At LATMOS (CNRS / IPSL), I designed the two-channel representation behind a diffusion-model research paper and built the data pipeline for Météo-France radar fields. I found and fixed a preprocessing cap that cut rain-coverage error from 19% to 0.8%; the paper is under review.
In my SQL Analyst Agent + MCP Server project, I built a tool-calling SQL agent and exposed its tools through a read-only MCP server. Database-layer guardrails blocked all 13 write attempts in a 20-attack prompt-injection test, and I measured how each component affected accuracy.
At Shul Technologies, I built Pega web applications and integrated them with external client systems before moving into data engineering, where I wrote SQL and Python for migration ETL. I’ve also built a BERT-powered mental-health analysis platform and a PyTorch CNN for MRI tumour detection.
