靳泽旭 · Jack · Shanghai, China · @jackjin1997 · email
Hi, I'm Zexu, Jack to friends. I'm an AI engineer at WuXi Biologics in Shanghai, working on LLM agents. Before that I was the server-side tech lead at Starblazer, where I led a five-person backend team and built agent systems that run, recover, and stay maintainable. Nearly six years in backend, the last three on LLM products. When a framework breaks, I read the source and send the fix upstream.
My architecture test is simple: boundaries should be clear, failure should be visible, and complexity must buy a real outcome. New technology that misses those tests is just overhead.
Working onIMEANAI-CoyageDenoiseJack Wiki
Creator ofAgentBench-LiveClawForgeRoadbookA little ocean
Contributing toLangChainLangGraphhermes-agentWebHarboropenclawWrenAI
Stack · Deep Agents · LangGraph · TypeScript · Python
Stack · LangChain · LangGraph · LangSmith · MCP · TypeScript · Hono · GraphQL · Kubernetes
Stack · Vue · React · Next.js · NestJS · PostgreSQL · Redis · AWS Lambda · Puppeteer
Zhejiang University City College · B.Eng. Software Engineering · Intermediate Software Designer
A daily tech-signal judgment product — doesn't mirror raw GitHub Trending, it tells you what's worth your attention vs. what's noise. For indie developers, tech leads, and investors: 3 minutes a day to stay on top of tech trends. Engineering: Python pipeline is the sole write path, Node (Hono + Next 15) stays strictly read-only; document-first + test-first workflow.
Python · Next.js 15 · Hono · PostgreSQL · pnpm Monorepo · LLM Summary · EvalsStarted as side-product · Becoming a productdenoise.info
AI-powered skill roadmap generator. Input JD/article/resume, LangGraph Agent auto-extracts skill trees, runs Tavily web research, generates structured learning paths. D3.js radial force graph visualization, multi-source merging, RAG chat, Obsidian export.
LangGraph · React 19 · D3.js · Express 5 · SQLite · Tavily · Vite · Tauri182 Tests · 98% Coverage · Multi-lang · Obsidian Exportroadbook.fly.dev
A reproducible benchmark for coding agents. The same agent on the same task can swing widely, so it reports min/max, task setup, Docker sandbox boundaries, and failure samples instead of only averages.
Benchmark · Docker Sandbox · Claude Code · Gemini CLI · Codex CLI · AiderVariance-aware · 10 Tasks · MITgithub.com/jackjin1997/AgentBench-Live
Personal AI knowledge base — RAG retrieval + multi-model chat (Claude/GPT/Gemini) + 12 AI personas + CXP context export protocol. Full-stack Clean Architecture.
Next.js 15 · Bun · tRPC · PostgreSQL · pgvector · LangChainRAG · 3 LLMs · 12 Personasjzxwiki.com
A curated resource index around OpenClaw and agentic intelligence: Skills, MCP servers, prompt/workflow patterns, and tool-protocol practice for personal AI assistants.
OpenClaw · MCP · Skills · Agent Tools · Resource IndexSkills · MCP · Agent Resourcesgithub.com/jackjin1997/ClawForge
I wanted an ocean, so I made a place to spend a little time. Sail, watch the sun and moon, listen to the sea, and draw a wave with your hand.
Three.js · WebGL · Water simulation · SunCalcInteractive water · Sun and moon · Sailinga-little-ocean.vercel.app
Framework fixes across checkpointing, streaming function_call chunks, Gemini / Vertex tools, metadata filters, and Bedrock streaming (langchain, langgraphjs, langchainjs, langchain-aws)
Main-branch fixes for agent compression, delegate deletion, gateway routing, CLI toolsets, and env sanitizer bugs
Core contributor: site mirror environments, task verifiers, and grading contracts, plus review and takeover of external contributions
SQLite session auto-compaction and Markdown parser fixes, a Kubernetes Helm chart, and Cosmos SDK proto cleanup
MDL relationship-traversal docs correction and a core multi-relationship calculated-column test
OSS Hunter skill plus Clean Code skill / registry stabilization
Coding-agent variance benchmark and OpenClaw skills / MCP / agentic resource index