AI-powered Gomoku with AlphaZero self-play and real-time multiplayer. Extended beyond the course requirement with self-play training and production deployment.

• Trained an AlphaZero agent using PyTorch with MCTS self-play, scaling across four parallelism backends from sequential to Ray-distributed for near-production training.
• Built a C++ Minimax engine with alpha-beta pruning, principal variation search, iterative deepening, transposition tables, and killer move heuristics for real-time move computation.
• Deployed on Google Cloud with Docker and Cloudflare Workers routing. Built a Nuxt 3 frontend with real-time WebSocket gameplay and move evaluation.

Python · PyTorch · C++ · pybind11 · Docker · Google Cloud

DQN reinforcement learning agent that learns to play Snake.

• Built a DQN agent with a 24-dimensional relative state representation encoding wall, food, and body distances in four directional axes from the snake's heading.
• Designed multi-signal reward shaping with proximity incentives, BFS-based trap detection, and starvation penalties. Fully YAML-configurable with checkpoint persistence.

Python · PyTorch · DQN · YAML

Neural network built from scratch - no framework, manual backpropagation.

• Implemented a multilayer perceptron from scratch using NumPy with manual backpropagation, four gradient descent optimizers (SGD, Adam, Adagrad, RMSProp), and numerically stable softmax with cross-entropy loss.
• Applied He initialization, mini-batch training, and early stopping for breast cancer classification on the WDBC dataset with model checkpoint persistence.

Python · NumPy