I make large models serve faster: attention and MoE kernels, quantized GEMMs, and the serving-engine plumbing around them. I also build and ship iOS apps on the side.
- SGLang contributor - merged work on Gemma 4, MoE and DeepSeek sparse attention:
- FA4 on Hopper (SM90) with head_dim 512 for Gemma 4
- Faster w4a8 MoE for GLM-5.2 on H200 - up to 1.23× faster
- int32 offset overflow fix in the CUTLASS MoE pre-reorder
- DSA ragged seed-index fix before MTP draft decode
- In review: per-layer attention backends for Gemma 4, FlashInfer SM90 w4a8 CUTLASS MoE
- GPU kernels - throughput-focused C++ kernel work on AMD hardware
C++ · CUDA · HIP · CUTLASS · Triton · Python · PyTorch · SGLang · vLLM Swift · SwiftUI · TypeScript · React
LLM inference / ML systems roles and collaborations.
