Projects

Built at Hack the Worlds, a hackathon organized under Yann LeCun . An action-conditioned latent world model (EB-JEPA) that predicts future latent states from wrist-camera video and proprioception on surgical data. Includes ablations on the anti-collapse terms, and a Transformer predictor that reduces autoregressive compounding error.

🏆 Won 1st place at Unboxed, a 30-hour hackathon with GE HealthCare and Mistral AI. Our pipeline reads a CT scan in 3D, segments it, and works out where each tumor is and how big it is. A Mistral agent then pulls the patient's previous reports and writes two follow-ups: one for the radiologist, one the patient can actually read.

Wrote a review paper on different ways to detect drones with a focus on spiking neural networks. Fine tuned several spiking neural network architectures (Spiking YOLOv8, Spiking SSD...) on a fraction of the FRED dataset for drone detection. Achieved 92% mAP with a 69M parameters model.

Designed and developed a low-cost autonomous embedded system (< €150) mounted on a drone for real-time ozone concentration measurement using an ESP8266 microcontroller. Integrated electrochemical ozone sensor (SEN0321), GPS module, SD card for local logging, and Wi-Fi connectivity to transmit data to a remote server built with Flask + InfluxDB + Grafana. Personally handled the full hardware design, wiring, embedded software development, and network integration. Achieved ~80% correlation with a professional €3000 reference sensor, demonstrating high reliability, energy autonomy, and lightweight design for drone deployment.