High Performance
GPU Kernels.
Built from the Metal Up.
I build and optimize CUDA kernels, profile at scale, and engineer ML infrastructure that delivers measurable performance.
cuda-ml-accelerator
Tiled shared-memory GEMM kernel (TILE_WIDTH=32, float32) with Nsight-driven tile size optimization, pybind11 Python bindings, and full benchmark validation against CPU, naive CUDA, and PyTorch/cuBLAS across five matrix sizes.
cuda-kernels
The first CUDA project of the summer — vector addition, naive GEMM, and tiled GEMM (double precision), profiled with Nsight Compute. Found the FP64 pipeline bottleneck that directly motivated the float32 rebuild in cuda-ml-accelerator.
CUDA Numerical Integration
Trapezoidal rule and Simpson's rule implemented with CUDA parallel reduction, built alongside Calc I coursework this fall.
GPU Kernel Library
A library of high-performance ML kernels — Flash Attention variant, Softmax, LayerNorm, and multi-GPU GEMM — benchmarked against cuDNN.