Helped the Smol team post-train SmolLM3 to be better at tool-use: this was my discovery of SOTA model training. It was really fun, and the model reached the Pareto Frontier on tool use. But there would still be lots to do!

Developed ScreenSuite, a comprehensive benchmarking suite for evaluating GUI agents across perception, single-step, and multi-step agentic behavior. Built on smolagents, it provides a standardized evaluation framework for Multi-Modal LLMs interacting with graphical interfaces.

GUI agents are agents where the LLM interacts directly with a GUI through vision: the first renowned one is Claude Computer Use. At Hugging Face, I participated to building an open version of it.

Built an open version of OpenAI's Deep Research in 24 hours, reaching SOTA results on GAIA benchmark. (to be fair I had started building on this idea before they published)

Smolagents is a leading framework for agents, with over 20,000 stars on GitHub. It's the first framework to propose agents that think in code.

Developed a predictive model for Bordeaux wine prices using weather data across grape growth stages. Despite agriculture experts telling this project was doomed to fail, this model predicts vintage quality more accurately than expert ratings.