Backend engineer · Reliability, distributed systems & runtime infrastructure
Contributing to bytedance/deer-flow across runtime reliability, persistence, sandbox infrastructure, CI, and security.
I work on the failure paths — ownership, cancellation, durable state, fencing, and the cases where a timeout doesn't mean the work actually stopped.
In DeerFlow, an open-source agent harness, I've worked on:
- Durable runtime correctness — cancellation-safe MCP task claims with bounded compensation, lease-token fencing, and stale-generation rejection. (#4966)
- Persistence & concurrency — safe durable cleanup for thread deletion while fencing concurrent event writers. (#5535)
- Sandbox reliability — command deadlines, no-replay semantics, and session fencing for ambiguous AIO sandbox outcomes. (#5634)
- CI infrastructure — duration-aware parallel sharding for a 13k+ backend test suite with exactly-once coverage. (#5137)
- Filesystem security — Windows credential hardening against ACL inheritance, reparse points, hard links, and path races. (#5141)
- Deployment infrastructure — Gateway/provisioner state-root consistency for shared PVC-backed sandbox storage. (#5625)
My current focus is on failure semantics and ownership boundaries in asynchronous and distributed systems.
- tiny-store — distributed commerce backend exploring inventory reservation, Redis atomic operations, idempotency, asynchronous persistence, compensation, and reconciliation.
- yahoo-stock-mcp — MCP server for structured financial-market data access, focused on API boundaries, reliability, and caching.
- MAGI-system — experimental multi-agent system for agent orchestration and LLM application infrastructure.
📖 Blog