LLM infrastructure engineer — Rust + CUDA inference engines, file/storage layers for training data, tooling for agents that operate real machines.

Research Scientist @ MindLab-Research · previously SRE @ DeepSeek, storage @ ByteDance / Volcengine · crypto grad out of HUST

Treading stars to exhaust the Truth.

Start here

• pegainfer · 730★ — an LLM inference engine in Rust + CUDA, no PyTorch on the box. I contributed its Qwen3 LoRA serving path.
• DSH-better-sidebar · 4783★ — an open sidebar where file editing, terminal, Git and sub-agent views are first-class pages, not bolted-on plugins.
• curvine · 1034★ — a CNCF Sandbox project I contribute to: a high-speed cache with file semantics in front of cloud object storage.

Reach me — axi.moe (blog) · WeChat Nolan-Ho · github.com/NolanHo

pegainfer · 730★ · Rust + CUDA

For teams that have to put a model behind an API without dragging a Python runtime into production. I contributed its Qwen3 LoRA serving path: request-level adapters, tensor-parallel adapter loading, mixed LoRA batches, and the router extension that keeps the vLLM-compatible front end in step.

DSH-better-sidebar · 4783★

For people who spend the day inside an agent sidebar: extensions register their own pages and stand on the same footing as the built-in file rendering/editing, terminal, side chat, Git and sub-agent views — so the sidebar is a platform, not a fixed set of tabs.

curvine · 1034★

For ML teams whose datasets outgrow local disks: cloud object storage gets a file-semantic layer plus a high-speed cache, so training jobs read it like a real filesystem. CNCF Sandbox project — I am a contributor here, not the originator.

Upstream & org contributions — authored PRs: pegainfer-project 6 · omdsh-dev 10 · CurvineIO 1 · vllm-project 1 · tile-ai 1 · ray-project 0. Pinned upstream repos: vllm, ray, tilelang.

Megumin poured her whole life into a single spell, explosion, and refused to diversify. Same instinct here: pick one system, take it apart to the last layer, make it run reliably.

何か成す者とは歩み続ける愚者である、成せぬ者とは歩めを止めた賢者である

"成一事者，矢志不渝之愚者；毁一事者，停滞不前之贤者"

_{—— ロクでなし魔術講師と禁忌教典}

• rayrust — a Ray SDK for Rust.
• termdeck — persistent PTY sessions for automation agents, with CLI control and observe-only web replay.
• obsharness — an observability harness that lets an agent search, narrow and inspect a production incident.
• DSH-shellout — runs agent commands in a supervised out-of-process runner, so a hung command can never block the host event loop.
• pi-secret-mask — keeps secrets out of the model context while still allowing ${secret:...} references at tool boundaries.

_{113 public repos — production infrastructure alongside experiments. The three pinned projects above are where I would start.}