Personal Projects

HPC recommender — tail latency vs request rate

HPC-Accelerated Music Recommender

CUDA · cuBLAS · OpenBLAS · OpenMP · C++ · Python

Matrix-factorization recommender for ~120K users and ~5.6M tracks from the Last.fm LFM-1b dataset. Both training and inference are implemented as a CPU baseline and a GPU-accelerated path, with end-to-end benchmarks on training wall-clock and tail latency.

View on GitHub
GPT-2 from scratch

GPT-2 From Scratch (124M)

PyTorch · Python · HuggingFace Transformers

From-scratch PyTorch implementation of GPT-2 (124M), including causal self-attention, transformer blocks, and autoregressive generation, without nn.Transformer or fused attention kernels. Verified against HuggingFace's reference to logit- and token-level parity.

View on GitHub