Fares Naoui

I built an end-to-end autonomous driving stack in CARLA for my M.Sc. thesis, combining SAC control, GRU safety arbitration, and Qwen semantic arbitration. The work became a peer-reviewed Zenodo publication on neuro-symbolic co-pilots for hierarchical driving decisions.

I designed the decision and control layer, implementing 50Hz continuous steering, throttle, and brake control with Soft Actor-Critic and tuning rewards through 648+ Optuna trials. I also built a real-time GRU override layer that reached 90.1% validation accuracy and fixed pipeline issues affecting sensor data, vehicle physics, and steering smoothing.

I extended the research system into a reproducible, production-shaped ML service using FastAPI, Docker, MLflow, DVC, and GitHub Actions. My work spans reinforcement learning, computer vision, simulation, MLOps, and applied model deployment.

I've also trained autonomous drone agents with PPO and SAC, developed transformer-based image super-resolution models, built ECG rhythm classifiers, and created ML-serving REST APIs with JWT authentication.