ML systems engineer working on LLM inference, GPU kernels, and reliable agent runtimes.
I like working close to the model/runtime boundary: finding out where a system actually spends its time and memory, then building around that. I mostly learn by building the thing and measuring it, and I try to be honest in the benchmarks about what they don't show.
PatchQuest is a durable runtime for long-horizon software agents. Runs are checkpointed, so if a worker dies another one picks the run up where it stopped, and the patch still lands once. Any run can be replayed or forked from a checkpoint. Commands go through a sandbox and an approval policy, with a workflow engine and an eval harness on top.
Inference-Kernels is my attempt to understand LLM inference by writing it: KV cache, paged attention, fused Triton kernels, continuous batching, quantization, speculative decoding. Each piece has a correctness check and a benchmark. It's educational, and the README says so.
WarpServe goes lower, into CUDA. Small custom kernels first, then the same ideas measured against PyTorch and CUTLASS in an end-to-end serving loop on a laptop GPU.
EigenTune freezes a layer's SVD bases and learns only a small rank-r scaling of the singular values, as a parameter-efficient fine-tuning method. pip install eigentune.
TransKV: transactional KV caching for speculative decoding under paged KV memory. TechRxiv preprint.
A few merged contributions to CocoIndex, and an open pull request to vLLM on mixed-dtype RMSNorm quant fusion.