About me

I am an Applied AI Scientist at Mistral AI. I work on post-training coding agents for cybersecurity (cybench, cybergym) using reinforcement learning. I am particularly interested in understanding training dynamics in reinforcement learning, including curriculum design and key metrics such as importance sampling ratios, gradient norms, entropy, reference KL, off-policy KL, reward signals, and pass@k, and how they relate to overall model performance.

I defended my Ph.D. thesis in “Multidomain Neural Machine Translation” in December 2021. I was supervised by François Yvon, Senior Researcher at CNRS and Sorbonne Université, and Josep Crego, Head of Research at SYSTRAN. Before my Ph.D. years, I graduated from Ecole Polytechnique (cursus d’ingénieur polytechnicien). I also obtained two other diplomas, including an M.S. in Data Science from Université Paris-Saclay and an M.Eng. in Telecommunication from Télécom ParisTech. Before my current job, I was a Research Scientist at Zoom Communications, working in on Speech Language Model.

I used to compete in maths when I was in college (in IMC 2012 https://www.imc-math.org.uk/imc2012/IMC2012TeamResults.pdf ) and high school.