Welcome! I'm an AI/ML Research Engineer @Confidential, based in Paris. My work focuses on building more efficient and trustworthy AI systems, with a particular interest in LLM optimization, AI safety, explainability, alignment, and neuro-symbolic AI.
An autonomous research agent for data-science competitions. From a competition link it grounds ideas in prior work, ranks hypotheses on an academic rubric, implements and validates the top pick against the real metric, and submits only if a contribution judge approves. Its AutoCompete kit lets any engine (OpenEvolve, LLM agents, AutoML) run on any platform (Codabench, Kaggle, local) through one protocol.
A dataset of 26k RGB images of wheat grains from 8 varieties, collected over two years of field trials in Saclay. Our classification pipeline reaches up to 94% accuracy and estimates varietal proportions with under 4% absolute error. Under the inter-year domain shift, fine-tuning the whole network always beats a frozen pre-trained feature extractor.
Most work treats political bias as a fixed property of a single model. We track how political framing and sentiment shift across model generations such as the LLaMA family, as training data, safety data and RL-based alignment change between releases.
A Codabench challenge on LLM judges that score job offers, find legal and quality flaws, and explain them. Built on 140 surgically edited offer pairs. I built the judge baseline and the multi-metric scoring, and benchmarked open-model inference on T4 GPUs.
Retrieval-augmented generation combined with AI agents, so that domain-specific LLMs for education give grounded answers. Presented at the 7th International Conference on Information Technology and Modeling.
Contributions to Hugging Face's kernel-builder, a toolkit for building and sharing custom compute kernels, and to gromo from GrowingNet, a library for neural networks that grow during training.
LinkedIn
GitHub
Google Scholar