Vincent Zhong — Resume

Vincent Zhong

GPU inference engineer and Computer Engineering student at the University of Waterloo.

emailgithubscholar

Professional work

Liquid AI — ML Engineer Intern, GPU Inference

May 2026 — present · San Francisco, CA

  • GPU-inference bring-up and performance work across SGLang, vLLM, TensorRT-LLM, llama.cpp, and Transformers.
  • Work spans model architecture, profiling, CUDA graphs, cache design, kernels, constrained decoding, and low-precision numerics.

Shopify — ML Engineer Intern, Search

Sep — Dec 2025 · Toronto, ON

  • Built a typo-correction pipeline that improved search relevance by XX% in A/B tests.
  • Improved query rewriting with SFT and RL, producing an XX% relative lift over the strongest baseline.
  • Ran retrieval-tuning experiments that improved relevance by XX% on head queries.

Yupp AI — Software Engineer Intern

Feb — May 2025 · Mountain View, CA

  • Built an agentic LLM tool-calling framework with a modular registry, external RAG connectors, and Google API integrations.
  • Built the backend for pairwise LLM preference collection: routing, judge evaluation, labeling, QA, and enrichment.

Open source

SGLang — Open-source contributor

Fall 2025 — present

  • 30+ pull requests across hybrid-model inference, speculative decoding, prefix caching, quantization, model integrations, and performance debugging.
  • Optimized GLM-5.x and DeepSeek-V3.2 with fused NeoX RoPE and RoPE–Hadamard indexer paths plus an opt-in MXFP4 index-K cache; worked on Kimi K2.x linear-attention stream syncs, MLA backends, and radix caching.
  • Brought up Spec V2 for Qwen3-Next and Qwen3.5; co-authored ReplaySSM's GDN speculative-verify path with Yuan Luo, including ring-state rollback, exact-fold numerics, and long-output validation.
  • Root-caused a FlashInfer sliding-window-attention numerics bug and integrated NVIDIA Model Optimizer MXFP8 weights, alongside text and multimodal model support.

Research

University of Waterloo · Data Systems Group — Systems Researcher

Nov 2024 — May 2025 · Waterloo, ON

  • Co-authored a SIGIR 2025 resource paper on learned dense and sparse retrieval in Anserini.
  • Integrated Snowflake Arctic text-embedding models and SPLADE encoders through ONNX.

Education

University of Waterloo — BASc Computer Engineering

Sep 2024 — Apr 2029 (expected)

Technical

Inference — SGLang, vLLM, TensorRT-LLM, llama.cpp, Transformers, FlashInfer

GPU — PyTorch, CUDA graphs, torch.compile, NCU, Nsight Systems, CuTe DSL, CUTLASS, Triton

Languages — Python, C++, Rust, Java, SQL, Bash