AI Product Manager. I turn AI agents into products that generate revenue — observable, evaluated, auditable.
Most AI pilots never prove their revenue impact. I ship the ones that do, and prove it in euros.
10+ years turning messy, regulated data into products that move the number.
I build and ship products with impact, not decks:
- +30M€ unlocked through data quality
- +8M€ from a pricing product at Decathlon
- 300M€/year billing platform at BNP Paribas
- −75% production time on core reporting
Built observability platforms tracking AI adoption across the BNP Paribas Group. Ran data teams, roadmaps, and stakeholders in regulated industries where failure costs millions.
Then I started building for myself.
| Project | What it does | Status |
|---|---|---|
| nanoAgent | Agent loop built from scratch, no framework — the nanoGPT of agents | 🟢 Live |
| EvalBlink | Benchmark prompts × models × tests — quality, cost, latency in one command | 🟢 Live |
| Awesome French AI | Curated map of the French AI ecosystem — labs, models, startups, people | 🟢 Live |
| The AI Compass | 10-day email course — AI for non-technical professionals | 🟢 Live |
| SASU IS Simple | Salary vs dividends optimizer for French freelancers | 🟢 Live |
| ValuePick | Value-investing screener with LLM-powered thesis generation | 🔨 Building |
Most people learn an agent framework before they understand what an agent loop actually is.
nanoAgent inverts that: a tool-calling loop written in plain Python, no framework. About 20 lines run the whole thing. Skills load on demand, the same progressive-disclosure pattern behind Claude's own Skills, rebuilt from first principles.
I ran it on financial tasks, across multiple models and providers, and logged every call. Three findings mattered more than the code:
- The provider is a second dial below the model. Same model, same question: one provider took 18s and followed every skill exactly. Another took 8s and drifted.
- Progressive disclosure isn't free. Loading skill descriptions on demand cost +95 tokens every turn. Loading full skill bodies upfront cost +266. On-demand only pays off once a good share of your skills go unused.
- One sentence beat every code change. A single line added to a skill file cut a 1,351-token turn down to under 90.
The loop is the easy part. What decides cost, speed, and correctness sits outside it — the model, the provider, and how the instructions are written.
That's the part most agent demos skip past.
→ Full repo, logs, and measurements
The gap between "we have an AI strategy" and "our AI product has 10,000 users" is where I work.
Real users. Real revenue. AI that earns its place in the business.
📍 Paris, France