AI engineerGeneva / Lausanne, Switzerland
GenAI engineer at Orange Business. I build RAG systems, multi-agent workflows and LLM infrastructure that run in production, and I ship my own products on the side.
Open to AI engineer, forward-deployed and solutions roles
Ship log
73,000+
fan interactions in 2 weeks, 71.5% completion rate
JIP - Basketball (FIBA World Cup 2023)202373,000+ fan interactions in 2 weeks, 71.5% completion rate, 5 languages. Led the official GPT-4 chatbot for the 2023 FIBA Basketball World Cup end to end, from PoC to a 4-month launch on Facebook Messenger and web.
GPT-4 · Real-time API · Facebook Messenger · Web Widget · Multilingual (5 languages) · Project Management
Demoed to the CEO
of Orange Business in July 2025, in production for product and sales teams
KARL2024-2026Automated cloud competitive-intelligence robot and RAG chatbot in production at Orange Business, demoed live to the CEO in July 2025. First orchestrated with N8N, paused in early 2026 for compliance, then rebuilt on LangChain + ChromaDB with local models (Llama 3.3 70B, DeepSeek R1, QwQ 32B) via vLLM on H100 NVL and L40S GPUs.
LangChain · ChromaDB · vLLM · H100 NVL · Llama 3.3 70B · DeepSeek R1
Paying users
live at jobswiper.ai, built solo
JobSwiper2026Live at jobswiper.ai with paying customers, built solo. AI-powered job-search SaaS: a match-scoring engine (vector embeddings + LLM reranking), a Canva-style CV editor, a real-time interview simulator, and a Chrome extension for 18+ job boards and ATS systems.
Next.js 16 · React 19 · TypeScript · Tailwind CSS · shadcn/ui · Supabase
20 PRs
merged into apache/airflow, through an agent pipeline
20 pull requests merged into apache/airflow (38,000+ stars), out of 42 submitted, through an autonomous agent pipeline I built to run the contribution cycle. Sharp bugfixes: scheduler, task-sdk, providers (Google, Oracle, Alibaba, SMTP, FAB), database migrations, including 2 security fixes.
Python · Apache Airflow · Autonomous agents · Claude Code · Open Source · pytest
1.27B
parameters, trained from scratch to understand
Koda, an LLM trained from scratch2026A 1.27 billion parameter language model trained from scratch, just to understand how they work inside. LLaMA-style decoder-only (24 layers, GQA, SwiGLU, RoPE), trained in JAX/Flax NNX on 2 L40S GPUs. Checkpoints published on Hugging Face, with HF, GGUF and MLX exports.
JAX · Flax NNX · Python · LoRA · SlimPajama · Hugging Face
Recall 1.000
across 40 screening cases, zero false negatives, held by a CI gate
A client-onboarding screening agent for a private bank, built on Microsoft Foundry. It screens a case against SECO sanctions and OpenSanctions PEP data, and returns a traced verdict backed by record identifiers. The real subject is not the agent but the apparatus that measures it: 40 ground-truth cases, a custom evaluator, and a CI gate that fails the build below 100% recall.
Microsoft Foundry · MCP · Azure Functions · Python · Entra ID / RBAC · OIDC
I was hired at Orange to do marketing. I taught myself to build, and turned my own role into an AI engineering one: nobody handed it to me. Now I build production GenAI by day and my own products at night. I care about one thing most: making AI you can actually trust, not AI that just looks good in a demo. I finish my Master's at Epitech in 2026. Until then, I let the work do the talking.
Idea to production, alone if I have to. Frontend, backend, infra, payments: I hold the whole chain.
"It feels good in the demo" is not a number you can ship on. I test my outputs against real cases.
The Epitech piscine every year. I even trained a small language model from scratch, just to understand how they work inside.
RAG workshops for PMs, mentoring interns, reusable patterns. What I build should be useful beyond me.
09/2024 - Present
Lyon, France
24-month apprenticeship (September 2024 to September 2026) within the Cloud Avenue product team, Orange's sovereign cloud. Hired as a junior Compute Product Owner, I grew the role into AI engineering: from a manual Excel benchmark and watch to AI and data tools in internal production. End-to-end development of « KARL », automated competitive intelligence and RAG chatbot demoed live to the Orange Business CEO in July 2025: LangChain + ChromaDB orchestration, multi-LLM (Llama 3.3 70B, DeepSeek R1, QwQ 32B) on H100 NVL and L40S GPUs via vLLM. Model proxy listed in the internal application catalog (February 2026). Sovereign AI demonstration prototype built with Claude Code in under a day for the global NVIDIA x Gcore webinar (June 2026). Competitive benchmarking platform covering European cloud providers, hyperscalers and LLM providers, results presented to COMEX, supporting a repricing review in spring 2026. Benchmarked 6+ LLM models across 3 GPU tiers (H100 NVL, L40S, L4). Automated AI newsletter pipeline (93+ weekly editions). Led RAG workshops for product managers, onboarded 2 interns on agentic AI projects. Handover prepared: documentation and tutorials left to the teams and incoming apprentices.
01/2023 - 06/2023
Mies, Switzerland
Pioneering generative AI project (from January 2023, just 2 months after ChatGPT's launch): end-to-end management of a GPT-4 chatbot for the FIBA Basketball World Cup 2023. Solution benchmarking (RFI), PoC, negotiation and contracting with BotsCrew. Coordinated a 2-3 person team + vendor, delivered in 4 months. Localization in 5 languages (EN, JP, ID, FR, ES), deployed on Facebook Messenger and web widget. 73,000+ fan interactions in 2 weeks, 71.51% conversation completion rate. Created analytical dashboards (BI) via GA4 and BigQuery (GCP). UI/UX improvements and digital marketing participation.
2018
Geneva, Switzerland
Summer internship at the Cantonal Social Insurance Office of Geneva.
Open to AI engineer, forward-deployed and solutions roles
Based : Geneva / Lausanne, Switzerland
Languages : French (native), English (C1), Spanish (B1)