Computer Science at the University of Waterloo. I build low-level systems in C++, CUDA, Rust, and Go, with a focus on performance, memory efficiency, and correctness.
Native LLM Inference Engine — C++ / CUDA
Built a custom inference runtime for a 27B-parameter Qwen model on 16 GB VRAM, implementing NVFP4/FP8 kernels, Gated-DeltaNet, attention, and asynchronous CPU-to-GPU weight streaming. Validated numerical correctness against a NumPy reference.
Optimized 27B LLM inference on NVIDIA Blackwell, improving single-request decoding throughput from 195 to 256 tokens/s (+31%) through speculative decoding, CUDA graphs, optimized GEMMs, and L2 prefetching.
Building an x86-64 to C decompiler from scratch, featuring an intermediate representation, instruction lifting, control-flow analysis, and translation validation. Verified x86 flag semantics through 9.7 billion hardware comparisons without mismatches.
Built a multi-agent runtime with concurrent task scheduling, provider abstractions, persistent execution state, tool orchestration, and MCP integration.
Developing a coordination platform enabling coding agents to exchange authorized context, negotiate changes, and produce evidence-backed proposals through authenticated APIs, durable storage, background workers, and isolated execution.
Repository currently private.
- LLVM — contributed a proposed TailCallElim correctness fix with regression tests (PR open).
- Haven-HV and SkinSight — 7 merged PRs across AI infrastructure, ML, and application development.
Interested in inference engineering, compilers, systems programming, and high-performance computing.