At Mirakl, I developed the AI catalog Onboarder system, which processes 350M products per year. I lead 10 ML engineers.
I led post-training for an LLM specialized in retail products, using synthetic datasets to improve fill rate by 30% and reduce cost by 60% compared with state-of-the-art LLMs. I also redesigned evaluation and annotation campaigns, reducing judge-human agreement error by 77%.
At Doctolib, I built a medical evaluation harness in Python used by 20 research scientists to run 40 experiments a day. It became a release-quality condition and caught two major issues before production.
Earlier, I delivered recommendation systems at Fubo that increased engagement by 400%, and bidding model improvements at Criteo that increased company revenue by 0.4%. My work also includes ML evaluation, agentic AI, causal inference, and computer vision.

