I'm a Pre-Doctoral Researcher at Inria Paris in the ALMAnaCH team, working under the supervision of Djamé Seddah on Mechanistic Interpretability and LLM safety. My research focuses on understanding the internal mechanisms of LLMs, how they represent and process language, and how we can use that understanding to improve them.

I worked with Armel Randy Zebaze, whose PhD focuses on low-resource machine translation, on disentangling sentence equivalence from target language in LLM-based translation. I also worked with Wissam Antoun, who worked on the training of the Gaperon French LLM suite, on studying how triggers hijack model representations to induce backdoor behaviors.

As part of my work at Inria, I also ran continuous pre-training of the Gaperon models across multiple nodes, and built an end-to-end evaluation library that automates LLM benchmarking on HPC, making it easier to run evaluations at scale with Slurm.

On the side, I'm building Straufi, a no-code platform for designing and backtesting automated trading strategies, which placed 3rd at the EFREI Innovation Project (awarded by Microsoft, Thales, and Capgemini officials). I hold an M.S. in Computer Science from EFREI Paris with a major in Data & AI, and spent a semester at UC Irvine.

When not doing research, I train and compete in kickboxing 🥊