I build AI agents and ship them as real products that people actually use.

Most of my work is hands-on: designing multi-agent architectures, wiring up tool calling and local context, and turning LLM capabilities into desktop apps, browser extensions, and workflow tools that solve real problems — not just demos.

Right now I'm building Deskhand, an AI desktop agent that reads your local files, apps, and system context to help you get things done — disk cleanup, data analysis, document translation, note organization — all through natural conversation with real tool execution.

What I work on

• AI Agents: desktop agents, browser agents, multi-agent teams, MCP-based tool calling, local-first agents
• Agent orchestration: hierarchical routing, sequential pipelines, planning loops, context passing between agents; tools: LangChain, LangGraph, Google ADK
• Prompt engineering: system prompt design, chain-of-thought, structured output, prompt management at scale
• RAG: embedding pipelines, hybrid search, chunking strategies, agentic RAG with tool retrieval
• LLM evaluation: A/B testing prompts, human-in-the-loop scoring, automated eval pipelines, vibe checks at scale
• Agentic engineering: spec-driven development, AI-assisted architecture, test-aware code generation, production-grade agent coding
• Product strategy: competitive analysis, user interviews, feature prioritization, go-to-market for developer tools

Deskhand — An AI desktop agent that actually does things on your computer. It reads local files, connects to your apps (Apple Notes, Excel, .docx), and executes tasks with your permission. Built with TypeScript + Anthropic API. Think of it as a local-first AI assistant that goes beyond chat.

PM-Agent — A multi-agent product manager assistant built with Google ADK. Feed it a product idea, it runs a structured interview, then generates a full documentation suite: market research, user personas, feature specs, wireframes, and interaction flows — all through a hierarchical 3-layer agent architecture.

Recently shipped a few smaller tools: Corn's Awesome Skills, PM-Agent, styled-resume, and BOSS Job Collector.

• AI Product Manager based in Shanghai — I design products and write the code to build them
• Shipped a Chrome extension from zero to 10k+ users as solo PM + developer
• Comfortable across the stack: Python for agents & backend, TypeScript/React for frontend, LLM APIs for the intelligence layer
• I care about building AI tools that work in the real world, not just in notebooks

Python TypeScript React LLM APIs Multi-Agent Systems Google ADK Anthropic API Tool Calling RAG