Projects

CNN built from scratch in C++/CUDA with hand-written backprop, im2col, and GEMM kernels. 99%+ on MNIST at ~10% the throughput of an equivalent PyTorch/cuDNN model on T4 GPUs.

124M-parameter decoder-only transformer with causal self-attention and learned positional encodings. Optimized for multi-GPU CUDA with minimized CPU-GPU sync — 11× speedup over the pre-optimized baseline.

A transformer built in pure NumPy — no autograd, manual backprop through attention and FFN blocks.

Convolutional neural net implemented from scratch in NumPy, including manual conv and pool gradients.

CNN that predicts age from brain MRI scans, trained on the IXI-T1 dataset.

CNN for classifying AI-generated vs. real images.

CAFA-6 protein function prediction pipeline.

Collaborative-filtering book recommendation system.

Classic Kaggle Titanic notebook — feature work and modeling.

Reinforcement learning agent for Settlers of Catan trained with Proximal Policy Optimization.

LLM + RL self-play for Codenames using SFT and GRPO on filtered rollouts.

Negamax + alpha-beta chess engine with quiescence search and a Zobrist-keyed transposition table. Tkinter UI.

Procedural game world generation engine built on Bevy ECS — GameSpec → validation → spawn rules → chunk output.

Real-time driver drowsiness detection using eye- and mouth-aspect ratios from a webcam feed.

RAG-based consulting assistant with a Weaviate vector store.

Automated job application tool with LLM-powered resume tailoring.

AI project planner on Cloudflare — Durable Objects for state, Workers AI for the LLM, React + WebSocket frontend.

Flask file hosting service with public/private visibility, CSRF, and session auth.

Flask web app for secure file storage with auth and hashed passwords.