I got into AI through a fascination for how language works, and that curiosity quickly turned into a specialization in NLP, transformers, and learning systems that can operate in messy, real-world conditions. I graduated top of my class from Université Sorbonne Paris Nord with a Bachelor's in Computer Science, which gave me a strong software engineering foundation before I dove into ML.

For the past two years, I've been working as a Data Scientist at Safran Aircraft Engines, where I fine-tuned a domain-specific language model (AeroModernBERT) on FAA technical manuals, improving multi-label classification F1-score by 16%. I architected and deployed a GenAI document extraction application on AWS using ECS, Textract, and Bedrock, achieving 80% recall on complex sourcing tasks. I built a preprocessing pipeline that cut processing time from 40 minutes to 15 seconds. And I spent a lot of time on what I think matters most: making sure models are actually trustworthy. That meant running SHAP explainability audits that uncovered critical biases, like a classifier relying on technician signatures instead of real features. Before Safran, at KIA France, I built two production web applications for vehicle fleet and IT asset management. That experience taught me how to write software that other people depend on.

Outside of work, I spend my time on research. My latest project, Privileged World Supervision, proposes a new way to train reinforcement learning agents on open-ended information retrieval and knowledge graph reconstruction. The core idea: generate synthetic environments where ground truth is fully accessible, so we can compute exact per-action reward signals instead of relying on noisy value approximations. I engineered the full pipeline, from synthetic world generation using Claude 4.5's API to fine-tuning Qwen3-30B with LoRA adapters and training with PPO. 

I also enjoy building under pressure: my team placed 5th at the AgentsIRL Hackathon by Google and Nvidia, where we built TruthTrace, an end-to-end AI debate fact-checking system combining speech-to-text, speaker diarization, and LLM-based claim verification, all in less than a day.