Projects

CNN built from scratch in C++/CUDA with hand-written backprop, im2col, and GEMM kernels. No deep-learning or matmul libraries. 99%+ on MNIST at ~50% the throughput of an equivalent PyTorch model on T4 GPUs.

124M-parameter decoder-only transformer with causal self-attention and learned positional encodings, following “Attention Is All You Need.” Optimized CUDA training for an 11× speedup over the unoptimized baseline.

Testing in-context memory compression to enable longer-term retention while preserving response quality.

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

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

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