Zoe
Zoe

About

I'm Zoe — an AI Infra Engineer based in Hangzhou, China. Ten years of code in Go, Python, Flutter, and Rust.

First half of my career: distributed systems, container scheduling, GPU clusters. Last two years: full-stack on AI inference. Today I lead an in-house LLM-serving platform at a top-tier internet company.

What I do now

AI Infra — I work with LLMs, vLLM, SGLang, and GPUs every day.

My day-to-day revolves around:

  • LLM serving: turning large models into production services, PD-disaggregation, KV-cache transfer, long-context optimization
  • GPU clusters: multi-GPU tensor parallelism, multi-node pipeline parallelism, NCCL/RDMA networking, topology-aware scheduling
  • Inference acceleration: vLLM / SGLang / TensorRT-LLM deployment and tuning, quantization, continuous batching
  • Platform engineering: making inference capacity stable, schedulable, and cheap for the teams that use it

Off-hours I build AI ops tools — distilling the repetitive parts of AI engineering into one-line CLIs and open-sourcing them.

Background

I started on Python and Go — crawlers, data pipelines, distributed application frameworks. Then I pivoted to cloud-native: container runtimes, Kubernetes, GPU scheduling.

From 2024 onwards, as LLM inference moved into thousand- and ten-thousand-GPU clusters, I went all-in on AI Infra: model deployment, inference optimization, large-scale serving. These days I'm wrestling with throughput numbers, VRAM, and network bandwidth.

What I'm building

I'm good at shipping things — and for the first decade I shipped them for other people. Starting in 2026, I'm building for myself.

The goal: a portfolio of small, profitable products. Real users, real revenue, no VC.

distro — a content distribution CLI

Write one Markdown post, and it rewrites for Twitter threads, Juejin, Zhihu, Xiaohongshu, English blogs. Bonus: generates social cards, renders Mermaid diagrams, and flags AI-writing tells.

automagent — mobile AI agent

Built in Flutter. It sees the screen, understands the UI, and drives your phone. Runs on-device, no laptop tether. Multi-model, cross-platform.

Multi-agent workflow

I run a small fleet of agents to manage day-to-day engineering and writing. Three roles, three lanes:

  • Z plans and dispatches
  • Watch monitors and gathers
  • Build writes code and ships projects

They cooperate via a file protocol, producing daily plans, summaries, and patches — fully auditable.

Stack

Primary: Python (inference / serving / training integration), Go (backend / platforms / CLIs), CUDA-adjacent tooling

AI inference: vLLM, SGLang, TensorRT-LLM, Triton, Ray, NCCL

Infrastructure: Kubernetes, Docker, RDMA / InfiniBand, GPU device plugins, NRI

Also using: Flutter (mobile agents), Rust (systems-level tools), React/Next.js (frontend)

Taste: clean architecture, use my own wheels when I have them, allergic to over-engineering, obsessed with wrapping hard problems behind a one-line command.

Open-source projects

What I write about

Mostly:

  • AI Infra in the wild: LLM-inference tuning, GPU scheduling, RDMA, platform-engineering war stories
  • AI ops tools: how to automate dev workflows with AI and good tooling
  • Passive-income experiments: building sustainable revenue from zero, with monthly public numbers

I write in Chinese first; distro fans the post out across platforms — this English version is one of those outputs.

This site

Built with my own nextjs-starter-zoe-app. Next.js + Tailwind + shadcn/ui, YAML-driven config, Markdown content.

Working with me

Outside my day job, I keep room for one or two external projects a year.

Usually one of these:

  • AI inference / LLM Serving — Help your team stand up or optimize an inference platform from scratch. GPUs, scheduling, RDMA, cost.
  • AI product 0→1 — Partner with founders to take an idea to a working v1. I move fast on the Next.js / Cloudflare / Supabase stack I use every day.

Not for: UI implementation gigs, pure front-end work, or staff-aug at large orgs.

If something resonates, [email protected] or Telegram — one line about "what you're building and what's stuck" is enough.

Contact

Happy to talk shop on AI infra, LLM inference, open-source tools, or indie SaaS. Skip the formalities and just ping me.