Projects

A turn-based RPG that runs entirely in the terminal, with a 51x51 procedurally generated world, combat, enemy AI, and looting.

Modernized a C implementation of the Diffeomorphic Demons medical image registration algorithm into C++, integrating FFTW for a large complexity reduction. Built during the CNRS internship to stabilize retinal video sequences.

A 5-agent pipeline that continuously audits course curricula against the job market and state of the art, hallucination-checked. Placed 3rd at Hackathon IA Agentique.

A real-time AI copilot wired into Google Workspace that briefs you before meetings, listens live during calls, and recaps afterward. Built in 6 hours at the Gemini 3 Hackathon Paris.

End-to-end segmentation and classification pipeline for retinal Doppler holography images, now running daily in a hospital diagnostic pipeline.

AI-powered covenant compliance review for loan agreements, results shown as an interactive knowledge graph. Built at the Paris Fintech Hackathon 2026.

A GIMP 3.0 plugin automating the save-close-open-next annotation cycle for manual image segmentation, built to support the EyeSegmentation dataset work.

A simulated festival management system solving forecasting, anomaly detection, resource allocation, and scenario evaluation as independent Kafka-connected microservices.

A MCP server giving Claude fine-grained, structured access to Jupyter notebooks instead of dumping the whole file into context.

Collision-free trajectory planning for a drone fleet in continuous 3D airspace, modeled as CP-SAT, Conflict-Based Search, and space-time A*, with a Three.js 3D frontend.

A feedforward neural network written from scratch in C++, no ML libraries, trained on MNIST to classify handwritten digits.

A stock market data pipeline merging two mismatched real-world sources into a Dash dashboard with Bollinger Bands and normalized performance comparison.

Detects empty shelf space from a photo, forecasts demand, and has an LLM agent decide whether to restock or order from a supplier.

A GraphRAG pipeline combining knowledge graphs with LLMs for multi-hop question answering, benchmarked against classic vector RAG.

A real-time multiplayer tank battle game, written entirely by hand without AI assistance, with a custom game loop and full gamepad support.