Hi, I'm Libin.

I'm Libin Cao, an engineer in Shenzhen who builds AI products. Not demos — the version that actually gets deployed, holds up under real users, and doesn't quietly cost a fortune per request.

I build agent systems end to end: execution and orchestration, MCP servers and tool integrations, and multi-channel gateways so one agent can be reached wherever your users already are. I know the agent platforms deeply enough to tell you what will and won't port between them — I took a document-processing agent built on hosted skills and code-execution containers, tested it across Anthropic, Cloudflare AI Gateway and OpenRouter, and wrote down exactly where each one stops working.

I also treat token cost as an engineering problem rather than a bill. On that same agent, prompt caching plus context management took a single task from 2.7M input tokens (about $8) down to roughly $0.15 — a 50× difference, with no change to what the user received. And having come up through infrastructure, I can put the result on Kubernetes with real observability instead of leaving it on one box.

Right now that work is at Kuse, an AI workspace, and Junior, the AI-employee product built on top of it, where I own backend and infrastructure: the move from a single box to a GKE cluster, a unified multi-channel gateway, agent execution orchestration, and the observability that keeps all of it debuggable. Before that, at BUD, I designed and maintained a real-time multiplayer engine on my own (Agones with state and frame sync, 50+ players on screen, millions of daily users), and owned the backend platform and the LLM engineering behind every AI project the company shipped.

The work I like best is forward-deployed: sitting with your actual problem, working out whether AI is even the right answer, then building and shipping the thing that solves it. Based in Shenzhen, happy to work remote. Email me and describe the problem — if I'm not the right person for it, I'll tell you that instead.