Built and open-sourced RecoAtlas, Criteo’s first LLM recommendation agent and benchmark, integrating retrieval, complementarity, and diversity tools that increased held-out product recovery from 1–3% to 51–54%; adopted by internal teams for production and research prototyping, with a live demo available upon request.
Co-developed and analyzed a ratio-free off-policy RL objective for LLM reasoning, maintaining stable reward and entropy across a 100× range of regularization strengths and policy lags from 32 to 256 optimization steps.
Developing an RL objective for LLM reasoning that reuses multiple reasoning traces through answer-level marginalization, designed to reduce gradient variance and improve sample efficiency.
May 2022 - Present · Paris Area, France
Developed scalable on-policy and off-policy RL methods for decision-making over millions of actions, resulting in 8 top-tier publications, including 7 as first author, an ICML Oral, and a NeurIPS Spotlight (top 2%), plus 6 workshop/tutorial papers and 15+ technical blog posts.
Fine-tuned a CodeBERT-style transformer for AWS CodeGuru to detect non-compliant code, improving detection of security-rule violations by 42% and best-practice violations by 5%.
Jan 2022 - May 2022 · San Francisco Bay Area, USA
Selected as one of three graduate students from France to receive generous funding and mentorship
Sep 2021 - Sep 2022 · Paris Area, France
Built probabilistic models for recommender systems, showcasing the effectiveness of integrating products ranking with reward feedback in two workshops at RecSys and KDD 2022
Apr 2022 - Sep 2022 · Paris Area, France
Improved flight-search prediction performance by 32% using sequence-to-sequence neural models
Apr 2020 - Sep 2020 · Nice Area, France
Designed time-series models using temporal word embeddings to detect semantic shifts over a century of text
Oct 2019 - Jun 2020 · Lille Area, France
Built a Word2Vec-CNN classifier for short, noisy bank transactions with 93\% accuracy; used to identify errors in the existing production rule-based system
Jul 2019 - Sep 2019 · Lille Area, France