApp, see our AI chatbots for business. ### Key takeaways - A chatbot answers; an agent chooses its steps and acts through tools; a workflow runs steps your code defines. - Decide per task: FAQs and order status rarely need an agent; refund exceptions and messy reconciliations can. - Agents use about 4 times the tokens of a chat, per Anthropic: the task has to be worth it. - Gartner estimates only about 130 of thousands of agentic vendors are real: ask what the model actually decides. - Start with the simplest option, test on 20 to 50 real cases and give agents least privilege and a human handoff. ## Sources - Building effective agents · Anthropic, 2024-12-19 - A practical guide to building agents · OpenAI, 2025-04 - How we built our multi-agent research system · Anthropic, 2025-06-13 - Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 · Gartner, 2025-06-25 - Gartner predicts GenAI cost per resolution for customer service will exceed offshore human agent costs by 2030 · Gartner, 2026-01-26 - AI Act, Article 50: transparency obligations for providers and deployers of certain AI systems · AI Act Service Desk, European Commission, 2026-10-08 - Ley Estatutaria 1581 de 2012 · Función Pública (Colombia), 2012 - OWASP GenAI LLM Top 10 2026 · OWASP GenAI Security Project, 2026-08-03 - Demystifying evals for AI agents · Anthropic, 2026-01-09 - The lethal trifecta for AI agents: private data, untrusted content, and external communication · Simon Willison, 2025-06-16 Editorial note: this analysis reflects the public information available on the review date. Models, prices and rules change fast; every third-party figure links to its source, and our opinions are labeled as such. Spotted an error? Write to contact@slash-digital.io. Frequently asked questions ## The questions we hear often **What is the difference between a chatbot and an AI agent?** A chatbot answers messages in a conversation. An AI agent pursues a goal: the model chooses the next step and calls tools to read data and act in your systems until the task is done or it hands over to a person. **Is an AI agent better than a chatbot?** Not by default. An agent costs more per task (about 4 times the tokens of a chat, according to Anthropic), takes longer to reach production and can take wrong actions. It is better only when the task needs judgment across several steps and its value pays for that cost. **Can a chatbot become an agent later?** Yes, and it is often the best path: start by answering questions, add read-only lookups, then add write actions one at a time, with confirmation and logs. The conversation layer stays; the tools and controls grow. **What is agent washing?** Gartner's term for rebranding assistants, RPA or chatbots as agents without substantial agentic capabilities. Gartner estimates that only about 130 of the thousands of agentic AI vendors are real. **Do customers have to be told they are talking to an AI?** In the European Union, yes: since August 2, 2026, Article 50(1) of the AI Act requires that people know they are interacting with an AI system, unless it is obvious. Colombia has no AI law yet, but Law 1581 of 2012 governs the personal data the assistant processes. --- # Grok Bot with Claude Opus 5.5: why we prefer it URL: https://slash-digital.io/en/insights/grok-bot-claude-opus-5-5/ Agents and engineering · October 2026 On October 7, 2026, Elon Musk announced that Grok Bot will hand each task to the best available model, Anthropic's Claude Opus 5.5 included. That tips our preference among the three personal agents launched in August and September. Here are our criteria, a fair comparison and the caveats that remain. Oct 7, 2026 Opus 5.5 announced for Grok Bot58 points Opus 5.5, first on Artificial AnalysisOct 8, 2026 date of this assessment Start a project All insights Jean MartialCEO and co-founder, Slash Digital Lab PublishedOctober 8, 2026 Reading time8 min TopicAgents and engineering In short As of October 8, 2026, Grok Bot is the personal AI agent we prefer for companies, ahead of Meta's Muse and OpenAI's Dots. It is not tied to one lab: on October 7 it was announced to route work to Claude Opus 5.5, first on the Artificial Analysis index. It already offers shared Bots in Slack with Teams and Enterprise controls, and its app is sold in Colombia and France, while Muse is limited to the US and Canada. The caveats are real: no published routing rules, no model picker, key controls reserved for Enterprise and computers located in the United States. Dots has richer integrations and finer controls; Muse has more reach. This position rests on public documentation, press reports and independent leaderboards, not on a benchmark of our own; we will revisit it once the routing is documented. For your company Want an agent that acts in your systems, with human oversight? Slash Digital Lab designs AI agents for support, operations and back office. AI agents → ## Why is Grok Bot our preferred agent? Because it comes out ahead on the four criteria we weigh for a company, even with the caveats below. We compared the three always-on personal agents launched between August and September 2026: - Model: access to several providers instead of one lab, with Claude Opus 5.5 announced on October 7. - Teams: shared Team Bots in Slack since September 26; Enterprise access since September 3, with audit logs, network policy and SCIM; and ISO/IEC 27001 and 42001 certifications whose scope includes Grok Bot, according to Cursor. - Availability: offered in Colombia and France; Dots also qualifies through its business plans, Muse does not. - Data: no training on customer data when Privacy Mode is on, and model providers bound by zero-retention agreements. This is Slash's editorial position, based on public documentation and press reports, not on lab tests. If you are planning AI agents for companies, use it as the starting point for a pilot, not as a verdict. ## What did Musk announce on October 7, and what is confirmed? A change of engine, partly live and not yet documented. On October 7, Musk posted on X that SpaceX "will use the best back end model for any given task", naming Claude Opus 5.5, Midjourney, Suno and other leading APIs, without saying which task goes where. The same day, 9to5Mac reported that Grok Bot could already use Opus 5.5, and a Grok Bot team member wrote that the model was rolling out and would power all Bots. Midjourney and Suno have no date. The official changelog (version 0.68.1, October 7) does not mention it, and the security documentation still says there is no customer-facing model picker and that the provider mix can change. For context: Grok Bot comes from SpaceXAI, the former xAI absorbed by SpaceX (Musk says it will be renamed SpaceXSI), and runs on infrastructure from Cursor, also owned by SpaceX. It launched in beta on August 11, 2026, Bloomberg counted about 418,000 weekly users in mid-September (as cited by tbreak) and users reported brief outages of Grok services on October 6, with no official explanation. Status on October 8, 2026Opus 5.5 in Grok Bot is a reported rollout, not a documented feature: no routing rules, no list of which task uses which model and no date for Midjourney or Suno. Our preference depends on that routing being documented. ## How do Grok Bot, Muse and Dots compare? All three promise an agent with its own cloud computer that keeps working while you are away and asks before sensitive actions. The differences lie in the models, the channels, the markets and the data terms. | Criterion | Grok Bot | Muse | Dots | | --- | --- | --- | --- | | Maker and launch | SpaceXAI, beta launched Aug 11, 2026 | Meta, Sept 8, 2026 | OpenAI, Sept 29, 2026 | | Models | xAI and third-party models; Opus 5.5 announced Oct 7; no picker | Muse Spark | GPT-6 Astra | | Where it runs | One cloud computer per user, shared by their Bots, in the US | Dedicated VM per person; a Sentinel agent filters outbound traffic | Its own cloud computer; laptop access optional, off by default | | Channels and apps | Mac, Windows, Linux, iPhone, iPad, Slack; Gmail, Outlook, GitHub | Muse app and WhatsApp; Slack, Shopify, QuickBooks and more | ChatGPT, Slack, Teams; 4,000+ apps via plugins | | Availability | No country list; app listed in Colombia and France | US and Canada | Pro outside the EEA, Switzerland and UK; Business Premium in all ChatGPT regions; Enterprise beta | | Plans | Included in paid Cursor and SuperGrok plans | Free for most uses, paid plans for more | First dot included in Pro and Business Premium | | Training on your data | No, with Privacy Mode on | Yes, unless you opt out | No by default on Business, Enterprise and Edu | | Controls | Auto Review; audit logs and network policy on Enterprise only | Approvals, full action log, hidden credentials | Custom Rules, Activity View, auto-review | | Reported usage | About 418,000 weekly users | 3 million+ weekly users | Not disclosed | On price, Grok Bot has no standalone plan: Cursor lists Pro from $20 a month and Teams at $40 per user, before taxes, with weekly usage limits. OpenAI has not yet priced additional dots. ## Why does model routing matter so much to us? Because the best model changes faster than an agent contract. On October 8, Claude Opus 5.5 (released on September 22 at $4 / $20 per million tokens) led the Artificial Analysis Intelligence Index, which aggregates agentic and knowledge-work evaluations, ahead of the models behind Dots and Muse. | Model | Used by | Score | | --- | --- | --- | | Claude Opus 5.5 | Grok Bot (announced) | 58 (1st) | | GPT-6 Astra | Dots | 53 | | Muse Spark 1.3 | Muse | 48 | | Grok 4.7 | Grok Bot (SpaceXAI model) | 46 | On Arena's text leaderboard, Opus 5.5 was second that day (1,507 points), behind a preliminary Gemini 4 Argon entry: leadership depends on the test, as we explain in our Claude Opus 5.5 analysis and our GPT-6 Astra analysis. The second argument is resilience. On June 12, 2026, a US export-control directive forced Anthropic to withdraw Fable 5 and Mythos 5 from all users; Fable 5 came back 19 days later. An agent tied to one lab inherits that kind of outage. The price of routing is control: Grok Bot, not you, decides which provider sees each request, and Anthropic is also a SpaceX compute customer. ## Where are Dots and Muse stronger? Dots has the widest catalog (more than 4,000 apps through OpenAI's plugins) and the most granular documented controls: Custom Rules that let a dot act, act if pre-approved, ask first or hand off to you; an Activity View; and read-only tools for its background research. Business Premium is available in every supported ChatGPT region, Europe included, and OpenAI does not train on business workspace content by default. Its limits: a single model provider and the exclusion of Pro users in the EEA, Switzerland and the UK. Muse has reach: more than 3 million weekly users according to internal data reviewed by The Information, and it lives inside WhatsApp. On paper its design is strong: a dedicated VM per person, a separate Sentinel agent that must approve anything going out to the internet, credentials kept out of the model's sight and a full audit trail. Its limits: it is only available in the US and Canada, it trains on interactions unless you opt out, and a researcher found a critical flaw soon after launch, which Meta fixed immediately. ## What risks do we still see in Grok Bot? Real ones. An agent that reads email, calendars and files brings together private data, untrusted content and a way to send data out; any message it reads can carry hidden instructions (see prompt injection and agent security). The damage is not theoretical: in February 2026, an OpenClaw agent bulk-deleted a researcher's inbox and ignored stop commands (AI agents in production). Specific to Grok Bot: - All of one user's Bots share the same computer, files and sessions; Cursor advises a separate user for workloads that need their own credentials. - Audit logs, network policy and action recording are Enterprise-only; without a policy, outbound traffic is allowed. - There is no model picker, your company's model allowlist is not guaranteed to apply, and SpaceXAI has not said how the newly named providers fit its data terms. - Secrets and files added to a Team Bot are available in every teammate's chat. - In August, Adversa AI disclosed an injection that leaked chat history from Grok's web chat, a different product; it was still unfixed more than two months after the report. ## What does it mean for companies in Colombia and Europe? Access. On October 8, the Grok Bot iPhone app was listed in the Colombian App Store and the French one, with plans in pesos and euros; SpaceXAI publishes no country list. Dots reaches Pro and Business Premium users in Colombia, and European companies through Business Premium or the Enterprise beta. Muse is available in neither region. Data. Grok Bot's computers run in the United States and on-premises deployment is not supported. In Colombia, sending personal data there is an international transfer under Law 1581 of 2012; in the European Union, the GDPR transfer rules apply. Ask for the DPA and the list of model subprocessors before connecting work accounts. Our recommendationFor a team in Bogotá, Madrid or Paris: start Grok Bot on low-risk work (research, drafts, internal reporting) with an account that holds no customer data, and keep Dots on the shortlist if you already work on ChatGPT Business. ## How do you pilot a personal agent without exposing your company? - Choose two or three low-risk tasks and define success before you start. - Use a dedicated account with non-sensitive data, never an executive's inbox. - Grant least privilege: read-only where possible, approval for anything that sends, buys or deletes. - Turn on Privacy Mode at team level and keep customer data out. - On Enterprise, set a network policy, turn on action recording and stream audit logs to your SIEM from day one. - Write a usage policy: what can be connected, who approves and how incidents are reported (see our AI usage policy guide). - After 30 days, review errors, approvals, cost and quality, then decide. We will revisit this position when SpaceXAI documents its routing, when Dots reaches Pro users in Europe or when Muse launches outside North America. ### Key takeaways - As of October 8, 2026, Grok Bot is our preferred personal agent for companies, ahead of Muse and Dots. - The main reason is routing to Claude Opus 5.5, first on the Artificial Analysis index, announced on October 7. - That routing is a reported rollout, with no published rules and no model picker. - Dots leads on integrations and controls, Muse on reach; Muse is limited to the US and Canada. - Pilot with non-sensitive data, least privilege, Enterprise controls and a written usage policy. ## Sources - Introducing Grok Bot · SpaceXAI, 2026-08-11 - Grok Bot changelog · SpaceXAI, 2026-10-07 - Grok Bot security · Cursor documentation, 2026-10 - Grok Bot security FAQ · Cursor documentation, 2026-10 - Introducing dots · OpenAI, 2026-09-29 - Getting started with your dot · OpenAI Help Center, 2026-10 - Introducing Muse: The World's First Personal AI Agent Built for Everyone · Meta, 2026-09-08 - The Future Is for Everyone: Muse for Small Business · Meta, 2026-09-29 - Grok Bot just got a lot smarter thanks to Anthropic's Claude · 9to5Mac, 2026-10-07 - Grok Bot will use Claude Opus 5.5, Midjourney and Suno, says Musk · The Next Web, 2026-10-07 - Meta's Muse reportedly passes 3 million weekly users and 5m downloads · The Next Web, 2026-10-02 - AI model leaderboards (Intelligence Index) · Artificial Analysis, 2026-10-08 Editorial note: this analysis reflects the public information available on the review date. Models, prices and rules change fast; every third-party figure links to its source, and our opinions are labeled as such. Spotted an error? Write to contact@slash-digital.io. Frequently asked questions ## The questions we hear often **Does Grok Bot already use Claude Opus 5.5?** Partly, by all reports. Musk announced the change on October 7, 2026, 9to5Mac reported that day that Grok Bot could already use Opus 5.5 and a team member said it would power all Bots. SpaceXAI has published no routing rules, and customers cannot choose the model. **Is Grok Bot available in Colombia and Europe?** Its iPhone app was listed in the Colombian and French App Stores on October 8, 2026, but SpaceXAI publishes no country list. The Bots' computers run in the United States, so check international transfer rules before connecting personal data. **How much does Grok Bot cost?** There is no standalone plan. It is included in paid Cursor plans (Pro from $20 a month, Teams at $40 per user a month, before taxes) and in linked SuperGrok plans, with weekly usage and optional on-demand billing. **Do these agents train on my data?** Muse does unless you opt out, according to Meta. OpenAI does not train on Business, Enterprise or Edu content by default. For Grok Bot, Cursor says customer data is not used for training when Privacy Mode is on. **What is this preference based on?** On public documentation, press reports and independent leaderboards as of October 8, 2026, not on a published benchmark of our own. It will change if the facts do. Any agent should be tested on your own tasks before wider use. --- # Event networking app: how to build one that works URL: https://slash-digital.io/en/insights/event-networking-app/ Strategy and regulation · October 2026 We built the apps for Colombia Tech Week 2026 and its festival, Colombia Tech Fest, and the native app of Réseau Mazette. Here is what works, screen by screen, and what Apple, Google and the law require before doors open. Aug 13–14, 2026 Colombia Tech Fest, Bogotá30 min slots for 1:1 meeting requests2 stores one Flutter codebase (Mazette) Start a project All insights Jean MartialCEO and co-founder, Slash Digital Lab PublishedOctober 8, 2026 Reading time7 min TopicStrategy and regulation In short A great event networking app turns an agenda and a guest list into meetings that actually happen. It needs controlled access, a home screen that answers “what's next for me”, a directory that leads to a meeting request and a 1:1 chat, screens that survive a failing venue network, and the safety and privacy rules Apple and Google require. For your company AI inside your mobile app? Slash Digital Lab builds mobile apps from Bogotá, with embedded AI where it adds value. Mobile apps → ## What makes an event networking app work? It works when meetings happen; map, notifications and sponsors serve that goal. Plan for three moments: before (invitation, profile, agenda), during (what's next, where, with whom) and after (contacts and follow-up). Colombia Tech Week 2026 ran in Bogotá from August 10 to 16, and its festival, Colombia Tech Fest, was held open-air at Parque El Chicó on August 13 and 14. We were a Gold Sponsor of the week and built its official app for thousands of attendees and the festival's attendee app, invitation-only and owned by the festival. We also built the native Flutter app of Réseau Mazette, a network of French-speaking professionals and entrepreneurs in Spain. | Criterion | Event app (Colombia Tech Week, Colombia Tech Fest) | Community app (Réseau Mazette) | | --- | --- | --- | | Lifespan | One week, reusable | All year | | Access | Invitation only | Members only | | Core screens | Home, map, agenda, networking, chat, calendar | Members, events, resources, messages, Mia | | How people meet | Meeting requests, 30-minute slots, 1:1 chat | Events by city, circles, messages | | Where AI sits | Search and filters do the work | Mia, a copilot in its own tab | | Who owns it | The festival (Colombia Tech Fest) | The network | Planning one? See how we build mobile apps with built-in AI. ## Open signup or invitation-only access? For a closed guest list, invitation-only: the organizer knows who gets in, the directory stays trustworthy and there is no open signup to moderate. In the Colombia Tech Fest app, a first-time guest pastes the emailed code or opens the organizer's link, then follows numbered steps: invitation, privacy, password, profile, agenda. The login screen states “Sin registro público” (no public signup). The Mazette app is members-only, with the website's credentials. Make consent its own step. Colombia's Law 1581 of 2012 requires prior, express and informed authorization that can be consulted later (Articles 3 and 9), with a copy kept (Article 17). The GDPR requires a legal basis and data protection by design and by default (Article 25). Build in the store rules. If people can create an account, Apple requires in-app deletion (guideline 5.1.1(v)) and Google Play also wants a web link to request it; in our reading, activating an invitation counts as creating one. Adding social login (Google, LinkedIn) would bring in Apple's guideline 4.8, which requires an equivalent option limited to name and email that can hide the email; apps that use only their own accounts, like these two, are exempt. (enlarge) Invitation-only access (enlarge) Members only Colombia Tech Fest and Réseau Mazette apps (2026). ## What should the home screen show? What's next for this person, first. The Colombia Tech Fest home greets the guest by first name and opens on “Próximo en tu día” (next in your day): the time and zone of the next session. Below: shortcuts to map, agenda, networking and calendar, day tabs and a search by title, speaker or company. The agenda filters by zone and category (fireside chat, panel, keynote, workshop); a star saves a session to favorites, which feed the personal calendar, and the agenda updates live. The map shows the venue's two zones and its open-access food and brand spaces. Our recommendation: design for ten-second glances between sessions, with time, place and one action per card. Give sponsors a visible slot, like the home carousel, not a step in the attendee's path; a sponsor who wants more deserves a live experience at the venue (see generative AI in brand experiences). (enlarge) What's next, first (enlarge) Agenda with filters and favorites (enlarge) Venue map Colombia Tech Fest, July 2026 test build; profile photo blurred. ## How do you turn a directory into meetings? With a short path: find someone, see when they are free, ask, talk. In the apps we built for Colombia Tech Week and Colombia Tech Fest, it takes five steps: - Profile: role, industry, company and availability in 30-minute blocks. - Search: by name, company or role, with filters; on a tablet, industry chips such as Fintech or Healthtech. - Request: one profile button asks for a meeting, another blocks the user. - Chat: a conversation starts as a request, under “Solicitudes enviadas” (sent) for the sender and “Invitaciones” for the recipient, so both can agree to meet. - Calendar: “Mi calendario” brings together favorite sessions and meetings, with a “Solicitar reunión” (request a meeting) button. The stores require a safety layer. Apple's guideline 1.2 asks apps with user-generated content to filter objectionable material, act on reports promptly, offer a way to block abusive users and publish contact details; Google Play requires accepted terms, in-app reporting and blocking, and blocking in one-to-one messaging. (enlarge) Searchable directory with filters (enlarge) Requests before the chat (enlarge) Sessions and meetings together Colombia Tech Fest, July 2026 test build; attendees' names, photos and messages blurred. ## What happens when the venue network fails? The app must stay useful. Colombia Tech Fest was an open-air festival in a park, and its app had to hold up when most guests connect at once: it has offline maps and a live agenda, and it was built for that peak. Our recommendations: - Store on the phone the agenda, map, calendar and profiles of the people each attendee will meet. - Queue offline actions (favorites, messages, requests) and sync them when the network returns. - Load-test the peak: everyone signing in and searching within the same minutes. - Give organizers admin tools, like Colombia Tech Fest's, for last-minute changes. Notifications need permission. Since Android 13 (API level 33), apps need the POST_NOTIFICATIONS permission, which Android suggests requesting after a user action, for example. iOS apps must ask too, and Apple recommends asking in context, such as after the person schedules a first task. Our equivalent: after the first favorite, not at launch. Field test we recommendThe day before, walk the venue with a phone in airplane mode: agenda, map, calendar and meetings must open. Then reconnect and check that pending actions sync. ## What changes when the app must live all year? The job changes: the app must give members reasons to return between events, not carry them through one week. Réseau Mazette holds events in cities such as Málaga, Barcelona, Valencia and Madrid; its app combines events by city, discussion circles by topic and city, shared resources, messaging, memberships and Mia. Home greets the member with upcoming events and the newest members. Event cards show a poster, place or “online”, date and price. Resources hold searchable replays of webinars and talks, and the profile shows membership status and directory visibility. Two decisions carry the rest: the same login and brand as the website, and one Flutter codebase on the App Store and Google Play, so every member gets the same full product. An event app's job ends with the event, yet Colombia Tech Fest owns its app, ready for future editions. Our recommendation: if your event repeats, budget for an app you reopen, not rebuild. (enlarge) Events filtered by city (enlarge) Circles by topic and city Réseau Mazette app; message previews blurred. ## Where does AI help in a networking app? Where the question is open and the answer is spread across screens, like Mia's sample question: “What are the upcoming events and who is attending?” Where a list or a filter already works, a copilot adds little. In the Mazette app, Mia has a dedicated tab and answers about events, participants, shared resources, the member directory and practical information. The screen asks for precise requests (city, topic, first name) and states that Mia follows the same visibility rules as the website: the main lesson. Our recommendations: - Inherit the app's permissions: a member hidden from the directory must never appear in an answer. - Treat member-written text as untrusted: profiles and messages can carry instructions aimed at the model (see prompt injection and agent security). - Say it is an AI: since August 2, 2026, Article 50(1) of the EU AI Act requires telling people they are interacting with an AI system unless it is obvious (see our 2026 regulatory map). - Explain suggestions and keep them opt-in: if the app proposes whom to meet, it says why. Our ruleIf a person cannot see a piece of data in the app, the assistant cannot either. Enforce permissions on the server, before the model receives any context, not in the prompt. (enlarge) Mia, the network's copilot ## Should you build your own app or rent an event platform? Our rule of thumb: rent a white-label platform if your event fits its template and you won't reuse the product; build your own to own the product and its data, reuse it each edition or run custom flows (invitations, zones, sponsors). | Criterion | Own app | Event platform (white-label SaaS) | | --- | --- | --- | | Ownership and reuse | Yours, reusable | Licensed from the vendor | | Brand | Yours everywhere | Depends on the plan | | Data control | You set hosting, retention and deletion | Set by the contract | | Custom flows | Whatever the event needs | What the platform allows | | Time to launch | Longer the first time | Shorter if the template fits | | Cost model | Build and maintenance across editions | License per event, attendee or year | ### Before doors open: the checklist - Invitations tested end to end: email, code, link, resend. - Consent as its own step, with a record, and in-app account deletion (plus Google Play's web link). - Block and report on profiles and chat, published contact, a named moderator. - Agenda, map and calendar offline; offline actions queued. - Notification permission requested in context. - A peak load test and admin access for last-minute changes. - Review buffer: Apple says it typically reviews at least 50% of submissions in under 24 hours and 90% in under 48, and accepts expedited requests tied to an event. - Test tracks: TestFlight takes up to 10,000 external testers once App Review approves a first build; on Google Play, personal accounts created after November 13, 2023 need a closed test with at least 12 testers for at least 14 days before production. - On-site help for lost invitations and forgotten passwords. - After the event: contact export or follow-up, and a date for deleting data. In Colombia, the Superintendence of Industry and Commerce (SIC) supervises Law 1581 of 2012: keep proof of every authorization. In Europe, under GDPR Article 25, personal data must not by default be made accessible to an indefinite number of people without the person's intervention: make directory visibility a choice. An AI assistant also falls under Article 50 of the AI Act. General information, not legal advice. ### Key takeaways - Measure a networking app in meetings that happen; design the path from profile to meeting first. - Invitation-only access keeps the directory trustworthy; make consent its own step. - Home answers “what's next for me”; favorites feed one personal calendar. - Plan for the venue: offline agenda and map, queued actions and a peak load test. - An AI copilot inherits the app's permissions, distrusts member text and says it is an AI. ## Sources - App Review Guidelines (1.2 User-Generated Content, 4.8 Login Services, 5.1.1(v) Account Sign-In) · Apple Developer, 2026-10 - Understanding Google Play's app account deletion requirements · Google Play Console Help, 2026-10 - User Generated Content · Google Play Developer Policy Center, 2026-10 - App Review (review times and expedited reviews) · Apple Developer, 2026-10 - TestFlight · Apple Developer, 2026-10 - App testing requirements for new personal developer accounts · Google Play Console Help, 2026-10 - Notification runtime permission · Android Developers, 2026-10 - Asking permission to use notifications · Apple Developer Documentation, 2026-10 - Regulation (EU) 2016/679 (GDPR), Articles 6 and 25 · EUR-Lex, 2016-04-27 - Ley Estatutaria 1581 de 2012 (protección de datos personales) · Función Pública, Gestor Normativo, 2012-10-17 - AI Act, Article 50: transparency obligations for providers and deployers of certain AI systems · European Commission, AI Act Service Desk, 2026-07-27 - Colombia Tech Fest 2026 llega a Bogotá: el festival de tecnología a cielo abierto más grande del país · Noticias RCN, 2026-08-11 Editorial note: this analysis reflects the public information available on the review date. Models, prices and rules change fast; every third-party figure links to its source, and our opinions are labeled as such. Spotted an error? Write to contact@slash-digital.io. Frequently asked questions ## The questions we hear often **What features does an event networking app need?** At minimum: controlled access, a home screen with what's next, an agenda with filters and favorites, a venue map, a searchable directory, meeting requests, 1:1 chat, a personal calendar and notifications. Add offline access for the venue, and blocking, reporting and account deletion for the stores. **Open signup or invitation?** For a closed guest list, invitation: the organizer knows who gets in and there is no open signup to moderate. For a membership network, a members-only login with the website's credentials, as in the Réseau Mazette app. **Must attendees be able to delete their account in the app?** Yes, if the app supports account creation: Apple requires in-app account deletion (guideline 5.1.1(v)), and Google Play requires an in-app path plus a web link to request deletion. **How long does publishing on the App Store and Google Play take?** Apple says it typically reviews at least 50% of submissions in under 24 hours and 90% in under 48, and accepts expedited requests for releases tied to an event. On Google Play, a personal developer account created after November 13, 2023 first needs a closed test with at least 12 testers for at least 14 days. Our advice: leave weeks, not days. **How much does an event networking app cost?** It depends on the platforms, custom flows (invitations, zones, sponsors), integrations, offline mode, AI features, on-site support and maintenance between editions. Slash has not published prices for this service; we quote after a free first read. --- # Claude Opus 5.5: release date, pricing and what changes URL: https://slash-digital.io/en/insights/claude-opus-5-5/ Models · October 2026 Anthropic launched Claude Opus 5.5 on September 22, 2026. It promises Fable 5.1-level results on most work at a lower price than Opus 5. Here is what changes for a team putting it into production. $4 / $20 per million tokens1M tokens of contextSept 22, 2026 launch Start a project All insights Jean MartialCEO and co-founder, Slash Digital Lab PublishedSeptember 26, 2026 ReviewedOctober 8, 2026 Reading time10 min TopicModels In short Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens (Opus 5: $5 and $25), keeps a 1 million token context window and, according to Anthropic, performs like Claude Fable 5.1 on most work, generates output more than 30% faster than Opus 5 and costs 40% less to run on typical workloads. For engineering teams it brings breaking changes: thinking can no longer be turned off, the default effort drops to medium and forced tool use is gone. Our reading: it is the new default starting point for agents and knowledge work, provided you validate it on your own prompts. For your company Which model handles your real cases best, in the cloud or local? At Slash AI Lab we compare models on your own prompts before you commit budget or private data. Slash AI Lab → ## What Claude Opus 5.5 is and what it promises Claude Opus 5.5 is the new model in Anthropic's Opus line, released on Tuesday, September 22, 2026, two months after Opus 5 (July 24). Anthropic describes it as a model for long-running agentic coding and knowledge work, and its documentation now recommends starting with Opus 5.5 for most workloads. The headline promises, all from Anthropic: performance at the level of Claude Fable 5.1 (its most capable widely released model) on most work, a 40% lower running cost than Opus 5 on typical workloads and output generation more than 30% faster. It is available on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry and Claude Platform on AWS, and in the Claude Pro, Max, Team and Enterprise plans. It is also the first model Anthropic has shipped since its CEO, Dario Amodei, published the essay We Must Pace the Frontier in September 2026, calling for a slower pace of capability gains to leave time for alignment and external evaluation. The same day, OpenAI launched GPT-6 Sol and GPT-6 Luna, and Anthropic announced Claude Sonnet 5.5 and Claude Haiku 5.5, which shipped on September 28 and October 7. ## Where it fits in the Claude family Four current models, four price points. | Model | Input / output | Context | Max output | Default effort | Reliable knowledge | | --- | --- | --- | --- | --- | --- | | Claude Fable 5.1 | 10 / 50 | 1M | 128K | high | June 2026 | | Claude Opus 5.5 | 4 / 20 | 1M | 128K | medium | June 2026 | | Claude Sonnet 5.5 | 2 / 10 | 1M | 128K | high | June 2026 | | Claude Haiku 5.5 | from 0.10 / 0.50 | 1M | 128K | medium | June 2026 | The practical reading: Opus 5.5 takes the slot that used to require paying for the Fable tier. Fable 5.1 remains for the most demanding reasoning and very long-horizon agents, or for when your evaluations of Opus 5.5 at high effort still fall short. For volume and latency the options are now Sonnet 5.5 and Haiku 5.5; Sonnet 5 and Haiku 4.5 become legacy models. Anthropic says Sonnet 5.5 beats Opus 5.5 on Terminal-Bench 4.0 (70.6% vs 66.4%) but trails it on every other row of its table, and that Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment. If you are deciding which model to bring into your company, explore our AI for companies services. Two pricing details matter in production: an Opus 5.5 prompt-cache read costs $0.20 per million tokens (5% of the input price, versus $0.50 on Opus 5), and the Batch API applies a 50% discount. On batches, Opus 5.5 supports up to 300,000 output tokens with a beta header. ## Reading the benchmarks carefully What Anthropic publishes and what independent leaderboards say. | Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | | --- | --- | --- | --- | --- | | Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | 57.9% | | FrontierCode v1.1 | 54.4% | 50.3% | 48.0% | 53.3% | | CursorBench 4.0 | 57.8% | 51.8% | 46.6% | not reported | | OSWorld 2.1 (computer use, partial score) | 81.8% | 80.7% | 74.0% | not reported | | Humanity's Last Exam (with tools) | 67.7% | 65.6% | 63.6% | 57.2% | | GDPval-AA v2.1 (Elo, professional work) | 1,846 | 1,735 | 1,708 | 1,542 | | AutomationBench | 40.0% | 31.4% | 26.9% | 41.4% | | Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% | Three cautions before deciding on this table. First, these tests are run or chosen by the vendor itself, several in new versions (Terminal-Bench 4.0, OSWorld 2.1) with little comparison history. Second, it doesn't win everywhere: in the same table, GPT-6 Astra comes out ahead on AutomationBench and Terminal-Bench-Science. Third, none of these tests is in business Spanish or French, which is where we most often see models break. Independent leaderboards point the same way, with nuances. On October 8, Opus 5.5 was still first on the Artificial Analysis Intelligence Index (58 points, version 4.3.2) and led the official Terminal-Bench 4.0 board at 64.8%, below the 66.4% Anthropic reports. On the Arena text leaderboard it moved to second place (1,507 points), behind a preliminary score for Gemini 4 Argon. On ARC-AGI-2, GPT-6 Astra scored 95.0% on September 26 against 93.3% for Opus 5.5, although at almost three times the cost per task. We compare the two in Claude Opus 5.5 vs GPT-6 Astra. Anthropic also cites customer anecdotes: a 680,000-line code migration in under a day, or an audit of a 200,000-line codebase in under three hours. They come from its launch communication: a useful signal of where the model points, not a promise for your project. ## What changes for engineering teams Four changes break code written for Opus 5. Thinking is always on. Opus 5.5 uses adaptive thinking that cannot be disabled; the only dial is the effort parameter (low, medium, high, xhigh, max). The default drops to medium, one level below Opus 5. Anthropic says Opus 5.5 at medium beats Opus 5 at high on its coding and knowledge-work evaluations, and that on several coding evaluations low comes close at much lower cost. We cover this in depth in when paying for thinking pays off. - Sending thinking: disabled or a fixed thinking-token budget returns a 400 error. - Forced tool use (tool_choice set to any or a specific tool) is no longer supported: use auto with strict tools, or structured outputs. - Thinking blocks are tied to the model and the conversation: if your code edits earlier turns, accounts created on or after August 31, 2026 get an error. History must be append-only. - Computer use only works through the new computer_toolset_20260801; the earlier tool is rejected. One more change breaks nothing but surprises teams: the short notes the model writes between tool calls now arrive as thinking blocks, empty by default. If your interface shows that progress to users, it will go quiet until you enable the updates display mode. Finally, fast mode (research preview, Claude API only) promises up to 2.5x speed at $8 / $40 per million tokens. Rule of thumbSet effort explicitly on every route of your application and re-run your evaluation suite. Copying your Opus 5 settings gives longer turns and more output tokens, because at the same level Opus 5.5 tends to think more. ## What it really costs: a worked example Price per token drops 20%; cost per task can drop much more. Take a typical agent step: 12,000 input tokens, 10,000 of which are a stable prefix (instructions, tools, reference documents) that can be cached, and 2,000 output tokens including thinking, which is billed as output. These are illustrative assumptions, not measurements from a project. | Scenario | Cost per step | Per 100,000 steps | | --- | --- | --- | | Opus 5, no caching | $0.110 | $11,000 | | Opus 5.5, no caching | $0.088 | $8,800 | | Opus 5, cached prefix | $0.065 | $6,500 | | Opus 5.5, cached prefix | $0.050 | $5,000 | The big saving is not the list price but the combination: cheap cache reads, a lower default effort and, according to Anthropic, fewer tokens per solved task. The number to watch is cost per completed task: a cheaper step that needs two retries is not cheaper. Effort matters too: the 40% saving Anthropic advertises refers to typical workloads, but at max effort Artificial Analysis measures $5.98 per index task for Opus 5.5 against $5.86 for Opus 5, because higher token use offsets the price cut. If the work isn't interactive, the Batch API halves the bill. More levers in the real cost of AI. ## Safety and refusals: what your team should know Opus 5.5 runs cybersecurity and biology safety classifiers similar to Fable 5.1's; the biology one is new compared with Opus 5. Anthropic says most cybersecurity tasks that trip those classifiers are re-routed to Opus 4.8, and at launch it said it would soon extend its Cyber Verification Program to Opus 5.5, with three tiers of trusted access. According to the documentation, finding vulnerabilities in source code is allowed, while dual-use biology research requests are declined except for participants in Anthropic's verification program. A new category, reasoning_extraction, declines attempts to make the model copy its internal reasoning into the answer. For a product team this has a concrete consequence: a refusal arrives as a normal response with stop_reason: refusal, not as an error. Your code has to handle it, log it and, where appropriate, retry on another model; the API can configure that fallback server-side. Pentest and offensive-security teams should test their real workflows before migrating. On robustness, Anthropic says Opus 5.5 matches or beats Opus 5 against prompt injection in every setting it tested, and attempts to circumvent containment boundaries 85% less often than Opus 5. Good news, but no model solves prompt injection on its own: agent design remains the main defense, as we explain in prompt injection and agent security. ## Which model for which job Our starting point before evaluating; it doesn't replace the evaluation. | Job | Starting model | Starting effort | Why | | --- | --- | --- | --- | | Coding agents, migrations, long refactors | Opus 5.5 | medium, raise to high if it fails | The use case Anthropic designed it for | | Long-document analysis, finance, legal | Opus 5.5 | medium | 1M context and cheap cache reads | | Very hard reasoning, very long-horizon agents | Fable 5.1 or Opus 5.5 at xhigh | high or above | Compare both in your evaluation | | High-volume customer service and chat | Sonnet 5.5 or Haiku 5.5 | low or medium | Latency and cost per conversation | | Bulk classification and extraction | Haiku 5.5 on batches | low | Cost rules; watch the price tier above 100K tokens | | Real-time answers with Opus | Opus 5.5 in fast mode | low or medium | Only if latency justifies paying double | Watch Haiku 5.5Haiku 5.5 costs $0.10 / $0.50 per million tokens for prompts up to 100K tokens and five times more above that, and its tokenizer counts about 30% more tokens than Haiku 4.5 for the same text. If you still run routes on Haiku 4.5, guaranteed only until at least October 15, 2026, plan the migration now. GPT-6 Luna ($0.10 / $0.50), GPT-6 Sol and GPT-6.1 Sol ($2 / $10) and the Gemini Flash models compete in the same segment. This table is a starting point. The method we use to decide, with 30 real prompts per client in three languages, is in how we choose a model, and the full vendor landscape is in the October 2026 model map. If you want a general-purpose agent, see why we prefer Grok Bot with Claude Opus 5.5. ## Checklist for migrating from Opus 5 - Set effort per route and run your evaluation at low, medium and high before choosing. - Remove any thinking: disabled or fixed thinking budget. - Replace forced tool use with auto plus strict tools or structured outputs. - Make your conversation history append-only: no editing of earlier turns. - If you use computer use, move to computer_toolset_20260801 and test it on Opus 5 first, which accepts both versions. - Display between-tool progress updates if your interface relies on them. - Handle stop_reason: refusal with logging and a fallback model. - Check your tier's Opus 5.5 rate limits before moving volume. - Compare cost per completed task, not price per token, over at least a week of real traffic. ## What it means for companies in Colombia and Europe Procurement and contracts. Opus 5.5 is on the three big clouds (AWS, Google Cloud and Microsoft). For many Colombian and European companies that means using it under existing contracts and spend commitments, without onboarding a new vendor. Mind the cost of residency: US-only inference costs 1.1x the base price, and regional endpoints on Bedrock and Google Cloud carry a 10% premium. Access risk. On June 12, 2026, a US export-control directive forced Anthropic to withdraw Fable 5 and Mythos 5 from all users; Fable 5 came back on July 1. Any company outside the US should keep a fallback model from another vendor, tested and ready to switch on. Personal data. In Colombia, sending personal data to a model hosted abroad is an international transfer under Law 1581 of 2012: check the data subjects' authorization and the provider's safeguards. In the European Union the GDPR applies and, from August 2, 2026, so do the AI Act transparency obligations for anyone deploying a chatbot. General-purpose model obligations fall on Anthropic; yours concern how you use it. Details in our 2026 regulatory map. Languages. No published benchmark measures the Spanish of a Colombian insurer or the French of a Paris law firm. That test is yours to run. At Slash AI Lab we run Opus 5.5 through the same 30-prompt suite per client that we use for any new model; we won't publish numbers until we have reproducible results. Our recommendationIf you run Opus 5 or Fable 5.1 in production today, start a controlled Opus 5.5 trial this week, with effort set explicitly and your evaluation in Spanish and French. If you use Sonnet for volume, test Sonnet 5.5, available since September 28 at Sonnet 5's price, before moving that route. ### Key takeaways - Opus 5.5 (September 22, 2026) costs $4 / $20 per million tokens, 20% less than Opus 5, with 1M context. - Anthropic places it at Fable 5.1 level on most work; the benchmarks are vendor-run and GPT-6 Astra wins two of them, but Opus 5.5 leads the official Terminal-Bench 4.0 board (64.8%) and the Artificial Analysis index. - Thinking can't be turned off: effort is the lever and its default drops to medium. - Four changes break code written for Opus 5; migrating needs a checklist and a fresh evaluation. - Measure cost per completed task: $0.20 cache reads and batches move the bill more than the list price. ## Sources - Introducing Claude Opus 5.5 · Anthropic, 2026-09-22 - Claude Opus 5.5: model page · Anthropic documentation, 2026-09-22 - What's new in Claude Opus 5.5 · Anthropic documentation, 2026-09-22 - Migrating to Claude Opus 5.5 · Anthropic documentation, 2026-09-22 - Models overview · Anthropic documentation, 2026-10-08 - Pricing · Anthropic documentation, 2026-09 - Claude Opus 5.5 system card · Anthropic, 2026-09-22 - Introducing Claude Sonnet 5.5 · Anthropic, 2026-09-28 - Introducing Claude Haiku 5.5 · Anthropic, 2026-10-07 - Claude Haiku 5.5: overview and pricing · Anthropic documentation, 2026-10-07 - Anthropic releases Opus 5.5 with lower prices and Fable-level performance · TechCrunch, 2026-09-22 - Anthropic Releases Claude Opus 5.5 · MarkTechPost, 2026-09-22 - We Must Pace the Frontier · Dario Amodei, 2026-09 - Update on access to Claude Fable 5 and Claude Mythos 5 · Anthropic, 2026-06-12 - GPT-6 Astra · OpenAI, 2026-09-03 - Text leaderboard · Arena, 2026-10-08 - AI model leaderboards (Intelligence Index) · Artificial Analysis, 2026-10-08 - Claude Opus 5.5: intelligence, performance and price · Artificial Analysis, 2026-10-08 - Opus 5.5 vs Opus 5 cost to run the Intelligence Index (post on X) · Artificial Analysis, 2026-09 - Terminal-Bench leaderboard · Terminal-Bench, 2026-09-26 - Terminal-Bench 4.0 leaderboard (mirror of tbench.ai) · Snorkel AI, 2026-10-08 - ARC-AGI leaderboard · ARC Prize, 2026-09-26 Editorial note: this analysis reflects the public information available on the review date. Models, prices and rules change fast; every third-party figure links to its source, and our opinions are labeled as such. Spotted an error? Write to contact@slash-digital.io. Frequently asked questions ## The questions we hear often **When was Claude Opus 5.5 released?** On September 22, 2026. Anthropic commits to keeping it available until at least September 22, 2027. **How much does Claude Opus 5.5 cost?** $4 per million input tokens and $20 per million output tokens; cache reads cost $0.20 per million and the Batch API is 50% off. **Is it better than Claude Fable 5.1?** According to Anthropic it performs at that level on most work for less than half the price. Fable 5.1 remains recommended for the most demanding reasoning. Only your evaluation confirms it for your case. **Is Claude Sonnet 5.5 the better buy?** It depends on the task. Sonnet 5.5 (September 28, 2026) costs $2 / $10, half the price of Opus 5.5, and Anthropic says it beats Opus 5.5 on Terminal-Bench 4.0 (70.6% vs 66.4%) while trailing it on every other row of its table. On the official Terminal-Bench 4.0 board, Opus 5.5 scores 64.8% and Sonnet 5.5 61.8%. Test both on your prompts. **Can I disable thinking to save money?** No. Thinking is always on; to save, lower the effort parameter to low or medium. **Is it available from Colombia and the EU?** Yes, through the Claude API and through AWS, Google Cloud and Microsoft. Check the available regions and your data-protection obligations before sending personal data. --- # Best AI models, October 2026: Opus 5.5, GPT-6, Gemini URL: https://slash-digital.io/en/insights/ai-models-2026/ Models · October 2026 Between September 1 and 22, 2026, Anthropic, OpenAI, Google and SpaceXAI released at least seven models, and at least five more arrived by October 8. This map sorts out what changed, what each tier costs and what the leaderboards really measure, with dates and sources, so you can build a shortlist on solid ground. $10 / $50 the new premium tier1M tokens of context, the normOct 8, 2026 snapshot date Start a project All insights Jean MartialCEO and co-founder, Slash Digital Lab PublishedSeptember 26, 2026 ReviewedOctober 8, 2026 Reading time9 min TopicModels In short As of October 2026, three US labs set the pace: Anthropic with Claude Fable 5.1, Opus 5.5 and Sonnet 5.5, OpenAI with GPT-6 Astra, Sol, Luna and GPT-6.1 Sol, and Google with Gemini 4 Argon, still in limited access, while Gemini 3.5 Pro stays delayed. Grok 4.7, Meta's closed Muse Spark and the open-weight labs complete the map. Prices now fall into clear bands, from $10/$50 per million tokens down to $0.30/$1.20, and leadership depends on the test. Our advice: shortlist by use case, keep a fallback from a second vendor and decide with your own prompts in Spanish and French. For your company Choosing a model or an AI architecture for your company? Slash Digital Lab helps you decide with your own data, costs and constraints. AI consulting → ## The closed labs and their flagships Who sells what, since when and at what list price. | Lab | Model | Released | Input / output | Context | | --- | --- | --- | --- | --- | | Anthropic | Claude Fable 5.1 | Sept 1, 2026 | 10 / 50 | 1M | | Anthropic | Claude Opus 5.5 | Sept 22, 2026 | 4 / 20 | 1M | | Anthropic | Claude Sonnet 5.5 | Sept 28, 2026 | 2 / 10 | 1M | | Anthropic | Claude Haiku 5.5 | Oct 7, 2026 | 0.10 / 0.50 up to 100K tokens; 0.50 / 2.50 above | 1M | | OpenAI | GPT-6 Astra | Sept 3, 2026 | 10 / 50 | 1.05M | | OpenAI | GPT-6 Sol | Sept 22, 2026 | 2 / 10 | 1.05M | | OpenAI | GPT-6.1 Sol | Sept 29, 2026 | 2 / 10 | unconfirmed | | OpenAI | GPT-6 Luna | Sept 22, 2026 | 0.10 / 0.50 | 1.05M | | Google | Gemini 4 Argon | Sept 30, 2026, limited access | 2 / 10 introductory, then 4 / 20 | 1M | | Google | Gemini 3.8 Flash | Sept 2, 2026 | 0.75 / 3.75 until Dec 31, 2026, then 1.50 / 7.50 | 1M | | Google | Gemini 3.1 Pro (preview) | Feb 19, 2026 | 2 / 12 up to 200K tokens | 1M | | SpaceXAI | Grok 4.7 | Sept 21, 2026 | 2 / 6 up to 200K tokens | 500K | Anthropic. Claude Fable 5.1 is its most capable widely released model; its twin, Mythos 5.1, is limited to a small set of vetted organizations. Anthropic's docs still recommend starting with Opus 5.5 for most workloads (details in our Opus 5.5 analysis). With Sonnet 5.5 and Haiku 5.5 out, Sonnet 5 and Haiku 4.5 are now legacy models, still available. OpenAI. GPT-6 Astra (see our analysis) is its first model rated "Critical" for cybersecurity under its Preparedness Framework, so the public version refuses advanced offensive tasks. GPT-6 Sol and Luna launched at half the promotional price of their GPT-5.6 predecessors; GPT-5.6 Terra ($2/$12) remains the middle option and the o-series is now legacy. On September 29 OpenAI launched GPT-6.1 Sol, which it says nearly matches Astra on agentic coding, computer use and professional work at one fifth of the price. Google. On September 30 it announced Gemini 4 Argon, its new flagship, for now only for trusted defenders in its Fairwind Program; API access is coming "soon", Google says. Gemini 3.5 Pro, expected in June, is still missing: Bloomberg reported on July 16 that its coding performance fell short. Meanwhile a new Flash model shipped almost every month (3.5 in May, 3.6 in July, 3.7 in August, 3.8 in September), and Gemini 3.1 Pro is still in preview. SpaceXAI. SpaceX absorbed xAI in February 2026, renamed it in July and bought the code editor Cursor in August. Grok 4.7 has slipped from sixth to tenth on the official Terminal-Bench 4.0 board (37.6%). Meta, Microsoft, Amazon. Muse Spark (April 8) was Meta's first model without open weights; version 1.3 scores 48 on the Artificial Analysis index. Microsoft unveiled MAI-Thinking-1, its first reasoning model, in June (price unconfirmed). Amazon moved Nova Premier and other Nova models to maintenance mode, according to Business Insider (via eWeek). SnapshotFigures checked on September 25 and 26, 2026, and updated on October 8 against the sources below. Prices, rankings and even availability change monthly: check again before signing a contract. ## Open weights in brief Open-weight models have grown huge (Kimi K3 has 2.8 trillion parameters, 104 billion active per token) but still trail the closed leaders: the best, Xiaomi's MiMo-V2.6-Pro, scores 46 on the Artificial Analysis index against 58 for Claude Opus 5.5. Mistral unveiled Mistral Large 4, a 1-trillion-parameter model, on October 6, for now API-only and without published benchmarks; it plans to release the weights about three weeks later. DeepSeek V4.1-Flash (MIT license) is the cheapest near-frontier API we found, at $0.30/$1.20 at peak hours and half that off-peak, and licenses are drifting toward custom terms with revenue thresholds. More in our open-weight guide. ## Price bands: four tiers, not one market The list price tells you the band; your cost per task tells you the bill. | Band | Typical price | Examples | Fits | | --- | --- | --- | --- | | Premium | 10 / 50 | Claude Fable 5.1, GPT-6 Astra | Hardest reasoning, long-horizon agents | | Upper mid | 4 / 20 | Claude Opus 5.5, Gemini 4 Argon (after its introductory price), GPT-5.6 Sol (promotional) | Default for agents and knowledge work | | Mid | about 2 / 10 | Claude Sonnet 5.5, GPT-6 Sol, GPT-6.1 Sol, GPT-5.6 Terra (2 / 12), Grok 4.7 (2 / 6) | Volume with solid quality | | Budget | 1 / 5 or less | Claude Haiku 5.5 and GPT-6 Luna (0.10 / 0.50), Gemini 3.8 Flash (0.75 / 3.75) | Classification, extraction, routing | | Cheapest near-frontier | 0.30 / 1.20 | DeepSeek V4.1-Flash, half price off-peak | Batch work on data that may leave your perimeter | The $10/$50 tier is a 2026 novelty (Claude Fable 5 in June, then Fable 5.1 and GPT-6 Astra), while the middle got cheaper: GPT-6 Sol costs half its predecessor's promotional price, Opus 5.5 is 20% cheaper per token than Opus 5, and Sonnet 5 kept $2/$10 instead of rising to $3/$15. Three details move a real bill more than the list price: long prompts (above 272K input tokens OpenAI bills the whole request at twice the input and cache rates and 1.5 times the output rate; Claude has no surcharge except on Haiku 5.5 above 100K tokens), cache reads (5% of the input price on Opus 5.5, Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon, 10% on GPT-6) and scheduled rises (Gemini 3.6 to 3.8 Flash double on January 1, 2027). Method in our AI cost guide. To learn how we put these models to work in each business, explore our AI for companies services. ## Leaderboards: what they measure, who leads Each ranking answers a narrower question than "which model is best". Vendor tables are run or chosen by the lab itself, usually at its highest effort and on benchmark versions a few weeks old. Independent boards apply one method to everyone but lag behind and give different numbers: Anthropic's reported 66.4% for Opus 5.5 on Terminal-Bench 4.0 was not yet on the official board on September 26, which showed 64.8% on October 8. Versions also break comparisons: OpenAI's Astra post quotes Artificial Analysis index v4.1.1 values of up to 65.7, while v4.3.2 tops out at 58, a new scale rather than weaker models. Arena, for its part, ranks by human preference votes, not by tasks completed. | Board | What it measures | Leader | Worth knowing | | --- | --- | --- | --- | | Arena, text | Human preference votes | Gemini 4 Argon (1,525, preliminary) | Claude Opus 5.5 is 2nd (1,507); GPT-6 Astra fell to 35th | | Artificial Analysis index v4.3.2 | Ten agentic and knowledge-work evaluations | Claude Opus 5.5, max effort (58) | Sonnet 5.5 follows (56); Fable 5.1, GPT-6 Astra and Gemini 4 Argon tie at 53 | | Terminal-Bench 4.0, official | Agentic tasks in a terminal | Claude Opus 5.5 (64.8%) | Sonnet 5.5: 61.8%; GPT-6 Astra: 58.2% | | ARC-AGI-2 | Abstract reasoning, with cost per task | GPT-6 Astra (95.0%, $1.12 per task) | Opus 5.5 at high effort: 93.3% for $0.408 | | Humanity's Last Exam | Expert-level questions | Claude Opus 5.5 (61.4%, measured by Artificial Analysis) | Anthropic reports 67.7% with tools | None of these tests uses business Spanish or French: they filter a shortlist, they don't decide it. Our method is in the evals guide. ## How to choose by use case A shortlist to test, not a verdict. | Use case | Shortlist to test | Why | | --- | --- | --- | | Customer service in Spanish | Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 3.8 Flash; GPT-6 Luna or Claude Haiku 5.5 for routing | Volume pricing; no public benchmark measures Latin American service Spanish, so test tone, accuracy and refusals | | Coding agents | Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra; GLM-5.3 if you need open weights | Leaders on Terminal-Bench 4.0; GLM-5.3 is the best open-weights entry on the official board | | Long documents | Claude Opus 5.5, GPT-6 Astra, Gemini 3.8 Flash | 1M-token windows; OpenAI reports 96.3% for Astra on MRCR v2 at 512K to 1M tokens; no long-context surcharge on Claude except Haiku 5.5 | | Voice | Gemini 3.8 Live, Amazon Nova 2 Sonic, Mistral Voxtral for transcription | Live audio models; Voxtral Mini Transcribe Realtime has Apache 2.0 weights (see our voice guide) | | Regulated or on-premise data | Self-hosted Gemma 4, Muse Glimmer, Qwen3.8-27B or Mistral Small 4 | Apache 2.0 licenses, and no data leaves your infrastructure | | High-volume classification | GPT-6 Luna, Claude Haiku 5.5, Gemini 3.1 Flash-Lite, DeepSeek V4.1-Flash | From $0.10 per million input tokens, half that in batch; Haiku 5.5 costs five times more on prompts over 100K tokens | Include a second vendor and an open-weight option in every shortlist. Our method (30 fixed prompts per client, five engines per run, three languages) is described in how we choose models. ## Seven verified trends of 2026 - A premium tier, yet cheaper capability. On ARC-AGI-2, a score of about 61 to 65% cost $2.25 per task with Claude Opus 4.6 in February and $0.042 with DeepSeek V4 Flash in July. - Reasoning effort is a setting. Claude, GPT-5.6, GPT-6 and Gemini 3.x expose effort levels; on ARC-AGI-2 on September 26, Opus 5.5 cost $0.241 per task at low effort and $1.85 at max (see our reasoning guide). - 1M tokens of context is the norm for flagships; Grok 4.7 stops at 500K. - Offensive cyber capability is gated. Mythos 5.1 is limited to a small set of vetted organizations, Gemini 3.8 Flash Cyber and Gemini 4 Argon without cyber guardrails to trusted defenders in Google's Fairwind Program, and OpenAI plans wider defensive access to Astra through its Daybreak program. - Products disappear quickly. OpenAI closed the Sora app in April, about seven months after launch, removed the Sora 2 API on September 24 and shut its Atlas browser on August 9; Nova Premier reached end of life on September 14. - US policy can cut access. A US export-control directive made Anthropic pull Fable 5 and Mythos 5 for all users on June 12 (Fable 5 returned on July 1), and Executive Order 14409 (June 2) calls for a voluntary framework giving the government up to 30 days of pre-release access to "covered frontier models". Since September 29 Washington has added a voluntary pledge by six AI companies (White House Accord on Super Intelligence), an executive order on "super intelligence", an FTC probe into rogue AI agents and a "Super Intelligence Force" (October 4). - Labs debate a slower pace. In We Must Pace the Frontier, Dario Amodei called for slowing capability gains and committed Anthropic to embedded third-party evaluators; Sam Altman matched that commitment on September 14. One trigger he cites: the July incident in which OpenAI agents under evaluation reached Hugging Face systems (OpenAI's report). On September 28 OpenAI dropped GPT-6.1 Astra, saying it fell short on scope, authorization and reporting back to the user. ## What it means for companies in Colombia, Latin America and Europe Keep a tested fallback. The Fable 5 withdrawal lasted 19 days and hit every customer. A production system in Bogotá, Mexico City or Paris needs a second model from another vendor, tested on the same prompts, and an open-weight option you could host. Budget for residency. OpenAI adds 10% for regional processing on models released since March 5, 2026; Claude costs 10% more on regional Bedrock or Google Cloud endpoints. Put retirement dates in contracts. Anthropic commits to Opus 5.5 until at least September 22, 2027, but to Haiku 4.5, now legacy, only until October 15, 2026. Ask every vendor for such dates, notice periods and a migration clause. Test Spanish and French yourself. At Slash AI Lab every candidate goes through 30 fixed prompts per client in Spanish, English and French before deployment; no leaderboard replaces that step. Check the reach of the EU AI Act. Article 2(1)(c) covers providers and deployers outside the EU when their system's output is used in the EU, so a Colombian firm serving French clients can be in scope. In Colombia, check how Law 1581 of 2012 applies before sending personal data abroad. More in our 2026 regulatory map. Our recommendationFor each route, name a default model, a fallback from another vendor and an open-weight option, then re-run your evaluation at every major release: in 2026, roughly once a month. ### Key takeaways - September 2026 brought Claude Fable 5.1, Opus 5.5 and Sonnet 5.5, GPT-6 Astra, Sol and Luna, GPT-6.1 Sol, Gemini 3.8 Flash, Gemini 4 Argon and Grok 4.7; October, Claude Haiku 5.5 and Mistral Large 4. Gemini 3.5 Pro is still missing. - Prices form bands, from $10/$50 for the premium tier down to $0.30/$1.20 for DeepSeek V4.1-Flash. - Leadership depends on the test: Anthropic leads the Artificial Analysis index and the official Terminal-Bench 4.0 board, Google tops Arena text (preliminary score) and OpenAI leads ARC-AGI-2. - Shortlist per use case, then decide with your own Spanish and French prompts. - Plan for access shocks and closures: a fallback from a second vendor, retirement dates in contracts and residency surcharges in the budget. ## Sources - Models overview · Anthropic documentation, 2026-10-08 - Introducing Claude Opus 5.5 · Anthropic, 2026-09-22 - Introducing Claude Sonnet 5.5 · Anthropic, 2026-09-28 - Introducing Claude Haiku 5.5 · Anthropic, 2026-10-07 - Claude Haiku 5.5: overview and pricing · Anthropic documentation, 2026-10-07 - Update on access to Claude Fable 5 and Claude Mythos 5 · Anthropic, 2026-06-12 - GPT-6 Astra · OpenAI, 2026-09-03 - Introducing GPT-6 Sol and GPT-6 Luna · OpenAI, 2026-09-22 - API pricing · OpenAI, 2026-10-08 - Gemini API pricing · Google AI for Developers, 2026-09 - Gemini 4 Argon · Google, 2026-09-30 - Gemini 3.5 Pro delayed over coding, Bloomberg reports · Search Engine Journal, 2026-07-16 - Models and pricing · xAI documentation, 2026-09 - Text leaderboard · Arena, 2026-10-08 - AI model leaderboards (Intelligence Index) · Artificial Analysis, 2026-10-08 - Terminal-Bench 4.0 leaderboard (mirror of tbench.ai) · Snorkel AI, 2026-10-08 - ARC-AGI leaderboard · ARC Prize, 2026-09-26 - OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less · TechCrunch, 2026-09-29 - OpenAI scraps release of latest AI model over safety concerns · Al Jazeera, 2026-09-29 - Mistral's new 1T model aims to leapfrog closed and open rivals · TechCrunch, 2026-10-06 - White House announces AI responsibility accord, executive orders and Super Intelligence task force · Crowell & Moring, 2026-10-07 - US regulator launches probe into AI companies · Al Jazeera, 2026-09-30 Editorial note: this analysis reflects the public information available on the review date. Models, prices and rules change fast; every third-party figure links to its source, and our opinions are labeled as such. Spotted an error? Write to contact@slash-digital.io. Frequently asked questions ## The questions we hear often **What is the best AI model in September 2026?** There is no single answer. On October 8, 2026, Claude Opus 5.5 led the Artificial Analysis index and the official Terminal-Bench 4.0 board, Gemini 4 Argon topped the Arena text board with a preliminary score and GPT-6 Astra had the best ARC-AGI-2 score (as of September 26). The best model for you is the one that wins on your own tasks at a cost you can defend. **How much does a frontier model cost?** $10 per million input tokens and $50 per million output tokens for Claude Fable 5.1 and GPT-6 Astra, $4/$20 for Claude Opus 5.5 (and for Gemini 4 Argon once its $2/$10 introductory price ends) and about $2/$10 for Claude Sonnet 5.5, GPT-6 Sol and GPT-6.1 Sol. Budget models start at $0.10/$0.50 with GPT-6 Luna and Claude Haiku 5.5 (the latter for prompts up to 100K tokens). **Has Google released Gemini 3.5 Pro?** Not as of October 8, 2026. Bloomberg reported a delay in July. Google's newest model is Gemini 4 Argon, announced on September 30 and for now limited to trusted defenders in its Fairwind Program; Google promises access "soon" for paid API customers and Google AI Ultra subscribers. Gemini 3.1 Pro is still in preview. **Are open-weight models good enough for business?** For many tasks, yes, and they are the basis of any on-premise option. On independent boards the best of them still trail the closed leaders (46 against 58 on the Artificial Analysis index). Our open-weight guide covers licenses and hosting. **Can companies in Colombia and France use these models?** Yes, through the vendors' APIs and the main clouds, but plan for residency surcharges, data-protection rules and the possibility that US policy restricts a model, as happened with Claude Fable 5 in June 2026. --- # Local AI hardware in 2026: what to buy, what runs URL: https://slash-digital.io/en/insights/local-ai-hardware-2026/ Local AI and open weights · September 2026 What you can run locally depends less on the brand than on two numbers: how much memory the model gets and how fast that memory can be read. Here are verified specs and prices, the memory formula with a worked example, and our picks by team size. 1,792 GB/s RTX 5090 and RTX PRO 6000 bandwidth512 GB maximum unified memory, Mac Studio M5 Ultra$1,999 RTX 5090 launch price Start a project All insights Rémy SanchezCTO and co-founder, Slash Digital Lab PublishedSeptember 26, 2026 Reading time8 min TopicLocal AI and open weights In short Memory capacity sets which model fits and memory bandwidth sets how fast it answers: single-user generation tops out near bandwidth divided by the bytes read per token. Mixture-of-experts (MoE) models read only their active parameters, so a 128 GB DGX Spark runs gpt-oss-120b at about 61 tokens per second but a dense 70B model at 2.7. Our picks: a 32 GB RTX 5090 ($1,999 at launch) for one developer, 96 GB RTX PRO 6000 cards for a team, and rented GPUs before buying for production. Buying pays only at high, steady volume or under strict residency rules. For your company Which model handles your real cases best, in the cloud or local? At Slash AI Lab we compare models on your own prompts before you commit budget or private data. Slash AI Lab → ## Why memory decides: capacity, then bandwidth Capacity decides whether a model fits: its weights plus the working memory of the conversation. Bandwidth decides how fast it answers, because every generated token means reading the weights again; NVIDIA notes that in this decode phase memory transfer, not compute, dominates latency. Rule of thumb: single-user speed tops out near bandwidth divided by the bytes read per token, and real systems reach only part of it (Databricks measures that share as bandwidth utilization). With the same 7B model at 4-bit, an M5 Pro (307 GB/s) generates about 66 tokens per second and an M5 Max (614 GB/s) about 120, in community benchmarks. Reading the prompt (prefill) depends more on compute than on bandwidth. Mixture-of-experts models read only their active parameters. On the same DGX Spark (273 GB/s), LMSYS measured a dense Llama 3.1 70B in FP8 at 2.7 tokens per second; llama.cpp benchmarks show gpt-oss-120b, with 5.1B active parameters, at about 61. Large-memory, modest-bandwidth machines suit MoE models, not big dense ones. To find out which model performs for your use case before you buy hardware, Slash runs model evaluation at Slash AI Lab, including local inference. ## The memory formula, with a worked example Memory needed is roughly the weights, plus the KV cache (the per-token memory of the conversation), plus a margin: weights (GB) ≈ parameters (billions) × bits per weight ÷ 8 KV cache per token (bytes) = 2 × layers × KV heads × head dimension × bytes per value total ≈ weights + KV cache per token × context tokens + about 20% margin Bits per weight: 16 in BF16, about 8.5 for Q8_0 and about 4.8 for Q4_K_M (per IBM's Granite 4.2 files). The margin is EleutherAI's heuristic; the KV cache comes from the model's config.json. | Item | 131K-token context | 262K-token context | | --- | --- | --- | | Weights at Q4_K_M | 16.5 GB | 16.5 GB | | KV cache at 16 bits (64 KiB per token) | 8.6 GB | 17.2 GB | | Total before margin | 25.1 GB | 33.7 GB | | Fits in 32 GB? | Yes, with room to spare | No: 8-bit KV cache (8.6 GB) or a larger card | Only 16 of the model's 64 layers use full attention, hence 64 KiB per token; the older Qwen3-32B needs 256 KiB, or 32 GiB at 131K tokens, more than the whole card. According to Ollama, an 8-bit KV cache halves that memory with a very small loss in precision. Rule of thumbStay at 4 bits or more for production: a Cohere study found that automatic metrics hid most of the loss in French (-0.3% on benchmarks, -16.6% in human evaluation, for a 4-bit 103B model). More in our quantization guide. ## Hardware tiers in September 2026 Vendor specifications and US launch prices. | Hardware | Memory | Bandwidth | Launch price (USD) | | --- | --- | --- | --- | | GeForce RTX 5060 Ti 16 GB | 16 GB | 448 GB/s | 429 | | GeForce RTX 5070 Ti | 16 GB | 896 GB/s | 749 | | GeForce RTX 5090 | 32 GB | 1,792 GB/s | 1,999 | | RTX PRO 6000 Blackwell (Workstation or Max-Q) | 96 GB, ECC | 1,792 GB/s | Not published | | DGX Spark (GB10) and OEM versions | 128 GB unified | 273 GB/s | Varies by vendor | | PCs with Ryzen AI Max+ | Up to 128 GB unified | Not in AMD's announcement | Varies by vendor | | Mac Studio M5 Max | 36 to 128 GB unified | 460 or 614 GB/s | From 2,499 | | Mac Studio M5 Ultra | 96 to 512 GB unified | 1,200 GB/s | From 5,499 | The RTX 5090 and RTX PRO 6000 share 1,792 GB/s and generate at similar speed on models that fit both (about 205 and 215 tokens per second with gpt-oss-20b, per LMSYS); the PRO card adds three times the memory, with ECC, and the RTX PRO 5000 (48 or 72 GB, 1,344 GB/s) sits in between. DGX Spark, Ryzen AI Max+ PCs and Macs trade bandwidth for capacity; only the M5 Ultra approaches a discrete GPU. Status: the M5 Mac Studio shipped on September 22, 2026, the new M6 starts in the Mac mini at up to 32 GB, and the RTX 5090 Laptop GPU has 24 GB. Above the desktop, Microsoft announced a Windows DGX Station (GB300) for the fourth quarter of 2026 that it says runs models of up to 1 trillion parameters. In datacenters: NVIDIA's H100 (80 GB), H200 (141 GB), B200 and B300, and AMD's Instinct MI300X (192 GB) and MI355X (288 GB). ## What runs on each tier Published file sizes; add the KV cache and a margin. | Tier | Examples that fit | What to expect | | --- | --- | --- | | 16 GB GPU | gpt-oss-20b (12.1 GB), Gemma 4 12B QAT (7.0 GB) | Fast small models | | 32 GB GPU (RTX 5090) | Qwen3.8-27B Q4_K_M (16.5 GB), Gemma 4 31B QAT (17.7 GB), Muse Glimmer Q4_K_M (16.8 GB) | Single-user sweet spot | | 96 GB GPU (RTX PRO 6000) | gpt-oss-120b (63.4 GB), Mistral Small 4 at about 4-bit (71.8 to 73.8 GB), Qwen3.5-122B-A10B Q4_K_M (76.5 GB) | Full speed; serves a team | | 128 GB unified (DGX Spark, Ryzen AI Max+, M5 Max) | Also Nemotron 3 Super NVFP4 (84.3 GB) and Step-3.5-Flash at 4-bit (111.5 GB, tight) | MoE at moderate speed; a dense 70B crawls | | 256 to 512 GB unified (M5 Ultra) | Qwen3.5-397B-A17B Q4_K_M (294.1 GB), Kimi K2.5 at 2-bit (375 GB) | Low concurrency; below 4 bits, test tool calling | | 8-GPU server (H100, H200, B200 class) | GLM-5 family (8 GPUs per Z.ai), Mistral Large 3 (one 8xH100 node) | Frontier class; Kimi K3 needs 64+ accelerators | Fitting is not speed: on 128 GB machines, prefer models with few active parameters. And vendor claims are ceilings: NVIDIA and AMD cite up to 200 billion parameters for DGX Spark and Ryzen AI Max+, already about 100 GB of weights at 4 bits. If you are looking for uncensored builds: they weigh the same as the original model, so they fit the same tiers. Our memory-tier map of uncensored models, from 32 to 512 GB, collects @sudoingX's picks with sizes verified on Hugging Face and explains the risks. ## Buy, rent or pay per token: a TCO method Compare the monthly cost of the same job at equal quality: hardware over its useful life, energy and maintenance hours, against the API bill for the same tokens. Illustrative example: hardware, energy and labor figures are our assumptions; API prices are list prices as of September 26, 2026. | Line | USD per month | | --- | --- | | RTX 5090 workstation, $3,500 over 36 months ($1,999 card plus $1,500 assumed) | 97 | | Energy: 0.7 kW, 176 hours, $0.15 per kWh | 18 | | Operations: 4 hours at $40 | 160 | | Local total | 275 | | Qwen3.8-27B on Groq ($0.80 / $4.00) | 128 | | gpt-oss-120b on Groq ($0.15 / $0.60) | 21 | | Claude Opus 5.5 ($4 / $20) | 640 | At this volume the local machine costs about twice the same model on Groq, and beats a frontier API only if a 27B model is good enough. Break-even with Groq comes near 170 million tokens a month with this mix, if a batching engine can serve that load. Operations, not the GPU, dominate: without them, local costs $115. Pan et al. find that a small model on one RTX 5090 pays back against commercial APIs in about three months, but that on-premises mainly pays above 50 million tokens a month or under strict residency rules. Peng et al. (2026) found that prompt caching made an API cheaper than a shared on-premises GPU, and GPU performance per dollar improves about 49% a year (Epoch AI). ## Power, noise and availability Power and noise. The RTX 5090 draws up to 575 W (NVIDIA recommends a 1,000 W supply) and the RTX PRO 6000 600 W, or 300 W in its Max-Q version with the same memory and bandwidth. DGX Spark runs on 240 W in a 15 cm box weighing 1.2 kg. We found no comparable noise data; our rule is that a 575 to 600 W card belongs in a ventilated room or a rack. Availability and software. DGX Spark also comes in OEM versions from Acer, ASUS, Dell, GIGABYTE, HP, Lenovo and MSI. Check format support: NVFP4 needs Blackwell GPUs, and MLX needs macOS 26.2 or later for the M5's accelerators. ## Buying from Colombia and Europe Colombia. Our prices are US launch prices; in Colombia, taxes, freight, the distributor's margin and the exchange rate move the final cost, so we publish no peso figure. Get quotes in USD and COP from at least two distributors, with warranty and delivery terms; prefer OEM versions with local invoicing and support, and check the circuit and UPS before installing a 1,000 W workstation. Renting is also legal: Law 1581 allows transfers to countries on the SIC's adequacy list, which includes the United States. Europe. Owned hardware or a SecNumCloud-qualified provider, which ANSSI designed against extraterritorial laws, answers the US CLOUD Act, which reaches data held by US providers wherever it is stored. For the models, see open-weight models in 2026. ## Our recommendations by team size Starting points, to check against your own models and volumes. | Profile | Hardware to start with | Models | Software | | --- | --- | --- | --- | | Solo developer | RTX 5090, or a 128 GB machine for larger MoE models | Qwen3.8-27B, Gemma 4 31B or Muse Glimmer at 4-bit; gpt-oss-120b on 128 GB | llama.cpp, Ollama or LM Studio | | Team of about 20 | Server with one or two RTX PRO 6000 (96 GB each) | gpt-oss-120b, Mistral Small 4 or Qwen3.5-122B-A10B at 4-bit | vLLM or SGLang behind authentication, with Open WebUI | | Production service | Rented 8-GPU nodes before buying | GLM-5.3, MiMo-V2.6-Pro or another large MoE | vLLM or SGLang, with Dynamo or llm-d | For a team, concurrency matters more than peak speed: in Red Hat's test on one A100, vLLM reached 793 tokens per second, while Ollama with default settings plateaued at 41. For production, rent first, measure real utilization and buy only if the GPUs stay busy; Lenovo's sponsored analysis puts the threshold near five hours of use a day against on-demand cloud. More in how to run a local LLM server. A local server is still a serverSentinelLABS and Censys counted 175,108 Ollama hosts exposed on the internet, and NVD lists dozens of 2026 vulnerabilities in vLLM, llama.cpp and Ollama. Add authentication, keep the port private and patch it like production. ### Key takeaways - Capacity decides which model fits; bandwidth decides generation speed. - MoE models read only active parameters: on a DGX Spark, gpt-oss-120b runs at about 61 tokens per second and a dense 70B at 2.7. - A 32 GB RTX 5090 ($1,999 at launch) runs Qwen3.8-27B at 4-bit with a 131K context; a 96 GB RTX PRO 6000 serves 120B-class MoE models to a team. - Count operations time: in our illustrative month it outweighs the hardware, and cheap open-model APIs often win at low volume. - Treat a local AI server as production infrastructure: authentication, no public exposure and regular patches. ## Sources - GeForce RTX 5090 · NVIDIA, 2025-01-06 - Compare GeForce graphics cards · NVIDIA, 2025 - RTX PRO 6000 Blackwell Workstation Edition · NVIDIA, 2025-03-18 - DGX Spark · NVIDIA, 2025-10-13 - Apple introduces new Mac Studio with M5 Max and M5 Ultra · Apple Newsroom, 2026-08-25 - AMD at CES 2026: Ryzen, Ryzen AI and ROCm announcements · AMD, 2026-01-05 - NVIDIA DGX Spark review · LMSYS, 2025-10-13 - llama.cpp performance on DGX Spark (discussion #16578) · ggml-org, 2026-02-05 - LLM inference performance engineering: best practices · Databricks, 2023-10-12 - A cost-benefit analysis of on-premise LLM deployment (Pan et al.) · arXiv, 2025-08-30 - Ollama vs. vLLM: performance benchmarking · Red Hat Developer, 2025-08-08 - Qwen3.8-27B model card · Qwen (Hugging Face), 2026-08-14 Editorial note: this analysis reflects the public information available on the review date. Models, prices and rules change fast; every third-party figure links to its source, and our opinions are labeled as such. Spotted an error? Write to contact@slash-digital.io. Frequently asked questions ## The questions we hear often **What is the best GPU for local AI in 2026?** For a single card, the RTX 5090 (32 GB, 1,792 GB/s, $1,999 at launch) gives the most speed per dollar in our sources. For models above about 30B, look at a 96 GB RTX PRO 6000 or a 128 GB unified-memory machine. **Can a DGX Spark run a 70B model?** It fits, but LMSYS measured Llama 3.1 70B in FP8 at 2.7 tokens per second on it. MoE models run far faster: about 61 tokens per second for gpt-oss-120b in llama.cpp. **How much VRAM does a 27B model need?** About 16.5 GB for Qwen3.8-27B weights at Q4_K_M, plus the KV cache: about 8.6 GB more for a 131K-token context at 16 bits. A 24 GB card works with shorter context; 32 GB leaves room. **Is a Mac good for local LLMs?** Yes for large models at low concurrency: the M5 Ultra Mac Studio offers up to 512 GB at 1.2 TB/s. For many simultaneous users, a GPU server with vLLM scales better. **Is local AI cheaper than an API?** Only at high, steady volume, against expensive frontier APIs or under strict residency rules. Cheap open-model APIs often win at low volume, and prompt caching can tip the balance. Run the numbers with your own volumes. --- # Uncensored AI models: what they are, uses and risks URL: https://slash-digital.io/en/insights/uncensored-models/ Local AI and open weights · October 2026 Stripping the safeguards from an open model now takes minutes. There are legitimate reasons to want fewer refusals, but the risks are documented and the law tightens in 2026. Here is what a CTO or CISO should know before using one. 3,471 uncensored models tracked25% of GitHub apps using them rated maliciousDec 2, 2026 EU nudifier ban applies Start a project All insights Rémy SanchezCTO and co-founder, Slash Digital Lab PublishedSeptember 26, 2026 ReviewedOctober 9, 2026 Reading time12 min TopicLocal AI and open weights In short An “uncensored” model is an open-weight model whose refusals were removed, by fine-tuning without refusals or by erasing the internal “refusal direction” identified in 2024. Over-refusal is a real, measured problem. But the change persists in every redistributed copy, can cost truthfulness and reasoning, and meets new law: the EU ban on nudifier and child-abuse generators applies from December 2, 2026. Our recommendation: most companies need a capable, well-calibrated model with an explicit policy, guard models, logging and human review, not a model without safeguards. When a low-refusal model is justified (authorized security work, research), treat it as a controlled lab tool. If you still need one, the memory-tier map (32 to 512 GB, October 2026), based on @sudoingX's picks and verified by us on Hugging Face, shows which build fits your hardware, how much it weighs and what each author claims. For your company Which model handles your real cases best, in the cloud or local? At Slash AI Lab we compare models on your own prompts before you commit budget or private data. Slash AI Lab → ## What “uncensored” means, technically One word, four mechanisms: each changes something different and calls for a different response. | Mechanism | What changes | Requires | Persists? | | --- | --- | --- | --- | | Fine-tuning without refusals | The weights, through new training | Weights or a fine-tuning API | Yes, in every copy | | Refusal-direction ablation (“abliteration”) | The weights, through a targeted edit | Weights (white-box access) | Yes, in every copy | | Base checkpoint | Nothing: never instruction- or safety-tuned | Weights | Not applicable | | Jailbreak | Only the inputs | Any access, hosted APIs included | No, one conversation at a time | Refusal-free fine-tunes came first. Community fine-tunes such as the Dolphin and Hermes families were tuned to refuse far less, on a “composable alignment” argument: publish a neutral model and let each deployer add its own policy. The original safety layer is thin anyway: Qi et al. (2023) removed most of GPT-3.5 Turbo's safety behavior with a handful of examples sent through OpenAI's fine-tuning API. Refusal-direction ablation comes from Arditi et al. (2024): in 13 open chat models of up to 72B parameters, refusal was mediated by a single direction in the model's activations. Erasing it suppresses refusals; adding it makes the model refuse harmless requests. Later work found a richer geometry, with several independent directions, and showed that a direction extracted in English suppressed refusals in other languages “with near-perfect effectiveness”, a direct concern for deployments in Spanish and French. A jailbreak changes only the inputs: it works against hosted APIs too, but must be repeated in each conversation. Instruct checkpoints learn to refuse in post-training; base checkpoints never did, so only safety built into the pretraining data protects them, and the Deep Ignorance study found it far more resistant to adversarial fine-tuning than post-training defenses. ## Why they exist: over-refusal is real The demand is not only malicious. Aligned models regularly refuse safe requests that merely look dangerous. In XSTest, 250 safe prompts built on homonyms, figurative language or historical events, Llama 2 70B Chat with its default system prompt fully refused 38% of them; GPT-4 refused 6.4%. OR-Bench tested 32 models on 80,000 seemingly toxic but safe prompts and found a rank correlation of 0.89 between refusing safe and toxic prompts: models tuned to be safer tend to over-refuse more, although newer models over-refuse less. Amazon's FalseReject showed that fine-tuning can reduce unnecessary refusals “without compromising overall safety”. The counterweight is SORRY-Bench, which checks that models still refuse 440 genuinely unsafe instructions across 44 categories; a useful model must do well on both sides. Anthropic reports that its 2026 classifiers refuse about 0.05% of harmless queries, 87% fewer than its first version. False refusals are an engineering problem with engineering fixes; removing safety is not one of them. If you want to measure both sides on your own cases, Slash runs model evaluation at Slash AI Lab. ## A large ecosystem that redistribution keeps alive Counts depend on the method, but they point the same way. Lin et al. (2025) identified more than 11,000 uncensored models on Hugging Face, one with over 19 million installs. With a stricter definition, 10a Labs counted 3,471 between January 2024 and March 2026, each repackaged 2.4 times on average, and 25% of the 1,643 GitHub applications that use them were classified as explicitly malicious. The same paper calls redistribution “the persistence layer”: compressed copies and mirrors across accounts, formats and platforms make upstream takedowns largely ineffective. Automation lowered the bar further. A Financial Times investigation with the safety firm Alice, published in May 2026, reported that a free, openly published tool had been used to create thousands of “decensored” variants, that reporters stripped Llama 3.3's safeguards in minutes without specialist hardware, and that Gemma 4's were removed within 90 minutes of its release. Google called the technique “a known technical challenge facing all open models”. We deliberately do not name the tool. ## What you lose when safeguards come off In Arditi et al., general benchmarks (MMLU, ARC, GSM8K) stayed close to baseline after the edit, but TruthfulQA accuracy “consistently drops”. A 2025 preprint comparing such edits on 16 instruction-tuned models from 7B to 14B found maths reasoning the most sensitive capability: GSM8K results ranged from a gain of about 1.5 points to a loss of 18.8, depending on method and architecture. The deeper risk is behavioral. The emergent misalignment study, extended in Nature in 2026, fine-tuned GPT-4o and Qwen2.5-Coder-32B on one narrow task, writing insecure code, and got broadly misaligned behavior on unrelated prompts; presenting the same data as material for a security class prevented it. The popular claim that they “hallucinate more” across the board is not established by any benchmark we could verify. The measured costs are reason enough to evaluate any modified model on your own tasks. ## Legitimate uses, and the Hugging Face lesson There are real reasons to want a model that does not refuse: authorized pentesting and red teaming, malware analysis and incident response, fiction about crime or violence, frank medical and legal questions, and research on refusal itself. The question is how to serve these uses safely. In July 2026, during an intrusion at Hugging Face run by an autonomous agent framework, commercial frontier models blocked the team's forensic requests because their guardrails “cannot distinguish an incident responder from an attacker”. The team analyzed more than 17,000 attacker action logs with the open-weight GLM-5.2 on its own infrastructure instead. The lesson: have a capable, vetted, self-hosted model ready before an incident, which also keeps attacker data and credentials in-house. Hugging Face adds that this is “not an argument against safety measures on hosted models”. The industry's answer is verified access, not removed safeguards. With Project Glasswing (April 2026), Anthropic restricted Claude Mythos Preview, strong at finding and exploiting vulnerabilities, to partners and more than 40 organizations that maintain critical software, with a Cyber Verification Program for security professionals. Specialized open models exist too: Cisco's Foundation-Sec-8B (Apache 2.0) targets SOC triage and threat simulation, excludes malware and phishing generation and requires human review. ## Which uncensored model fits your memory? The October 2026 map It depends on memory: from 32 to 512 GB there is an uncensored build of almost every open family, and since it weighs the same as the original, it runs at the same speed. On October 8, 2026, developer @sudoingX published a map of the most downloaded uncensored and abliterated builds for each memory tier, from one 32 GB RTX 5090 to four DGX Sparks with 512 GB, with speeds measured on their own machines. We summarize it here, credited to its author, and verified it: on October 9 we checked every repository on Hugging Face, every file size, the access terms and what each author claims. | Memory | Typical hardware | Top pick | Measured speed (base model) | | --- | --- | --- | --- | | 32 GB | RTX 5090, Radeon AI PRO R9700 or two 16 GB cards | Qwen 3.6 40B Deckard Heretic, Q5, 28.2 GB | 155.5 tok/s: 27B at Q4 with MTP, RTX 5090 (community) | | 48 GB | Two RTX 3090 or 4090, RTX A6000, RTX PRO 5000 or two RX 7900 XTX | Qwen 3.6 35B-A3B Genesis Hermes, Q8, 43.6 GB | 52.5 tok/s: 27B at Q4 with MTP, RTX A6000 (community) | | 96 to 128 GB | RTX PRO 6000 96 GB, AMD Strix Halo, DGX Spark or Mac Studio 128 GB | Qwen 3.8 Flash-Next 180B uncensored, IQ4_XS, 97.5 GB | 44.4 tok/s: Flash-Next at IQ4 with MTP, one DGX Spark | | 256 GB | Two DGX Sparks or Mac Studio 256 GB | GLM 5.3 Flash 321B uncensored, NVFP4, 194.7 GB | 49.5 tok/s: Flash-Next at FP8 with MTP, two DGX Sparks | | 512 GB | Four DGX Sparks or Mac Studio 512 GB | GLM 5.3 Flash uncensored, FP8, 328.3 GB | Not run yet | ### 32 GB: RTX 5090, Radeon AI PRO R9700 or two 16 GB cards - Qwen 3.6 40B Deckard Heretic: Q5, 28.2 GB. A 40B model expanded from a 27B and trained on a reasoning dataset generated with another model, according to its author. - Qwen 3.8 27B Aggressive: Q6, 25.9 GB, with an MTP head; 0 refusals out of 465, according to its author. - Gemma 4 31B Heretic: Q6, 25.2 GB; 10 refusals out of 100 versus 99 for the original, according to its author. - Qwen 3.6 35B-A3B Aggressive: Q6, 30.6 GB; 0 out of 465 and more than 11.9 million downloads all time. ### 48 GB: two RTX 3090 or 4090, RTX A6000, RTX PRO 5000 or two RX 7900 XTX - Qwen 3.6 35B-A3B Genesis Hermes: Q8, 43.6 GB. Built on the 0-of-465 Aggressive build, with data added from an agent fine-tune. - Llama 3.3 70B abliterated: Q4, 42.5 GB, the classic 70B. - Qwen3-Next 80B-A3B abliterated: IQ4_XS, 42.6 GB, a mixture of experts with 3B active. - With three RTX 3060s (36 GB), Gemma 4 31B Heretic fits at Q8 (32.6 GB), and Qwen 3.8 27B at Q4 runs at 17.3 tok/s, 24.5 with MTP (community). ### 96 to 128 GB: RTX PRO 6000, Strix Halo, DGX Spark or Mac Studio - Qwen 3.8 Flash-Next 180B uncensored: IQ4_XS, 97.5 GB, gated (you accept the terms on the page). - Flash-Next GSQ-RCO abliterated: IQ3_S, 83.9 GB, fits a 96 GB card; the refusal direction was removed from an already-quantized model. - Qwen 3.5 122B-A10B Aggressive: Q4, 74.2 GB; 0 out of 465, according to its author. - DeepSeek V4 Flash abliterated: Q2, 86.7 GB; its author presents it as a proof of concept. - gpt-oss-120b Aggressive: 65.4 GB in MXFP4, the model's native precision. - Measured by @sudoingX: a Strix Halo runs Qwen 3.6 35B-A3B at Q8 at 53.5 tok/s; one DGX Spark runs the base 122B at 28.2 tok/s (34.9 with MTP) and DeepSeek V4 Flash at IQ3 at 16.5 tok/s. ### 256 GB: two DGX Sparks or Mac Studio 256 GB - GLM 5.3 Flash 321B uncensored: NVFP4, 194.7 GB, keeps the MTP head; 0% refusals on HarmBench-320 and 85.28% MMLU, within a point of the base model, according to its author. - Qwen 3.8 Flash-Next uncensored: FP8, 185.5 GB, gated. - DeepSeek V4 Flash abliterated: Q4, 164.6 GB. ### 512 GB: four DGX Sparks or Mac Studio 512 GB - GLM 5.3 Flash uncensored: FP8, 328.3 GB, gated. - dealignai's FP8 build: 328.3 GB; 87.33% MMLU versus 86.74% for the base model, according to its author. Three things to read into the map. Demand is huge. According to Hugging Face, the Aggressive build of Qwen 3.8 27B logged more than 2 million downloads in 30 days, and the abliterated Flash-Next build, published on October 2, was already past 680,000. Refusal figures are self-reported. Each author uses its own test set (465 prompts, 100 prompts or HarmBench-320), and some admit keeping refusals on purpose to protect quality: none of these figures is an independent evaluation. Not everything is a clean ablation. Deckard and Genesis Hermes mix in other training, so they no longer behave like the original model. The license travels with the base model. The Qwen 3.8 Flash-Next derivatives declare Apache 2.0 on their repositories, but the base model uses the Qwen Community license, with its own conditions for some businesses (see open-weight models in 2026): check the original license before any commercial use. In practice, GGUF files run on llama.cpp (with Vulkan or ROCm on AMD) and on a Mac with llama.cpp or MLX where a build exists. To pick hardware, see local AI hardware in 2026; to understand Q4, Q6 or IQ4_XS, the quantization guide. No guardrails means every answer is yoursThat is how the map's author puts it, and it is the rule. If you truly need one of these models, run it isolated, with no customer data, with logging and with your own guard model on the output, and read the legal section below first. For almost everything else our recommendation stands: a well-calibrated model with a written policy. ## What the law and the licenses say in 2026 Most rules target abusive tools and outputs, above all sexual imagery, not models as such. | Jurisdiction | Rule | What it targets | Key date | | --- | --- | --- | --- | | United States | TAKE IT DOWN Act | Publishing non-consensual intimate images of a real person, AI forgeries included; 48-hour removal duty for platforms | Signed May 19, 2025 | | Texas | TRAIGA (HB 149) | Developing or deploying AI with intent to produce child sexual abuse material or unlawful sexual deepfakes | January 1, 2026 | | United Kingdom | Crime and Policing Act 2026, section 99 | Making, adapting or supplying tools that generate intimate images; up to 3 years in prison | Enacted 2026 | | European Union | AI Act, Art. 5(1)(ba) and (bb) | Systems generating non-consensual intimate images or child sexual abuse material; fines up to €35 million or 7% of turnover | December 2, 2026 | | France | Code pénal, art. 226-8-1 | Sharing a person's sexual content without consent, AI-generated included: up to 2 years and €60,000 (3 years and €75,000 online) | In force since May 23, 2024 | | Colombia | Ley 2502 de 2025 | Impersonation using AI, deepfakes included: the fine rises by up to one third | July 28, 2025 | The EU ban also reaches general-purpose tools: placing one on the market is prohibited where such output is a reasonably foreseeable and reproducible outcome and the system lacks reasonable and adequate safeguards to prevent it reliably. An image model whose safeguards were stripped fits that description. Article 50 transparency duties, such as telling people they are interacting with an AI, apply since August 2, 2026, whatever the model's size or license. Licenses add a layer: Llama 3.3's acceptable use policy bans child exploitation content and malicious code, and models under the Gemma Terms of Use forbid “attempts to override or circumvent safety filters” (Gemma 4 moved to Apache 2.0). Hugging Face's content policy prohibits non-consensual sexual content. More in our 2026 regulatory map. ## Better options than an uncensored model For almost every legitimate need there is a safer path. | Need | Better option | | --- | --- | | Pentest, red teaming, malware analysis | Verified access, security-specialized models and an explicit authorization context | | Incident response on attacker data | A capable open-weight model, vetted and self-hosted before the incident | | Content moderation, trust and safety | Guard models built to read harmful content and classify it against your policy | | Fiction, medical or legal questions refused | A better-calibrated model and a written content policy, with false refusals measured | | Too many refusals in a customer chatbot | Adjust the system policy and the guard thresholds, not the model's safety | | Research on refusal and alignment | Open weights in an isolated lab, under a research protocol, never redistributed | The common thread: fix calibration and access, not safety. Guard models and policies are covered in guardrails that work in production, and authorized offensive work in AI in offensive security. ## Our recommendation and a checklist In Slash we recommend three rules: no model with removed safeguards in client-facing products; for authorized security work, verified access or a security-specialized model first; and a low-refusal model only as a controlled lab tool, never redistributed. Before using one, check: - Authorization: a written scope, a named owner and a legitimate purpose (test, research, incident response). - Provenance: weights from a known publisher, with checksums; anonymous copies are a supply-chain risk. - Isolation: a separate environment with no production data, no internet egress and no unneeded tools. - Logging: prompts, outputs, operator and model version, under your retention policy. - Output review: a guard model on outputs and a person before anything leaves the lab. - Legal check: the license's acceptable use policy and the law where you operate and where outputs go. - Exit: delete the model and its outputs when the engagement ends. Red lineNever point a low-refusal image or video model at photos of real people: that is where legal exposure concentrates in every jurisdiction in the table, and in the EU such tools are banned from December 2, 2026 when they lack adequate safeguards. ### Key takeaways - “Uncensored” covers four mechanisms (refusal-free fine-tunes, refusal-direction ablation, base checkpoints, jailbreaks); weight edits persist in every copy. - Over-refusal is real (Llama 2 70B Chat fully refused 38% of XSTest's safe prompts) and is fixed by better calibration, not by removing safety. - Takedowns fail: 10a Labs counted 3,471 uncensored models, each repackaged 2.4 times, and 25% of the GitHub apps using them were rated malicious. - Removing safeguards has measured costs: lower TruthfulQA accuracy, reasoning losses of up to 18.8 points and a risk of emergent misalignment. - The EU bans nudifier and child-abuse generators from December 2, 2026; for legitimate needs, use verified access, specialized models and a controlled lab. - If you need one, there is a build for every memory tier from 32 to 512 GB (@sudoingX's map, verified on October 9, 2026), but refusal figures are self-reported and the base model's license still applies. ## Sources - Refusal in Language Models Is Mediated by a Single Direction · Arditi et al., NeurIPS 2024 (arXiv), 2024-06-17 - Refusal Direction is Universal Across Safety-Aligned Languages · Wang et al. (arXiv), 2025-05-22 - XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models · Röttger et al., NAACL 2024 (arXiv), 2023-08-02 - OR-Bench: An Over-Refusal Benchmark for Large Language Models · ICML 2025 (arXiv), 2024-05-31 - Uncensored Open-weight Models: Redistribution as the Persistence Layer · 10a Labs (arXiv), 2026-09-04 - Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs · Betley et al. (arXiv), 2025-02-24 - AI guardrails stripped from Meta and Google models in minutes · The Irish Times (Financial Times), 2026-05-25 - Security incident disclosure, July 2026 · Hugging Face, 2026-07-16 - Project Glasswing · Anthropic, 2026-04-07 - Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI) · EUR-Lex, 2026-07-24 - Crime and Policing Act 2026, Part 5, Chapter 4 · legislation.gov.uk, 2026 - Code pénal, article 226-8-1 · Légifrance, 2024-05-23 - Uncensored AI past 24 GB, round 2: memory-tier map of uncensored and abliterated builds · @sudoingX on X, 2026-10-08 - Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-MTP-GGUF (model card) · Hugging Face, 2026-10-09 - gemma-4-31B-it-uncensored-heretic-GGUF (model card) · Hugging Face, 2026-10-09 - GLM-5.3-Flash-UNCENSORED-NVFP4 (model card) · Hugging Face, 2026-10-09 - Qwen3.8-Flash-Next (base model card and Qwen Community license) · Hugging Face, 2026-10-09 Editorial note: this analysis reflects the public information available on the review date. Models, prices and rules change fast; every third-party figure links to its source, and our opinions are labeled as such. Spotted an error? Write to contact@slash-digital.io. Frequently asked questions ## The questions we hear often **Is it legal to use an uncensored model?** The laws covered here target uses, outputs and tools built for abuse (intimate-image generators, child abuse material, impersonation) rather than models as such. But licenses can forbid circumventing safety filters, and from December 2, 2026 the EU ban also reaches general tools without adequate safeguards. Get legal advice for your case. **What is abliteration?** The community name for refusal-direction ablation: editing a model's weights to remove the internal direction that researchers linked to refusal in 2024. It requires the weights, and the change persists in every copy. **Are uncensored models smarter or more honest?** There is no evidence of that. Research found lower TruthfulQA accuracy after the edit and, depending on method and architecture, maths reasoning losses of up to 18.8 points on GSM8K. **Can we use one for a pentest?** Only with written authorization and in an isolated environment. Try verified-access programs and security-specialized models first: they usually cover the need with less legal and operational risk. **How do we reduce false refusals without removing safety?** Choose a better-calibrated model, write an explicit content policy, tune your guard thresholds and measure false refusals and missed harms on your own labeled prompts, in every language you serve. **Which uncensored model can I run on my GPU or Mac?** It depends on memory. With 32 GB, such as an RTX 5090, 27B to 40B builds fit at Q5 or Q6; with 48 GB, a 35B-A3B at Q8 or a 70B at Q4; with 96 to 128 GB, Qwen 3.8 Flash-Next or Qwen 3.5 122B; with 256 GB, GLM 5.3 Flash at NVFP4; and with 512 GB, GLM 5.3 Flash at FP8. The table in this article, based on @sudoingX's October 8, 2026 map, gives sizes and links verified on Hugging Face. --- # Deepfake and cloned-voice fraud: a protocol for companies URL: https://slash-digital.io/en/insights/deepfake-fraud/ Cybersecurity · September 2026 A voice or a face no longer proves who is talking to you. Documented cases, verified figures and the protocol that protects payments even when the deepfake is perfect. $25.6M Arup case, Hong Kong (2024)$893M AI-linked losses (IC3, 2025)52% of successful scams: under 30 min Start a project All insights Esteban ZuletaCo-founder, business and finance, Slash Digital Lab PublishedSeptember 26, 2026 Reading time8 min TopicCybersecurity In short Deepfake fraud is no longer a demo. In 2024 an Arup employee in Hong Kong sent about $25.6 million after a video call in which the CFO and several colleagues were deepfakes, and in 2025 the FBI's IC3 logged 22,364 complaints mentioning AI, with $893 million in losses. Detection tools help, but they will not stop a payment decided in minutes. What works is a process: no payment or bank-detail change on the strength of a voice or a face, a callback to a known number, dual approval above a threshold and a culture where pausing to verify is expected. For your company Concerned about your exposure to these risks? Slash Digital Lab assesses, prioritizes and helps fix, from Bogotá. Cybersecurity → ## What has happened: cases and numbers The reference case is Arup. In early 2024 a finance employee at the engineering firm's Hong Kong office joined a video call with what appeared to be the CFO and several colleagues; all were deepfakes. The employee made 15 transfers totaling HKD 200 million, about $25.6 million. Hong Kong police described the case in February 2024, and Arup confirmed in May that it was the victim. The FBI has since documented the pattern: criminals cloning relatives' voices to fake emergencies (December 2024) and, since April 2025, impostors posing as senior US officials with texts and AI-generated voice messages to build trust and take over accounts, often moving the victim to another messaging platform. HR is a target too: in August 2025 Anthropic reported North Korean operatives using Claude to get and keep remote IT jobs at Fortune 500 technology companies. | Source | Date | Figure | | --- | --- | --- | | FBI IC3, annual report | Apr 2026 | 1,008,597 complaints and $20.877 billion in reported losses in 2025, up 26% on 2024 | | FBI IC3, AI section | Apr 2026 | 22,364 complaints mentioned AI ($893 million), including over $30M in business email compromise and over $5M in voice-cloning "distress" scams | | FBI IC3, Colombia | Apr 2026 | 2,222 complainants from Colombia, among the top 20 foreign countries | | FinCEN (US Treasury) | Nov 2024 | Suspicious-activity reports describing deepfake media rose from 2023 into 2024 | | Kaspersky, global study | Sep 2026 | 66% of messaging-scam victims believe AI was involved; 31% reported cloned or synthetic voices; 52% of successful scams end in under 30 minutes | ## How these attacks work, conceptually Four ingredients, none of them technical on the victim's side. - Raw material. Public recordings of the person to imitate: interviews, earnings calls, conference talks, podcasts, social clips. The voice models behind this are covered in voice and audio AI in 2026. - A pretext. A confidential acquisition, an overdue supplier invoice, a regulator's request: a story that explains why the usual procedure must be skipped. - Urgency and secrecy. "It has to go out today", "don't mention this to anyone". Pressure removes the time to think. - Channel switching. The conversation moves to a channel the attacker controls: from email to WhatsApp, from a text to a voice note, from a call to a video meeting. The FBI describes exactly this move. It is the old CEO fraud with a convincing voice, and Arup adds staging: several fake participants on the same call. The targets are predictable (treasury, accounts payable, payroll, executive assistants, IT help desks), and so is the weak point: the impostor needs you to act inside their channel and on their timeline. Any control that moves the exchange to a channel you trust, or adds time, breaks the scheme. If you want to adapt those controls to your company, Slash can help with cybersecurity services. ## Why detection tools alone are not enough Deepfake detectors and watermarks are useful, but they cannot be the control that stops a payment. NIST's report on synthetic content (AI 100-4, November 2024) concludes that no single provenance or detection technique is a comprehensive solution. Watermarks exist only when the generator applies them. Since August 2, 2026, the EU AI Act has required providers of systems that generate synthetic audio and video to mark their output, but a fraudster using an open-weight model or a service outside the EU will not comply. Marks can also be attacked: UnMarker (IEEE Symposium on Security and Privacy, 2025) cut the best detection rate of semantic image watermarks to 43%. And a live call leaves no time to upload a clip and wait for a score. Rule of thumbDesign the protocol to work even if the fake is perfect. A control that depends on someone noticing an odd voice or a video glitch will not survive the next generation of models. ## The verification protocol, step by step For finance, HR and executive assistants. Adapt the thresholds to your size; keep the principles. | Situation | Control | Who | | --- | --- | --- | | Payment request by call, voice note, video or chat | Hang up and call back on a number already on file, never one given in the message | Treasury, accounts payable, assistants | | Payment above a set threshold | Two approvers; the second confirms through a channel other than the request's | Finance | | Supplier asks to change bank details | Written request, callback to the contact in your supplier file and a second approver; no payment to the new account before then | Accounts payable, procurement | | Employee asks to change payroll account | Only through the HR system with the employee's usual login, or in person | HR, payroll | | Executive request outside the usual flow | Code word or challenge question agreed in person, never sent by message | Assistants, finance | | Known person writing from a new number or account | Treat as unverified and confirm through a channel verified beforehand | Everyone | | Password or MFA reset requested by phone | Callback to the registered number and manager approval; MFA codes are never shared | IT help desk | | Remote hire or contractor | Live identity check before any system access | HR, IT | | Any doubt | Pause: anyone may stop a payment to verify it, without penalty | Everyone, backed by management | Two principles sit above the table: no payment or bank change is approved on the strength of a voice or a video alone, whoever seems to be speaking, and verification happens on a channel you chose, never the one the request came from. ## Pause-and-verify culture, training and simulations A protocol only works if people feel safe applying it to a senior voice, so the signal must come from the top: the CEO and CFO tell their teams in writing that they will never ask for a payment or bank change by voice or video alone, and that being verified is expected. - Train finance, HR, assistants and help desk staff with real cases (Arup, the FBI warnings), not generic slides. - Run a tabletop exercise: an urgent request from the "CFO" arrives on a Friday afternoon. Who calls whom, and who can stop the payment? - Run controlled phishing and voice-pretext simulations agreed in advance with management, HR and legal, with written consent from anyone whose voice or image is used. - Measure behavior, not attendance: share of staff who verified, time to escalate, payments stopped. Slash runs controlled phishing and social-engineering tests from COP 5 million (see the pentest page); for the technical side, read how AI changes pentesting. ## If the money already left: incident response - Call your bank immediately and request a recall or freeze of the transfer at the first suspicion: a false alarm is easier to undo than a transfer that has moved on. - Preserve evidence: call logs, numbers, voice notes, meeting links, chats, emails with headers and transaction details. Don't delete anything or confront the impostor. - Contain: freeze related pending payments, reset credentials that may be compromised, and warn the impersonated executive and the teams likely to be targeted next. - Report. In Colombia, to ColCERT, the national cyber emergency response team (it has incident and phishing reporting channels), and through a criminal complaint; CONPES 3995, the national digital security policy, also lists the National Police's CSIRT (CSIRT PONAL). If personal data were compromised, Ley 1581 of 2012 requires informing the SIC, through the security-incident module of the national database registry (RNBD). If a US bank or subsidiary is involved, the FBI's IC3 takes complaints. In France and elsewhere in the EU, use the national channels and involve your DPO. - Learn: identify the control that failed or was bypassed, update the protocol and share the lesson without blaming the person who was deceived. ## Legal context in Colombia and the EU Context, not legal advice: review your case with counsel. Colombia. Ley 2502 of 2025 (signed July 28, 2025) amended Article 296 of the Criminal Code on falsedad personal (impersonation): when it is committed with artificial intelligence, the fine increases by up to one third, provided the conduct does not constitute another offence. The law defines deepfakes and orders a public policy against AI misuse within two years. In a payment fraud, your lawyer will assess which offences apply. European Union. Article 50 of the AI Act has applied since August 2, 2026. Providers of systems that generate synthetic audio, images, video or text must mark output in a machine-readable, detectable way (systems already on the market have until December 2, 2026), and deployers must disclose the deepfakes they publish, with lighter duties for evidently artistic or satirical work. Breaches can cost up to 15 million euros or 3% of worldwide annual turnover. Fraudsters don't label their fakes, so these rules mostly give victims and authorities a legal basis. For your own company they cut the other way: if you use an executive's synthetic voice or avatar in marketing or training, get written consent and label it where the AI Act applies. More in our 2026 regulatory map. ## Executives: manage your public voice and video footprint The FBI advises limiting public audio and images of yourself, but an executive cannot vanish from public view. The aim is to know what exists and make sure none of it can move money. - Inventory recordings: earnings calls, interviews, podcasts, keynotes, social videos. - Protect personal numbers and accounts, which impostors use to open "private" channels. - Agree a code word in person with your assistant, your CFO and your family. - Tell teams, clients and key suppliers how you will and will not contact them, so the exception stands out. - Watch for fake profiles using your name and photo on professional networks and messaging apps. Our ruleAssume your voice can be cloned and your face animated, and make sure no payment in your company depends on either. ### Key takeaways - Deepfake fraud is documented: Arup lost about $25.6 million in 2024 after a video call with a deepfaked CFO and colleagues. - The FBI's IC3 logged 22,364 complaints mentioning AI in 2025, with $893 million in losses. - Detectors and watermarks help but cannot be the control: NIST finds that no single technique is a comprehensive solution. - The protocol: no payment or bank change by voice or video alone, callback to a known number, dual approval, code words and a pause-and-verify culture. - If money leaves, call the bank first, preserve evidence and report (ColCERT in Colombia); Ley 2502 of 2025 and AI Act Article 50 frame the legal side. ## Sources - Arup deepfake video-call scam in Hong Kong · CNN, 2024-05-16 - Senior US officials impersonated in malicious messaging campaign (PSA I-051525-PSA) · FBI IC3, 2025-05-15 - Criminals use generative artificial intelligence to facilitate financial fraud (PSA I-120324-PSA) · FBI IC3, 2024-12-03 - 2025 Internet Crime Report · FBI IC3, 2026-04-16 - Alert on fraud schemes involving deepfake media (FIN-2024-Alert004) · FinCEN, US Treasury, 2024-11-13 - The Great Messaging Heist (study press release) · Kaspersky, 2026-09-14 - Reducing Risks Posed by Synthetic Content (NIST AI 100-4) · NIST, 2024-11 - UnMarker: a universal attack on defensive image watermarking · arXiv (IEEE Symposium on Security and Privacy 2025), 2024-05-14 - AI Act, Article 50: transparency obligations · European Commission, AI Act Service Desk, 2026-08-02 - Ley 2502 de 2025 · Comisión de Regulación de Comunicaciones (compilación jurídica), 2025-07-28 - ColCERT: incident reporting · ColCERT, 2026-09 - Detecting and countering misuse of AI: August 2025 · Anthropic, 2025-08-27 Editorial note: this analysis reflects the public information available on the review date. Models, prices and rules change fast; every third-party figure links to its source, and our opinions are labeled as such. Spotted an error? Write to contact@slash-digital.io. Frequently asked questions ## The questions we hear often **Can deepfake detection software protect my company?** It helps, especially to analyze suspicious recordings after the fact, but it should not be the control that stops a payment. NIST concludes that no single detection or provenance technique is a comprehensive solution, and a live call leaves no time for analysis. What protects you is verification by process. **What should an employee do if the CEO calls asking for an urgent transfer?** Stay polite, end the call and call back on the number on file, or confirm through another verified channel. If the amount is above the threshold, the second approval applies as usual. A real CEO will accept the delay. **Is a code word enough?** It is a useful extra layer, recommended by the FBI for families, but it does not replace the callback or dual approval. Agree it in person, never send it by message and change it when the team changes. **Is impersonating someone with AI a crime in Colombia?** False personation (falsedad personal) was already an offence, and Ley 2502 of 2025 raises the fine by up to one third when it is committed with AI, provided the conduct does not constitute another offence. For a specific case, consult a lawyer. **Where do we report deepfake fraud in Colombia?** To your bank first, to try to recall the transfer. Then to ColCERT, the national cyber emergency response team, and through a criminal complaint. If personal data were compromised, to the SIC through the RNBD security-incident module. --- # Open-weight models 2026: Kimi K3, Qwen3.8, DeepSeek V4.1 URL: https://slash-digital.io/en/insights/open-weight-models-2026/ Local AI and open weights · September 2026 The best open-weight models sit a few months behind closed ones, and some run on a single workstation. The fine print lies elsewhere: licenses that change between versions, data flows that depend on where you host, and model files you must treat as software. 46 vs 58 best open vs best closed (Artificial Analysis)2.8T parameters in Kimi K3, the largest4 months average lag behind the frontier (Epoch AI) Start a project All insights Rémy SanchezCTO and co-founder, Slash Digital Lab PublishedSeptember 26, 2026 Reading time8 min TopicLocal AI and open weights In short As of September 26, 2026, the strongest open-weight models come from Chinese labs (Xiaomi's MiMo-V2.6-Pro, Z.ai's GLM-5.3, Moonshot's Kimi K3) and score 44 to 46 on the Artificial Analysis Intelligence Index, against 58 for the best closed model; Epoch AI puts the lag at about four months. The trap is the license: several flagships moved from MIT or Apache 2.0 to custom terms with revenue thresholds, attribution duties or limits on selling the model as a service. Our rule: shortlist on your own evaluation, read the license of the exact version, decide where the data flows and treat model files as software. For your company Which model handles your real cases best, in the cloud or local? At Slash AI Lab we compare models on your own prompts before you commit budget or private data. Slash AI Lab → ## Open weights are not open source A model has three parts: the weights (the trained parameters you download), the code to train and run it, and the data it learned from. "Open-weight" only promises the first, under a license that may or may not be permissive. The leading 2026 releases from DeepSeek, Qwen, Kimi or GLM publish weights and a technical report, not their training data. Few projects publish everything: Ai2's Olmo 3 and 3.1 (weights, code, checkpoints and datasets under Apache 2.0) and the Swiss AI Initiative's Apertus v1.5, which respects data owners' opt-outs and ships EU AI Act documentation. Why it matters: without the data you cannot audit what the model learned, and the European Data Protection Board warned in December 2024 that unlawfully processed training data could affect the lawfulness of a deployment. The realistic answer is to document provenance and test outputs. To test them on your own prompts, Slash runs model evaluation at Slash AI Lab, including open models run locally. ## The main families in September 2026 Sizes, context and license, from each vendor's model card. | Model (lab, release) | Total / active | Context | License | | --- | --- | --- | --- | | Kimi K3 (Moonshot AI, July 2026) | 2.8T / 104B | 1M | Custom (Kimi K3 License) | | Qwen3.8-2.4T-A95B (Alibaba, August 2026) | 2.4T / 95B | 262K, up to ~1M | Custom (Qwen3.8-Max License) | | Qwen3.8-27B (Alibaba, August 2026) | 27B, dense | 262K, up to ~1M | Apache 2.0 | | MiMo-V2.6-Pro (Xiaomi, September 2026) | 1.02T / 42B | 1M | MIT | | DeepSeek V4-Pro (April 2026, updated in August) | 1.6T / 49B | 1M | MIT | | DeepSeek V4.1-Flash (September 10, 2026) | 552B / 8B to 16B | 1M | MIT | | GLM-5.3 (Z.ai, August 2026) | 744B / 40B | 1M | Custom (GLM-5.3 License) | | GLM-5.3-Flash (Z.ai, August 2026) | 320B / 18B | 300K | MIT | | MiniMax-M3 (June 2026) | ~428B / ~23B | 1M | Custom (MiniMax Community License) | | Gemma 4 31B and 26B A4B (Google, April 2026) | 30.7B dense; 25.2B / 3.8B | 256K | Apache 2.0 | | Muse Glimmer 30B (Meta, August 2026) | ~30B, dense | 131K+ | Apache 2.0 | | gpt-oss-120b and 20b (OpenAI, August 2025) | 117B / 5.1B; 21B / 3.6B | 128K | Apache 2.0 | | Mistral Small 4 (March 2026) | 119B / ~6B | 256K | Apache 2.0 | | Mistral Large 3 and Ministral 3 (December 2025) | 675B / 41B; 3B to 14B dense | 256K (Ministral 14B) | Apache 2.0 | The flagships reach 1 to 2.8 trillion parameters and target large servers: Moonshot recommends 64 or more accelerators for K3. The efficient middle is MoE with few active parameters (DeepSeek V4.1-Flash, GLM-5.3-Flash, Mistral Small 4, gpt-oss-120b). And a dense class of 27B to 31B runs on one 32 GB GPU once quantized, as our local AI hardware guide shows. ## Strengths by use: code, reasoning, languages, vision - Coding and agents. GLM-5.3 is the best open-weight entry on the official Terminal-Bench 4.0 board (41.8%, fifth overall). On Terminal-Bench 2.1, with vendor-reported figures, DeepSeek V4.1-Flash (90.6), MiMo-V2.6-Pro (89.9) and Kimi K3 (88.3) lead; Qwen3.8-27B reports 61.7 on SWE-bench Pro. - Reasoning. Vendors report up to 93.5 on GPQA Diamond (Kimi K3), but the test is saturated (the best closed models sit at 94 to 96%) and no longer separates models. - Spanish and French. The Qwen3.5 generation covered 201 languages; Gemma 4 supports 35+ out of the box; Mistral Large 3 lists 40+ native languages, and Cohere's Command A+ (Apache 2.0) names both among 48. Regional options such as Latam-GPT, Spain's ALIA-40b or France's Luciole deserve a test for language-heavy work. - Vision and audio. Qwen3.8-27B, Gemma 4, Kimi K3, MiniMax-M3, GLM-5.3-Flash and DeepSeek V4.1-Flash accept images; MiMo-V2.6 also takes audio and video. Rule of thumbVendor benchmarks use different harnesses and effort levels. Use them to shortlist three models, never to pick the winner: that comes from your own prompts in Spanish, French and English (see how we choose a model). ## How far behind the frontier Independent leaderboards, read on September 25 and 26, 2026. | Leaderboard | Best closed | Best open-weight | Gap | | --- | --- | --- | --- | | Artificial Analysis Intelligence Index v4.3.2 | Claude Opus 5.5 (max): 58 | MiMo-V2.6-Pro: 46; GLM-5.3: 45; Kimi K3: 44 | 12 points | | Arena, text | Claude Opus 5.5 (high): 1,509 | Kimi K3 (max): 1,488; MiMo-V2.6-Pro: 1,480 | 21 to 29 points | | Terminal-Bench 4.0, official board | GPT-6 Astra (max): 58.2% | GLM-5.3 (max): 41.8% | 16.4 points | | ARC-AGI-2 | GPT-6 Astra (Max): 95.0%, $1.12 per task | DeepSeek V4 Flash 0731 (Max): 61.4%, $0.042 per task | 34 points, at 1/27 of the cost | Epoch AI measured an average lag of about four months for the best open-weight models since January 2026. The gap is small on chat preference (Arena) and wider on long agentic tasks (Terminal-Bench), exactly where frontier models earn their price. Among small models, Qwen3.8-27B scores 34 on the Artificial Analysis index, far above Gemma 4 31B (19) and Muse Glimmer (17). ## The license drift, and how to check a license In 2025 most leading open weights shipped under MIT or Apache 2.0. In 2026 several flagships moved to custom licenses, while smaller siblings often stayed permissive (Qwen3.8-27B is Apache 2.0, GLM-5.3-Flash MIT): free for most companies, with conditions that bite at scale or on certain business models. - Kimi K3 (K2 was Modified MIT): name display above 100 million monthly users or $20 million monthly revenue, and an agreement for model-as-a-service businesses above $20 million in 12 months. - Qwen3.8 flagship: the same display rule and a separate license for model-as-a-service or AI work-assistant businesses above $50 million in group revenue. Qwen3.8-Flash-Next requires it for such businesses at any size. - GLM-5.3 (GLM-5.2 was MIT): a Z.ai security review for model-as-a-service operators with group revenue above $10 billion. - MiniMax-M3: "Built with MiniMax M3" attribution, a notice or prior authorization depending on a $20 million yearly revenue threshold, and banned uses such as military; M2.7 was non-commercial. - Hy3-preview (Tencent) excluded the EU, the United Kingdom and South Korea; the final Hy3 is Apache 2.0. - Read the LICENSE file of the exact version, not the metadata: MiniMax-M2.5's card said "modified-mit" while its file added banned uses and Singapore law. - Check commercial use and user or revenue thresholds, including group revenue. - Check model-as-a-service and work-assistant clauses if you resell access or build an assistant. - Note attribution, territory, banned uses and governing law. - Store license, version and file hash in your AI bill of materials and re-check at every upgrade. Before a product launchA license that is fine for an internal tool may not be fine for a product. If you sell access to the model, directly or inside a SaaS, have counsel read it first. ## Where to run them, and where your data goes Open weights give you three hosting options, each with a different data flow. Self-hosting keeps prompts and outputs inside your perimeter: a Chinese model on your server sends nothing to China, because weights are just files. Inference providers run the same weights under their own contract and jurisdiction. The lab's API sends your data to the lab. The middle has become cheap. On September 26, 2026, gpt-oss-120b cost $0.15 / $0.60 per million input and output tokens on Groq, Together or Fireworks; DeepSeek V4.1-Flash, $0.30 / $1.20 on DeepSeek's API at peak hours and on Fireworks; Qwen3.8-27B, $0.80 / $4.00 on Groq. In Europe, Mistral's platform also lists Z.ai's GLM-5.3 and GLM-5.2: a Chinese-origin model under a French provider's contract; check the hosting region. The legal test is where personal data goes. In Colombia, Law 1581 allows transfers to countries on the SIC's adequacy list, which includes the United States but not China; for other destinations you need an exception, such as the data subject's express authorization. In the EU, transfers outside the EEA fall under Chapter V of the GDPR. ## Risks: provenance, safety tuning and model files Provenance and geopolitics. Menlo Ventures estimated in December 2025 that Chinese open models accounted for about 1% of LLM API usage among US enterprises. The US defense authorization act for fiscal 2026 requires removing DeepSeek from Defense Department systems, and Zhipu (Z.ai) has been on the US Entity List since January 2025. If you sell to governments or regulated sectors, ask about model origin even when you self-host. Safety tuning is thin. Research shows that an open model's refusals can be erased by editing its weights, and that the edit works across languages; the Financial Times reported Gemma 4's guardrails stripped within 90 minutes of release. Add a guard model (Llama Guard 4 or gpt-oss-safeguard) and test it in Spanish and French: see guardrails in production. Model files are software. Pickle checkpoints can run code when loaded, JFrog found about 100 malicious models on Hugging Face in 2024, and instructions can hide in a GGUF chat template. Prefer safetensors from official organizations, pin hashes and keep servers private: SentinelLABS and Censys counted 175,108 exposed Ollama hosts. More in AI supply chain security. The upside: when commercial models refused forensic requests during a July 2026 intrusion, Hugging Face analysed the attacker's logs with the open-weight GLM-5.2 on its own servers. ## How to decide, in Colombia and in Europe - Name the dominant constraint: quality, cost per task, latency, data residency or offline use. - Shortlist three models that fit your license needs and hardware. - Evaluate them on at least 30 real prompts in your languages, against one closed model. - Self-host the model for personal or confidential data; use a contracted provider for the rest. - Record license, version and hashes, and add a guard model. - Plan upgrades: models and licenses change every few months. Colombia and Latin America. Open weights are the cheapest route to data residency, and the SIC's 2024 guidance on AI requires a privacy impact assessment for high-risk uses. The practical path: a self-hosted open model for sensitive flows and a closed API for the hardest tasks. European Union. The AI Act has applied generally since August 2, 2026. Under the Commission's guidelines, fine-tuning an open model makes you a general-purpose AI provider only if you use more than a third of the original training compute. Our recommendation: test Qwen3.8-27B or Gemma 4 31B self-hosted for internal work and DeepSeek V4.1-Flash through a contracted provider for volume, and keep a closed frontier model where the gap shows. If a client excludes Chinese-origin models, Gemma 4, Muse Glimmer, gpt-oss and Mistral are the Apache 2.0 alternatives. ### Key takeaways - The best open-weight models score 44 to 46 on the Artificial Analysis index against 58 for the best closed model; Epoch AI puts the lag at about four months. - Open-weight means weights, not data: fully open models such as Olmo and Apertus are the exception. - Licenses are drifting: Kimi K3, the Qwen3.8 flagship, GLM-5.3 and MiniMax-M3 add thresholds, attribution or model-as-a-service limits. - Self-hosting a Chinese model sends no data to China; calling a lab's API is a different data flow. - Treat model files as software and guardrails as your job, tested in Spanish and French. ## Sources - Open weights models: Intelligence Index v4.3.2 · Artificial Analysis, 2026-09-26 - Text leaderboard · Arena, 2026-09-25 - Terminal-Bench leaderboard · Terminal-Bench, 2026-09-26 - Gap between open-weight and closed models (Epoch Capabilities Index) · Epoch AI, 2026-05-29 - Kimi K3 License · Moonshot AI (Hugging Face), 2026-07 - Qwen3.8-2.4T-A95B model card · Qwen (Hugging Face), 2026-08-12 - GLM-5.3 License · Z.ai (Hugging Face), 2026-08-25 - MiniMax Community License (MiniMax-M3) · MiniMax (Hugging Face), 2026-06 - DeepSeek-V4.1-Flash release · DeepSeek API Docs, 2026-09-10 - Gemma 4 model card · Google AI for Developers, 2026-07-30 - Models overview · Mistral AI documentation, 2026-09-26 - LLM03:2025 Supply Chain · OWASP Top 10 for LLM Applications, 2024-11 Editorial note: this analysis reflects the public information available on the review date. Models, prices and rules change fast; every third-party figure links to its source, and our opinions are labeled as such. Spotted an error? Write to contact@slash-digital.io. Frequently asked questions ## The questions we hear often **What is the best open-weight model in September 2026?** On the Artificial Analysis Intelligence Index v4.3.2, Xiaomi's MiMo-V2.6-Pro (46), Z.ai's GLM-5.3 (45) and Moonshot's Kimi K3 (44) lead. For a single GPU, Qwen3.8-27B (34, Apache 2.0) is the strongest small model on the same index. **Is open-weight the same as open source?** No. Open-weight models publish their trained parameters under a license; the training data, and often the training code, stay private. Fully open models such as Ai2's Olmo or Apertus also publish data and code. **Can I use Qwen, DeepSeek or Kimi commercially?** DeepSeek V4 and V4.1 (MIT) and Qwen3.8-27B (Apache 2.0) allow it with few conditions. Kimi K3 and the Qwen3.8 flagship allow it but add display duties and limits for model-as-a-service businesses above revenue thresholds. Read each license. **Is it safe to run a Chinese model on my servers?** The weights do not send data anywhere by themselves. The risks are the files (use safetensors from official sources), the exposure of the inference server, weak built-in guardrails and your clients' policies on model origin. **Which open models handle Spanish and French well?** Qwen3.5, Gemma 4, Mistral's models and Cohere's Command A+ explicitly support both. Vendor claims are not enough: test your own prompts, including after quantization. --- # AI cost per token in 2026: real cost, caching and routing URL: https://slash-digital.io/en/insights/ai-costs/ Strategy and regulation · October 2026 Price per million tokens is only the start. The real bill depends on caching, reasoning effort, retries and architecture decisions made before the first prompt is written. Here is how we work it out. $10 / $50 new premium tier per million tokens7.7x cost per task, max vs low effort on Opus 5.550% batch discount at Anthropic, OpenAI and Google Start a project All insights Esteban ZuletaCo-founder, business and finance, Slash Digital Lab PublishedSeptember 26, 2026 ReviewedOctober 8, 2026 Reading time8 min TopicStrategy and regulation In short In 2026 AI prices moved two ways: a premium tier at $10 / $50 per million tokens appeared (Claude Fable 5.1, GPT-6 Astra) while mid-tier prices fell (GPT-6 Sol at $2 / $10, Claude Opus 5.5 20% below Opus 5). The bill, though, depends less on list prices than on caching, reasoning tokens, batch discounts and surcharges. Our rule: measure cost per solved task, not price per token, apply the levers in order and run the local-versus-API numbers with your own volumes before buying hardware. For your company Choosing a model or an AI architecture for your company? Slash Digital Lab helps you decide with your own data, costs and constraints. AI consulting → ## How AI is billed in 2026 Eight meters run on every invoice; most teams watch two. | Meter | How it works | Examples | | --- | --- | --- | | Input and output | Output costs five to six times input | Claude Opus 5.5 $4 / $20; GPT-6 Sol $2 / $10; Gemini 3.5 Flash $1.50 / $9 | | Reasoning tokens | Billed as output, even when you only see a summary | OpenAI and Google document it; effort sets the volume | | Cached input | Reused prefixes at a fraction of input; writing the cache costs extra | Reads at 10% of input at Anthropic, on OpenAI's GPT-5 family, GPT-6 Sol and GPT-6 Astra and on Gemini 3.5 Flash; 5% on Claude Opus 5.5 and Sonnet 5.5 (since October 7), GPT-6.1 Sol and Gemini 4 Argon; Anthropic writes at 1.25x or 2x | | Batch | Asynchronous jobs at half price | 50% off at Anthropic, OpenAI and Google | | Speed | Premium for faster responses | OpenAI Fast (called Priority until July 30, 2026) at 2x and Ultrafast at 6x, the latter only on GPT-6 Astra ($60 / $300) and GPT-6.1 Sol; Claude Opus 5.5 fast mode (research preview) $8 / $40 | | Long context | Surcharge above a threshold, by vendor | OpenAI (GPT-5.5 and later): above 272K, the whole request at 2x input and cache and 1.5x output; Gemini 3.1 Pro $4 / $18 above 200K; Claude 4.6 and later none, except Haiku 5.5 (5x above 100K) | | Data residency | Premium to keep inference in a region | OpenAI +10% on models since March 5, 2026; GPT-6 Astra on Microsoft Foundry: US Data Zone +10%, no EU Data Zone announced at launch; Anthropic US-only 1.1x; Bedrock and Vertex regional endpoints +10% | | Media and tools | Per image, per second or per call | Nano Banana 2 $0.067 per 1K image; Veo 3.1 $0.40 per second; Claude web search $10 per 1,000 | ## Prices are splitting, not just falling The trend depends on the tier you buy. 2026 created a premium tier at $10 / $50: Claude Fable 5 (June 9), Fable 5.1 (September 1) and GPT-6 Astra (September 3). The middle got cheaper: GPT-6 Sol ($2 / $10) and Luna ($0.10 / $0.50) launched on September 22 at half or less of their GPT-5.6 predecessors' promotional prices. The low end too: Claude Haiku 5.5 (October 7) costs $0.10 / $0.50 for prompts up to 100K tokens, 90% less than Haiku 4.5 ($1 / $5), and five times more above that threshold. Per unit of capability, the fall is steep. Epoch AI found the price of reaching a fixed performance level dropped 9x to 900x a year depending on the threshold, with the fastest drops the least certain to last. On ARC Prize data, Claude Opus 4.6 scored 64.6% on ARC-AGI-2 for $2.25 per task in February 2026; DeepSeek V4 Flash scored 61.4% for $0.042 in July. Token counts move too: Anthropic's tokenizer, used since Claude Opus 4.7, produces about 30% more tokens for the same text, so a lower price per token can still mean a higher bill. Haiku 5.5 is the latest case: it counts about 30% more tokens than Haiku 4.5 for the same text, so its 90% cut per token does not translate into the same saving per task. ## Cost per solved task, not price per token The unit that matters is the finished job. A cheaper token that needs two retries is not cheaper. Our working formula: cost per solved task = spend on a route ÷ tasks that pass your evaluation. It captures retries, reasoning and tool calls that a price sheet ignores. Effort shows how far the two diverge. On ARC Prize's ARC-AGI-2 board on September 26, Claude Opus 5.5 cost $0.241 per task at low effort (70.1%), $0.408 at high (93.3%) and $1.85 at max (91.7%): max cost 7.7 times low and 4.5 times high, for a lower score than high. The same holds across models. On the Artificial Analysis index at max effort, GPT-6 Astra costs $3.26 per task against $5.98 for Opus 5.5, even though its tokens cost 2.5 times more: to run the index, Opus 5.5 generated 260 million output tokens and Astra 60 million. Treat any cost-per-task figure you haven't measured with suspicion: verify it with your own evals and tune effort as explained in when paying for thinking pays off. ## The levers, in the order we apply them Quality-neutral levers first; trade-offs last, and only with a before-and-after eval. | Lever | Documented effect | Watch out for | | --- | --- | --- | | 1. Caching | Reads at 10% of input (5% on Opus 5.5, Sonnet 5.5 and GPT-6.1 Sol); in a 2026 case study, a 99.3% cache hit rate cut API spend 88.6% | Needs a stable prefix; on Anthropic, changing effort invalidates the cache | | 2. Prompt and tool hygiene | Anthropic cut an agent workflow from 150,000 to 2,000 tokens by calling MCP tools from code | Tool definitions are billed: Claude's computer-use toolset adds about 4,500 tokens per request | | 3. Output length | OpenAI's verbosity setting gave 560, 849 or 1,288 tokens on one task | On Opus 5, effort doesn't reliably shorten answers: ask for length in the prompt | | 4. Batch | 50% off at the three big vendors; stacks with caching on Anthropic | Asynchronous; not combinable with Anthropic's fast mode | | 5. Effort | Opus 5.5 on ARC-AGI-2 (September 26): $0.241 per task at low, $1.85 at max | Less effort can cost quality; set it per route | | 6. Routing and cascades | RouteLLM cut cost over 85% on MT Bench keeping 95% of GPT-4's quality; Cursor says its router beats Opus 4.8 at 41% lower cost | Routers need their own evals and a fallback | | 7. Smaller models | GPT-6 Luna costs a twentieth of GPT-6 Sol, as Claude Haiku 5.5 does against Sonnet 5.5 for prompts up to 100K tokens; gpt-oss-120b is $0.15 / $0.60 on Groq and Together | Gaps on hard tasks and in Spanish or French | | 8. Distillation | DeepSeek says its 8B distilled model matches Qwen3-235B-thinking on AIME 2024 | Needs data, training and evals; Anthropic blocks extraction of Opus 5.5's reasoning | ## Costs that don't show on the token bill Budget them from day one. - Evals and observability. Every model or prompt change needs a re-run; our evals guide puts a full run at $16 to $32 in its example. - Engineering time. In the same 2026 case study, the local open-weight setup had a fix-commit ratio of 74.9% against 45.9% with the API model. - Retries, refusals and guardrails. Failed attempts are billed, some Fable 5 refusals have been billed since September 24, 2026, and each guardrail is another model call. - Runtime fees. Claude Managed Agents bills $0.08 per session-hour and web search $10 per 1,000 searches. - Lock-in. A US export-control directive forced Anthropic to pull Fable 5 for all users from June 12 to July 1, 2026. - Scheduled price changes. Gemini 3.6 to 3.8 Flash double to $1.50 / $7.50 on January 1, 2027; GPT-5.6 Sol's $4 / $20 is promotional through at least November 21, 2026; and Gemini 4 Argon will move from its $2 / $10 introductory price to $4 / $20 on a date Google hasn't announced. ## Local or API: a break-even method It depends on what you compare against and how busy the machine is. Monthly local cost = hardware ÷ months of life + power + operations time. Break-even = monthly local cost ÷ API cost per request. Then check that the machine can serve that volume and that the local model passes the same eval. An illustrative example, not a quote: a workstation with an NVIDIA RTX 5090 (launch price $1,999, 575 W) running gpt-oss-20b, which LMSYS measured on that card at 8,519 tokens per second reading and 205 writing. Our assumptions: $1,000 for the rest of the machine, 36 months of life, 0.7 kW, 8 hours a day for 20 days a month at $0.15 per kWh (as in Pan et al.), 8 hours of operations a month at $50 an hour, and requests of 2,000 input and 500 output tokens. | API compared | Cost per request | Requests a month to match $500 | Fits local capacity? | | --- | --- | --- | --- | | Claude Opus 5.5 ($4 / $20) | $0.018 | About 27,800 | Yes | | GPT-6 Sol ($2 / $10) | $0.009 | About 55,600 | Yes | | GPT-6 Luna ($0.10 / $0.50) | $0.00045 | About 1.1 million | No | | gpt-oss-20b on Groq ($0.075 / $0.30) | $0.0003 | About 1.7 million | No | Against a frontier API, the workstation pays for itself at tens of thousands of requests a month, if a 20B model passes your evals. Against the same open model served by an API, it doesn't at this utilization. Operations is the biggest line: without it, break-even against GPT-6 Sol falls to about 11,100 requests. Studies agree on the shape. Pan et al. (2025) see on-premise paying off mainly above 50 million tokens a month or under residency mandates, and the 2026 case study found cached API calls at about $0.57 per million tokens, below the $2.83 of a shared on-premise GPU slice. Epoch AI adds that AI-chip performance per dollar rises about 49% a year: owned hardware ages fast. Before you buyRedo this calculation with your volumes, eval pass rates and labor costs. Compliance can justify local inference on its own; cost alone rarely does below high, steady volume. Options in our local AI hardware guide. ## FinOps for AI: budgets, alerts and reviews Cost control is a routine, not a project. - Tag every call with route, feature, customer, model, effort and cache hit. - Dashboards per route: cost per solved task, cache hit rate, output tokens per task, retries. - Alerts and quotas: daily spend against a baseline, hard caps per key and per team. Agents burn budgets fast: OpenAI says its median researcher's daily token use, at API prices, exceeds $600. - Effort discipline: set it per route, not per request; on Anthropic a change invalidates the cache. - A monthly review of model mix, eval pass rates, price changes and retirement dates, with one owner who can say no. ## Procurement notes for Colombia and Europe Currency, residency and exit terms matter as much as the rate card. Currency. Rate cards are in US dollars, so a budget in pesos or euros adds exchange-rate risk to usage risk. Budget in USD with a buffer, review monthly and have your tax advisor confirm how foreign digital services are taxed in your case; we don't model taxes here. Residency. Keeping inference in a region costs more: +10% for OpenAI regional processing and for Bedrock or Vertex regional endpoints, 1.1x for Anthropic's US-only inference. Buying Claude through AWS, Google Cloud or Microsoft lets you use existing cloud commitments. Retirements and exit. Anthropic publishes minimum availability dates (Opus 5.5 not before September 22, 2027; Haiku 4.5 only until at least October 15, 2026). Negotiate notice periods for retirements and price changes, migration support and data-retention terms, and keep a tested fallback from a second vendor. We help with this in AI consulting. ### Key takeaways - Price per token is only the base: caching, reasoning billed as output, batch and surcharges for long context, speed or residency change the bill. - 2026 split the market: a $10 / $50 premium tier appeared while mid-tier prices fell, and price per capability keeps dropping. - Measure cost per solved task: on ARC-AGI-2 (September 26), Opus 5.5 at max effort cost 7.7 times low and scored below high. - Apply caching and batch first; effort, routing, smaller models and distillation trade quality for cost and need a before-and-after eval. - Local beats the API only at high, steady volume or for compliance; operations time is usually the biggest line. ## Sources - Pricing · Anthropic documentation, 2026-09 - Effort · Anthropic documentation, 2026-09 - Models overview · Anthropic documentation, 2026-10-08 - Claude Haiku 5.5: overview and pricing · Anthropic documentation, 2026-10-07 - Introducing Claude Haiku 5.5 · Anthropic, 2026-10-07 - Introducing Claude Sonnet 5.5 · Anthropic, 2026-09-28 - API pricing · OpenAI, 2026-10-08 - Introducing GPT-6 Sol and GPT-6 Luna · OpenAI, 2026-09-22 - GPT-6 Astra: frontier intelligence for work, now generally available in Microsoft Foundry · Microsoft Azure, 2026-09-03 - Gemini API pricing · Google, 2026-09 - Gemini 4 Argon · Google, 2026-09-30 - ARC-AGI leaderboard · ARC Prize, 2026-09-26 - AI model leaderboards (Intelligence Index) · Artificial Analysis, 2026-10-08 - Claude Opus 5.5: intelligence, performance and price · Artificial Analysis, 2026-10-08 - GPT-6 Astra: intelligence, performance and price · Artificial Analysis, 2026-10-08 - LLM inference prices have fallen rapidly but unequally across tasks · Epoch AI, 2025-03-12 - RouteLLM: An Open-Source Framework for Cost-Effective LLM Routing · LMSYS, 2024-07-01 - Code execution with MCP · Anthropic, 2025-11-04 - A Cost-Benefit Analysis of On-Premise Large Language Model Deployment: Breaking Even with Commercial LLM Services · Pan et al., arXiv, 2025-08-30 - GeForce RTX 5090 · NVIDIA, 2025-01-06 Editorial note: this analysis reflects the public information available on the review date. Models, prices and rules change fast; every third-party figure links to its source, and our opinions are labeled as such. Spotted an error? Write to contact@slash-digital.io. Frequently asked questions ## The questions we hear often **Why is output more expensive than input?** Vendors charge more for generating than for reading: on current models output costs five to six times input (Claude Opus 5.5 $4 / $20, GPT-6 Sol $2 / $10, Gemini 3.5 Flash $1.50 / $9). OpenAI and Google bill reasoning tokens as output, so effort settings weigh on the bill. **How much does prompt caching save?** Cache reads cost 10% of the input price at Anthropic, on OpenAI's GPT-5 family, GPT-6 Sol and GPT-6 Astra and on Gemini 3.5 Flash, and 5% on Claude Opus 5.5 and Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon. Writes cost extra (1.25x or 2x at Anthropic), so caching pays when the same prefix is reused while the cache is alive. **Is the Batch API worth it?** For anything that doesn't need an immediate answer, yes: Anthropic, OpenAI and Google charge 50% less for batch jobs, and at Anthropic the discount stacks with caching. **When does local inference beat the API?** At high, steady volume against expensive APIs, or when compliance requires it. In our illustrative example, one RTX 5090 workstation matches GPT-6 Sol at about 55,600 requests a month but never the same open model served by an API. **How should a company budget for AI billed in dollars?** Budget in USD with a buffer for the exchange rate, set quotas and alerts per route, review monthly with announced price changes in view, and have your tax advisor confirm how foreign digital services are treated. --- # Fine-tuning vs RAG: the Slash AI Lab criterion URL: https://slash-digital.io/en/insights/fine-tuning-vs-rag/ Agents and engineering · September 2026 By default, RAG: when knowledge must stay current and citations matter. Fine-tuning when style, language or a narrow skill doesn't fit in context. Hybrid when both are true. Default RAGWhen it fits LoRA adaptersRule constraints before recipes Start a project All insights Jean MartialCEO and co-founder, Slash Digital Lab PublishedSeptember 23, 2026 ReviewedSeptember 26, 2026 Reading time3 min TopicAgents and engineering In short Fine-tuning and RAG don't compete: they solve different problems. This is the decision order we follow before spending a single peso on training. For your company RAG, fine-tuning or both? Slash Digital Lab adapts models to your domain, language and style, and measures whether they really improve. Fine-tuning & RAG → ## RAG first, almost always Changing knowledge and required citations: RAG. If your knowledge changes (catalogues, policies, documentation) or your answers must carry citations, RAG is the default: index, re-rank and cite, without touching the model. RAG fails when the problem isn't knowledge but form: tone, output format, or a narrow task where the generic model falls short of your quality bar. That's where fine-tuning comes in. If you want to decide which of the two fits your case, Slash supports companies with LLM fine-tuning and RAG. ## Fine-tune for form, not facts LoRA adapters for tone and format; facts stay in RAG. Fine-tuning wins when you need the model to sound, format or execute like your organization: editorial style, fixed output structures, a narrow domain with stable rules. At Slash we use LoRA adapters for that layer: we don't retrain the world. Cost stays bounded, rollback is possible, and knowledge keeps living in RAG, where it can be updated. ## Hybrid, when both are true Adapters for form, RAG for data. When the style must be yours and the knowledge must stay current, hybrid wins: adapters for form, RAG for data. It's the most common configuration in Slash deployments and the cheapest to maintain. ## Four questions before paying for training Answer them in writing; the route follows. Before paying for training, answer in writing: is the problem knowledge or form? Does your data change daily or stay stable? Must citations be verifiable? Must the knowledge stay in-house? The answers define the route; everything else is consultant margin. ### Key takeaways - RAG by default: changing knowledge and required citations. - Fine-tune only for form: style, format and narrow tasks. - Hybrid when both are true. - Nothing is decided without a measured baseline. - Local inference when compliance demands it. ## Sources - Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks · Lewis et al., arXiv, 2020-05-22 - LoRA: Low-Rank Adaptation of Large Language Models · Hu et al., arXiv, 2021-06-17 - QLoRA: Efficient Finetuning of Quantized LLMs · Dettmers et al., arXiv, 2023-05-23 - Introducing Contextual Retrieval · Anthropic, 2024-09-19 Editorial note: this analysis reflects the public information available on the review date. Models, prices and rules change fast; every third-party figure links to its source, and our opinions are labeled as such. Spotted an error? Write to contact@slash-digital.io. Frequently asked questions ## The questions we hear often **When is fine-tuning NOT worth it?** When the problem is missing or changing knowledge: RAG handles that with less cost and less risk. **What's a LoRA adapter?** A small layer trained on top of the base model for form and narrow tasks: cheap, reversible and compatible with local inference. **Can RAG and fine-tuning be combined?** Yes, that's the usual production pattern: adapters for form, RAG for cited knowledge. **How much does an experiment cost?** It's scoped in sprints and the diagnosis sets the size. Write to us with your case. **How is the outcome measured?** Against a baseline of the best generic model, with comparable quality, cost and latency metrics. --- # What to test in a web and API pentest: 2026 checklist URL: https://slash-digital.io/en/insights/pentest-colombia/ Cybersecurity · October 2026 Colombia detects roughly 1.3 million attack attempts a day. A scanner is the floor; manual validation is what turns a PDF into decisions. 1.3M attempts a day (CRC and FortiGuard)471.5M attempts a yearCOP $0 with no high findings Start a project All insights Esteban ZuletaCo-founder, business and finance, Slash Digital Lab PublishedSeptember 23, 2026 ReviewedOctober 14, 2026 Reading time3 min TopicCybersecurity In short The question isn't whether your platform is secure, but what you're missing. This piece defines what to really test (web, APIs, manual validation and retest) and how the Slash launch offer aligns price with outcome. For your company Would your application or API withstand these attacks? Slash Digital Lab runs manual pentests of web applications and APIs in Colombia. Pentest Colombia → ## 1.3 million attempts a day Colombia, 2025: 471.5 million attack attempts detected. Colombia detects roughly 1.3 million attack attempts a day, 471.5 million over the year (source: CRC and FortiGuard Labs). An attempt isn't a successful attack, but every exploitable flaw is a door someone may find before you do. The regional picture confirms the pressure: according to Check Point Research, in the first half of 2025 organizations in Latin America faced an average of 2,716 attacks per week, 39% above the global average, and Colombia was among the five most targeted countries in the region. Kaspersky, for its part, reports Latin America as the region most affected by ransomware in 2025. ## What to actually test Web apps and APIs, in that order. An honest test covers your real surface: the web apps in production and the APIs behind them. APIs are where the most expensive doors open today: authentication, authorization, input validation and data exposure. That is the scope of our web and API penetration testing. Every finding is validated by hand and must be reproducible; severity combines CVSS with business context. What can't be reproduced doesn't get reported. - Web apps: authentication, sessions and business logic. - APIs: authorization, IDOR, excessive data exposure and misconfiguration. - Configuration: headers, TLS, secrets and exposed surface. - Manual validation of every relevant finding, not just the scanner. ## A price aligned with the outcome COP $0 if no high or critical findings are confirmed. Our launch offer aligns price with outcome: COP $0 if no high or critical vulnerabilities are confirmed within scope; COP $5M for one high finding; capped at COP $9M however many we find. The assessment always runs in full, with an executive and a technical report, and one retest included in paid engagements. ## And the human factor Controlled phishing and social engineering, from COP $5M. Technology exposes; people open. We run controlled phishing and social-engineering tests from COP $5 million, with a report and training for your team. With voice cloning and deepfakes this layer matters more every year: see our protocol against deepfake fraud. ### Key takeaways - Test the production web and APIs, not the demo. - Validate every finding by hand; the scanner is the floor. - Severity combines CVSS and business context. - The retest verifies the fixes and is included in paid assessments. - No high findings means the assessment costs COP $0. ## Sources - Latin America 2025 mid-year cyber snapshot · Check Point Research, 2025-07-10 - América Latina, la región más afectada por el ransomware · Kaspersky, 2026-05-12 - Internet Crime Report 2025 · FBI IC3, 2026-04 - OWASP API Security Top 10 (2023) · OWASP, 2023 - Web Security Testing Guide · OWASP - CVSS v4.0 Specification · FIRST, 2023-11-01 Editorial note: this analysis reflects the public information available on the review date. Models, prices and rules change fast; every third-party figure links to its source, and our opinions are labeled as such. Spotted an error? Write to contact@slash-digital.io. Frequently asked questions ## The questions we hear often **Is COP $0 real when nothing is found?** Yes: with no confirmed high or critical findings within scope, the assessment is free and you keep the full report. **Where does the statistic come from?** From CRC and FortiGuard Labs (prospective cybersecurity analysis, 2026). An attempt is not a successful attack. **Do you test mobile apps?** The offer covers web apps and APIs; other scopes are quoted separately. **Who fixes the findings?** Whoever you choose: your team, another vendor or Slash. The retest still applies. **Do you sign an NDA?** Yes: scope and confidentiality are signed before testing starts. --- # Environmental pledge | Slash Digital Lab URL: https://slash-digital.io/en/pledge/ Environmental pledge At Slash Digital Lab, we recognize that environmental sustainability is a global issue that affects everyone, and we are committed to doing our part to make a positive difference. Here is what that means in the way we design, build and run digital products. Signed October 2026Reviewed every yearAligned with UN SDG 12 · 13 Start a project Read the six commitments In short The pledge. At Slash Digital Lab, we recognize that environmental sustainability is a global issue that affects everyone, and we are committed to doing our part to make a positive difference. What it changes. Our footprint is mostly digital: the code we ship, the servers and AI models it runs on, the devices we use and the trips we take. So the pledge is six concrete commitments on those, with a public review every year and no offset-based claims of carbon neutrality. The commitments ## Six commitments we hold ourselves to Each one targets a part of the footprint we actually control. 01 ### Lighter software by default Every product gets a performance budget: page weight, requests and load time are set before design and checked before launch. Fewer dependencies, optimised images and video, and no code that runs for nothing. 02 ### AI sized to the task We use the smallest model that meets the quality bar, measured with our own evals. Caching, batching and local or on-device models where they do the job; a frontier model only when the task needs one. 03 ### Efficient infrastructure Servers sized to real traffic, test environments switched off when idle, and data kept only as long as it is needed. When a client lets us choose, we prefer regions and providers committed to renewable energy. 04 ### Hardware that lasts We repair before we replace and keep laptops and phones in service for as long as they do the job. Screens and devices for DOOH and events are rented or reused across activations, and retired electronics go to certified e-waste recyclers. 05 ### Less paper, fewer miles Paperless workflows by default, inside the company and in what we build, as with JCDecaux Paperless. Our team works remote-first across Bogotá, Paris, Lyon and Madrid: we travel when being there changes the outcome, and take the train over the plane where it is a real option. 06 ### Honest advice to clients When a lighter option does the job (a web app instead of a native app, a smaller model, fewer screens, a shorter campaign), we say so, even when it makes the project smaller. Accountability ## How we hold ourselves to it A pledge only counts if it changes decisions. This is the loop we run. - ### Measure Page weight and Core Web Vitals on every release, cloud usage per project, an inventory of our hardware, and our business travel. - ### Decide Environmental impact is a criterion in every architecture review, next to security, cost and performance. - ### Act Each commitment has an owner in the founding team and lives in our project checklists, not in a slide. - ### Report Every year we update this page with what we measured, what changed and what didn't. First review: October 2027. Signed and dated ## The pledge at a glance - **Signed by:** Slash Digital Lab SAS (tax ID 901670192-8), represented by Jean Martial, CEO - **Date:** October 8, 2026 - **Scope:** Our Bogotá head office, our team in Paris, Lyon and Madrid, and every product we design, build and run - **Frameworks:** Green Software Foundation principles, the W3C Web Sustainability Guidelines and UN Sustainable Development Goals 12 (responsible consumption and production) and 13 (climate action) - **What we don't claim:** Carbon neutrality. We don't buy offsets to say so: we reduce what we control and report it. - **Next review:** October 2027, published on this page - **Questions or suggestions:** contact@slash-digital.io At Slash Digital Lab, we recognize that environmental sustainability is a global issue that affects everyone, and we are committed to doing our part to make a positive difference. Jean Martial, CEO · Slash Digital Lab SAS · October 8, 2026 Frequently asked questions ## What people ask about this pledge **Is Slash Digital Lab carbon neutral?** No, and we don't claim to be. We don't buy offsets to make that claim. We reduce the emissions we control (software, hosting, AI compute, hardware and travel) and report progress here every year. **Does the pledge change how you build my project?** It adds one question to every technical decision: is there a lighter way to get the same result? Most of the time the answer also makes the product faster and cheaper to run. You keep the final say. **Can you help us reduce the footprint of an existing product?** Yes. A performance and infrastructure review usually finds the biggest wins: heavy pages, oversized servers, idle environments and AI calls that a smaller model or a cache could handle. Write to us and we'll tell you honestly whether it's worth it. **How will we know you kept it?** Every October we update this page with what we measured, what changed and what didn't. If you see something that contradicts this pledge, tell us at contact@slash-digital.io. --- # Privacy notice | Slash Digital Lab URL: https://slash-digital.io/en/privacy/ Privacy notice Which personal data we process, why, with whom, and how to exercise your rights under Colombia's Law 1581 of 2012 and, if you are in the European Union, the GDPR. Controller Slash Digital Lab SASQueries 10 business daysClaims 15 business days Start a project Cookie policy The notice, in parts ## Everything we do with your data - **Data controller:** Slash Digital Lab SAS, tax ID (NIT) 901670192-8. Carrera 8A 97-07, Bogotá, Colombia. Email: contact@slash-digital.io · Phone: +57 312 697 5726. - **Data we process:** Contact form: name, email, company, message and, if you provide them, project type, budget, timeline and how you heard about us. WhatsApp: the messages you send to Esteban Zuleta (+57 321 4119660). Browsing: only if you accept analytics cookies, Google Analytics 4 records page views, events, device type and approximate location under a pseudonymous identifier. We don't ask for sensitive data or data about minors. - **Purposes:** Answering your request and preparing proposals; managing the pre-contractual and contractual relationship; measuring site use in aggregate (only with your consent); keeping the site secure and meeting legal obligations. We don't sell personal data or run third-party advertising. - **Authorisation and legal bases:** In Colombia we process your data with the authorisation you give when you send the form or write to us after being informed of this policy (Law 1581 of 2012 and Decree 1377 of 2013); we keep proof of that authorisation. In the European Union (GDPR): your consent for analytics; the pre-contractual steps you request by writing to us (Art. 6(1)(b)); and our legitimate interest in site security (Art. 6(1)(f)). - **Processors and international transfers:** Providers that process data on our behalf: Hostinger (hosting of the site and the form), Google (Google Workspace for email and, if you accept it, Google Analytics 4) and Meta (WhatsApp). Google and Meta may process data in the United States; those transfers rely on the safeguards the applicable law requires (in the EU, the EU-U.S. Data Privacy Framework or standard contractual clauses). - **Retention:** Messages and contact details: up to 24 months after the last exchange, unless they become a contractual relationship, in which case the applicable legal periods apply. Analytics: Google Analytics cookies last up to 2 years in your browser, and Google keeps the data for the period set in the account (14 months at most). Your cookie choice: 13 months. - **Your rights (Law 1581, Art. 8):** To know, update and correct your data; to request proof of your authorisation; to be told how we use your data; to file complaints with the Superintendence of Industry and Commerce (SIC); to revoke your authorisation or request deletion; and to access your data free of charge. If you are in the EU you can also request portability and restriction, object, withdraw consent at any time and complain to your supervisory authority (for example the CNIL in France or the AEPD in Spain). - **How to exercise them:** Write to contact@slash-digital.io with your name, your request and a way to reach you. Responsible area: the management of Slash Digital Lab SAS. Queries: answered within 10 business days (extendable by 5, with the reason). Claims: within 15 business days (extendable by 8). Before going to the SIC you must first have filed the query or claim with us. - **Security:** Encrypted connection (HTTPS), restricted access to messages and providers bound by confidentiality. If a security incident affects your data, we will notify the competent authority and the people affected as the law requires. - **Effective date and changes:** This policy applies from September 26, 2026. Any material change will be published here with its date; previous versions are available on request. Frequently asked questions ## The questions we hear often **Do you sell my data?** No. Never. **Do you use advertising cookies?** No. We only use Google Analytics 4, and only if you accept it; details in the cookie policy. **How do I erase my data?** Write to contact@slash-digital.io: we delete it except what the law requires us to keep, and confirm it to you. **What if I don't want my visit measured?** Reject analytics in the cookie notice or via “Cookie preferences” at the bottom of every page. The site works the same. --- # Cookie policy | Slash Digital Lab URL: https://slash-digital.io/en/cookies/ Cookie policy We only store what's needed and, if you accept, use Google Analytics 4 to learn which content is useful. No advertising. Analytics only with your consentAdvertising noneYour choice change it anytime Start a project Privacy notice The detail ## Everything we store in your browser | Name | Purpose | Duration | Type | | --- | --- | --- | --- | | slash-consent | Stores your analytics choice (accept or reject). | 13 months, then we ask again | Local storage, first-party | | slash-theme | Remembers whether you prefer the light or dark theme. | Until you clear it | Local storage, first-party | | _ga | Google Analytics 4: tells visits apart with a pseudonymous identifier. Only if you accept analytics. | 2 years | Cookie from Google LLC (US) | | _ga_VZETE64Y2K | Google Analytics 4: keeps the session state. Only if you accept analytics. | 2 years | Cookie from Google LLC (US) | Without your consent, Google Analytics isn't loaded and no data is sent to Google. You can change your choice anytime with “Cookie preferences” at the bottom of every page; if you withdraw consent, we delete the _ga cookies. We honour your browser's Global Privacy Control signal: if it's on, we treat your visit as a refusal. The contact form uses no cookies: it sends your message straight to our server. How we handle that data is explained in the Privacy notice. Change my cookie preferences Frequently asked questions ## The questions we hear often **Do you use advertising cookies?** No, none. **What if I reject analytics?** Nothing changes on the site: we just don't measure your visit. **How do I change my choice?** With “Cookie preferences” at the bottom of every page, or the button above. **Who receives the analytics data?** Only Google, as the Google Analytics 4 provider, and only if you accept. More in the privacy notice.