Full Stack Engineer & AI Developer @ Microsoft · Cloud-Native veteran turned vibe-coding believer
- 🛠 Software Engineer at Microsoft — working on developer tools for the AI-native era.
- ☁️ Previously: years of Cloud Native — Kubernetes, Spring Cloud, Azure container platforms, the whole microservices zoo.
- 🪄 Now: full-on vibe-coding believer. The most fun I've had as an engineer is sketching software with an agent at my side, in plain language, in flow.
- 🌏 Based in China.
One human + a team of AI agents — that's how I'd like to ship software in 2026. Most of my recent work explores what that team looks like: the tools, the editor integrations, the orchestration glue, and the lived workflow.
| Project | One-liner | Stack |
|---|---|---|
| 🚁 dev-pilot | AI Multi-Agent engineering team — one developer, full team power. Automates the entire dev lifecycle from coding to CI | Claude Code · GitHub Copilot |
| 🛡️ DevSquire | AI Agentic Engineer for VS Code — autonomously develops issues, reviews PRs, monitors CI with real-time dashboard | TypeScript · VS Code Extension API · Claude Code |
| 🏺 code-archaeology | Why is this code here? — a VS Code extension that aggregates git + PRs + issues into a streaming three-layer answer | TypeScript · VS Code Extension API · GraphQL |
| 🪵 codezen | Turn every keystroke into merit — a mindful coding companion with wooden fish animation, rank system, and anti-cheat | TypeScript · VS Code Extension API · SVG |
| 📚 ai-reading-mate | A warm, mobile-friendly reading dashboard + AI book recommender — bring your own reading data | Next.js · Claude Code · MCP |
| 🐛 mcp-server-weread | MCP server for WeRead, with extensions for booklists | Node.js · MCP |
- Software is becoming a conversation, not an artifact. The interesting question stops being "what did you build?" and becomes "what's the loop you and your agents run together?"
- Most agent failures are taste failures. The model can technically do the thing; the system around it just had no opinion on how the thing should feel. Latency, source-attribution, when to stay silent — these are product decisions, not model decisions.
- The unit of work is shrinking. A PR used to be a feature; it's becoming a paragraph. Tooling needs to catch up to that pace — explanations, reviews, and context retrieval all have to land in seconds, not minutes.
- Cloud-Native taught me observability; agents need the same. If you can't trace why an agent did what it did — which tool, which doc, which prior turn — you don't have a product, you have a vibes generator.
- 🛡️ Shipped DevSquire v0.1 — one developer + DevSquire = a full engineering team
- 🏗 Building code-archaeology — M0 shipped, M1 will integrate the VS Code Language Model API (Copilot)
- 🪵 Shipped codezen v0.1 — write code, earn merit, become enlightened
- 🧪 Iterating on dev-pilot — pipeline orchestration for AI engineering teams
- 📖 Reading: agent design patterns, evals, and the long tail of "tool use" failures
"Talk is cheap. Show me the agent."