A U-Net rebuilt from scratch in PyTorch (≈ 7.8M parameters) to segment cats and dogs pixel by pixel on Oxford-IIIT Pet, with an ablation study measuring what skip connections really bring.

mIoU 0.719 · skip connections: +0.096 IoU on boundaries

A reinforcement learning lab where every algorithm is hand-coded before being compared to Stable-Baselines3: Q-Learning (FrozenLake, Taxi), DQN (MountainCar), Actor-Critic PPO (LunarLander), and a Flappy Bird rebuilt from scratch as a Gymnasium environment.

Flappy Bird: 80 pipes per game, 20 out of 20 games without a crash

A complete system combining an embedding database with FAISS-optimised search: the face is detected with MTCNN, encoded into a vector by InceptionResNetV1 (FaceNet), then the closest matches to a query are retrieved in milliseconds.

Real-time vector search

The sequel to Deep Scan: telling a real face from an attack (printed photo, screen). Fine-tuning an ImageNet-pretrained ResNet18 for spoof / non-spoof classification, with a Streamlit interface for testing.

Team Kaggle competition: identify the animal species present in field sound recordings by fine-tuning a pretrained ResNet. A demanding evaluation: macro ROC-AUC, which weighs rare species as much as common ones.

Macro ROC-AUC: 0.87 (team)

An intelligent chatbot able to answer Bible questions using a RAG architecture (Retrieval-Augmented Generation). Verses are indexed as semantic embeddings and an open-source LLM answers with context.

Sourced answers, 100% open-source