AI Agent platform lead in Shanghai. I build infrastructure for running AI agents in production, and the tooling to run fleets of coding agents in parallel.
Most of my work lives in private repositories, which is why the contribution graph below is busy and the public repo list is not.
- North Agent Cloud: a platform for building, deploying, and running AI agents. Scheduling, sandbox isolation, egress control, credential custody, elastic scaling, and end-to-end tracing. In production across several deployments.
- North Sandbox: a multi-tenant sandbox service compatible with the E2B protocol. Per-template warm pools bring sandbox acquisition down to milliseconds, with JuiceFS over object storage as the sandbox file system.
- North Firewall (built solo): transparent egress interception with iptables, HTTPS inspection through CA injection and MITM, and Vault-backed credential isolation, so agents only ever see placeholders.
- Agent Team (built solo, 1,400+ commits): a control plane that runs 5 to 10 Claude Code / Codex agents in parallel across machines, each in its own git worktree and Kubernetes namespace. It gives every agent a status board, an activity timeline, and message delivery with receipts; it patrols them every minute; and its custom Stop hooks catch specific failure modes, for example an agent that says it will keep going and then ends its turn. Nothing is called done until a second agent has tried to prove it broken.
Public bits: agent examples and fixes in nexau-cookbook.
AI Scientist at Devz AI (US): enterprise agents, a general-purpose web agent, and fine-tuned domain LLMs. MA in Computational Social Science, University of Chicago.