About
Eight years putting ML and then LLMs into production, first on pricing at Amazon, then on product catalogs in SaaS. Today, agents in production every day, and room kept every month to work inside other teams’ repos, where the agentic setup has been installed for months and still returns a fraction of what it could.
The path
I started at Amazon, in Luxembourg, five years and most of them on European pricing. I came in through supply chain optimization and then AWS architecture, with four certifications passed in four months of internship, including the Solutions Architect Professional, the highest level there is.
Then pricing, as the only data scientist on a team that went from three people to thirteen. Causal inference on datasets of tens of terabytes, hierarchical Bayesian price elasticity models, hedonic prices estimated on text and image embeddings with ELMo and then fine-tuned BERT back in 2019, and a cross-country competitive price intelligence platform worth around €400M of additional annual revenue in European markets.
Out of that came the price volatility monitoring program, which I started and which ended up adopted across the group worldwide. Thirty billion price changes analyzed, an anomaly caught across a billion visits, 500,000 items with no reference price recovered right before the Black Friday peak, a dozen million euros of revenue brought back, and the European Just Do It Award for it.
Then Unifai, where I was the company’s first ML engineer. An end to end MLOps pipeline on GCP for retail groups, and language models from before ChatGPT, quantized FLAN-T5 to categorize industrial catalogs nobody could normalize by hand. That work is what put the company in a position to be acquired, and I carried the technical due diligence on the architecture and ML infrastructure side through to the Akeneo acquisition in 2023.
Then two and a half years at Akeneo as Tech Lead of the Core AI team. I architected and ran the internal inference platform every product team depended on, and behind them more than 900 enterprise retailers: around thirty million product attributes generated by LLM every month, 99.999% uptime, p95 under 150 milliseconds, and all model traffic centralized behind a single proxy, with failover and cascading across GPT, Claude, Gemini and Mistral, which took 40% out of the inference bill. And the Data Architect Agent, a multi-agent system with human gates that took catalog onboarding for an enterprise retailer from several months down to a few days.
Independent since 2026.
From the tool to the harness
At Akeneo I got my team onto Cursor, then onto Claude Code as soon as it shipped. I was the one other teams came to when they wanted to know which tool was worth it, and I pushed the subject with leadership.
Then I moved from the tool to the harness, which is where most of it is decided: CLAUDE.md files versioned in the repo, so that everyone’s harness gets better without everyone having to look after their own on the side. My setup spread to other teams and ended up as the basis for internal training.
What I think
The tools are here, and they are underused everywhere I go
I have never yet seen a team where the tool was the limiting factor. The trap is elsewhere: you install the agent, you have it write a few tests, it works, and that first win is enough to convince everyone the tool has been mastered. That is usually where it freezes. From that point on, what is missing is what you give the agent to read before it starts.
What I am aiming at is an agent that becomes a second brain, responsive, one you delegate to for real. The gain comes back as ambition, as going after things you would not have let yourself start before. The Claude Code team describes its own usage in roughly those terms, which lines up with what I see without proving it.
Harnesses are roughly equivalent today
That is not “the tool doesn’t matter”, which is false, and which sounds like a salesman talking down what you already know so he can talk up what he is pushing. It is a dated observation about the state of the market: harnesses have converged, so the gap has moved somewhere else, to what you put inside them. The CLAUDE.md files, the hooks, the skills, the structure of the repo, what the agent can check on its own before handing back to you.
If a tool takes a clear lead again in six months, the sentence falls and I will say so.
What counts is the date, not the artifact
Describing what I built is not worth much on its own any more. Automatic review of pull requests by an agent, GitHub has offered that natively since April 2025, and turning it on is a checkbox in a branch ruleset. Presenting it today as a feat means presenting yourself as someone who has just discovered the subject.
So the part of my work that mattered most is not the review tooling, which GitHub has made ordinary since. It is CLAUDE.md files versioned across a whole team, and that part is still rare.
Searching publicly, the only teams I found doing it seriously are GitLab, Camunda and Ledger, and across ten thousand or so public feedback comments my own agents went through, no team names an owner for those files. I am saying that from the outside and from public sources, so if your team already does it, write to me, that interests me more than the opposite.
What I am working on
I am full time on an enterprise AI mission at a CAC 40 company, under NDA. I will say nothing more about it here, and that is the same discretion you get from me if we work together. At a large company, what blocks agent adoption is almost never technical: it is who decides, on what grounds, and at what point somebody writes the decision down somewhere.
Alongside that, a few days a month stay free for short engagements, for teams that want their harness to actually pay off. An audit of the agentic setup to find out where you stand, a workshop day inside your repo, or regular office hours to hold the thing together over time. The audit is the simplest way in.
I also write now and then about what I break and what I fix along the way.