I'm a software engineer working mostly on AI systems, and I care about one thing above all: shipping software that people actually use, often and safely.
I started in data science, so I'm comfortable with models and evaluation. But I've learned that the hard part of an AI product is rarely the model. It's everything around it: tests, evaluation in CI, architecture that lets you swap a model without breaking the product, and a team that can still ship fast six months after the first release.
- How great teams ship. Accelerate changed how I see my job: continuous delivery, small batches, trunk-based development, release decoupled from activation. I keep wondering why so few teams work this way, and what it takes to get there.
- Developers close to users. On eMush, talking directly with players changed our roadmap overnight. I'm convinced that the shortest path between the people who build and the people who use is one of the most underrated engineering practices.
- Making AI products boring to operate. Evaluation in CI, deterministic tools around the model, architectures where swapping an LLM doesn't break everything. The model is the easy part.
- Craft as a team sport. TDD, DDD, refactoring, code reviews that make everyone better, not just the code.
Where I'd like to grow: toward a hands-on tech lead / engineering manager role, helping a team build an ambitious AI product the way the best teams in Accelerate do.
I helped turn an internal RAG prototype into a production service used by 500+ people across three subsidiaries.
I built its automated evaluation pipeline around a business-curated Q&A dataset and LLM-as-a-judge, and worked on improving answer quality up to an 80% correct-answer rate. I also worked on the API, indexing pipelines, monitoring and reliability.
An agent built on OpenClaw that handles my emails, CRM and prospecting every day. I wrote its core myself, hexagonal and test-driven: the LLM decides, deterministic tools act, and nothing leaves the machine without my approval.
I was a lead maintainer of this open-source multiplayer game remake from 2022 to 2025.
I drove architecture, technical and product decisions, reviewed and mentored 20+ contributors, and helped grow the project to 2,500+ users.
Archéologie logicielle : reconstruire un jeu sans son code source
I've contributed to open source projects including uv, pytest, Haystack and Hatch.