I make large models run well on real GPUs.
I study computer engineering at the University of Waterloo. Most of my work is in SGLang, on the kernels, schedulers, and caches that decide how fast a model runs once it is served.
This summer I was a GPU inference intern at Liquid AI in San Francisco, where I brought Liquid's new hybrid models up in SGLang, vLLM, and llama.cpp and made them fast to serve. I also built a custom production stack for serving an LFM at low latency for a semantic-ID use case, from the kernels and the C++ engine up to the serving layer (a writeup is coming soon), and chased down performance and correctness problems across the stack, on NVIDIA and AMD GPUs alike. Before that I worked on search at Shopify.
I like problems where you have to understand the model and the hardware at the same time: low-precision numerics, speculative decoding, and the overhead that hides between kernels. I measure before I believe, and I write up the experiments that failed along with the ones that worked.
May–Aug 2026
ML Engineer Intern, GPU Inference
Brought Liquid's new hybrid models up in SGLang, vLLM, and llama.cpp, then made them fast to serve. Built a custom low-latency production stack for a semantic-ID use case, from kernels to serving.
2025–present
Open-source Contributor
New model support and serving speed: speculative decoding, fused kernels, prefix caching for hybrid models, and the correctness bugs found on the way.
Sep–Dec 2025
ML Engineer Intern, Search
Used LLMs to correct typos and rewrite queries for Shop.app search. The best version raised search relevance by 2% in A/B tests.
Quantized checkpoints of Qwen3.5, Gemma 4, and Kimi K2.5 that load straight into SGLang. They have been downloaded more than two million times.
An NVFP4 matrix multiply for the B300, written in CuTe DSL. At 9.23 PFLOP/s it beat the seven other implementations I benchmarked, cuBLASLt, cuDNN, and CUTLASS among them.
The short causal convolution inside hybrid models, rewritten in CuTe DSL and tuned shape by shape for the B300. It installs with pip and plugs into SGLang.
369 runs of vLLM and SGLang on the same GPUs and workloads, and writeups of the 20 bugs that turned up along the way.
A reading map of 2,154 paper summaries, arranged by date, by topic, and by which paper builds on which.
magic-trace silently dropped GPU kernels that overlap through PDL, so traces showed less work than the GPU did. This Perfetto build keeps them. Perfetto fixed the same bug upstream in v57.
Does diagonal preconditioning improve the memory update in Gated DeltaNet-2? Five variants, paired runs, from 200M to 1.3B parameters. It doesn't.
Interactive diagrams of how data moves through a model, for when reading the modeling code is too slow.
How low can a language model trained on OpenWebText in 45 minutes on one GPU go? Stanford CS336's leaderboard task, measured on a B200: 2.957 validation loss, against 3.035 for the best entry. It uses outside kernels the rules forbid, so it measures the ceiling rather than competing.
Python bindings for AI2's Rust tool that checks whether a benchmark leaked into a model's training data.
Not finished, not published yet.
DeepGEMM's fused router GEMM and top-k, rewritten in CuTe DSL. On DeepSeek-V4 shapes it beats every other kernel at small and medium token counts, and still loses at large ones.
Tuning FlashInfer's MegaMoE kernel at DeepSeek-V4-Pro shapes with IKET, an experimental profiler that records events from inside the kernel. It is within a few percent of DeepGEMM, and the rest of the gap is now mapped.
In asynchronous RL, new weights can arrive in the middle of a rollout. What does it cost to keep the old KV cache instead of recomputing it? Measured bit-exactly on Qwen3.5-9B over about 24M tokens.
The public source code for NVIDIA's next GPU architecture, SM107, collected and compared against Blackwell to find what changes for people who write kernels.