Projects

Projects by Haoling Pu: a sparse attention CUDA kernel for Blackwell, a self-writing LLM wiki, hybrid retrieval, and more.

• Sparse attention on Blackwell· NVIDIA MLSys 2026 competition· code private

  A CUDA kernel that makes DeepSeek-style sparse attention fly on Blackwell. Passed every workload in NVIDIA's competition, dozens of times faster than the reference.

• SGLStreaming audio in SGLang-Omni· open source · sgl-project/sglang-omnicode ↗

  Contributor to the open-source serving framework for speech models. My work is on making the audio arrive on time, not just arrive.

• Agent Anchor· CMU capstone · BNY· code private

  Keeping production AI agents from quietly getting worse: a loop that notices when quality drops, finds the cause, and checks the fix.

• ML Notes· personal projectlive ↗code ↗

  Everything between a prompt and a token, one mechanism at a time: an interactive map of the LLM serving stack, from attention and the KV cache to agents and GPU kernels.

• LLM Wikicode ↗

  A wiki that writes itself: an agent reads sources, links pages, and cleans up its own contradictions. Karpathy's LLM Wiki idea, made real.

• CMUHybrid retrieval-augmented generation· CMU Advanced NLPcode ↗

  Keyword search and embeddings, fused so each covers the other's blind spots, with a local open-source model doing the talking.

• Agentic team knowledge system· Google internship· code private

  The agent that knows what the team knows. Shipped for real at Google, for teams beyond my own.

• EchoFrame· code private

  Hand it a faded old photo, get it back in color half a minute later. Built for my family's albums.

• PetFinder Procode ↗

  Lost a pet? Post it, and let matching do the searching.

• CMUSimultaneous MT with anticipated futurescode ↗

  An early bet that translating in real time gets easier if you guess the future first.

Nothing tagged that yet.