Agent runtime & memory engineer · MSc AI, researching agent memory systems · Shenzhen, China
Collaborator on bytedance/deer-flow (82k★) · 100+ merged upstream PRs · Open to AI agent / LLM infra internships
I work on the layer under the agent — the harness, the sandbox, the memory, the parts that break in production.
In DeerFlow, a LangGraph-based long-horizon agent harness, I've designed and shipped:
- Long-running agents — a durable task runtime that lets tool work outlive the agent turn, and managed subagents with bounded, recoverable fan-out. (design)
- Sandbox & agent security — controlled network egress with human approval, plus prompt-injection and secret-containment hardening across the harness. (egress)
- Memory & context — long-term memory hygiene and eviction, and context compaction for the lead agent. (RFC)
- Streaming performance — found and removed the main-thread bottlenecks behind long streamed runs. (analysis)
Also contributing to infiniflow/ragflow · Canner/WrenAI · Effect-TS/effect · mem0ai/mem0
- TechSpar
— memory-driven adaptive interview coach: one candidate profile drives training, resume/JD prep, a real-time copilot and review in a self-improving loop (LangGraph + 3-tier memory). Live
- arxiv-mcp — MCP server for searching and reading arXiv papers (PyPI)
- EvoGraph — minimalist graph-based long-term conversational memory
- MemRepo — incremental, hierarchical codebase memory for AI coding assistants
- Patent — Microkernel-architecture context manager for large language models
- 2nd Prize — ASC25 Student Supercomputer Challenge
- Kaggle Bronze — Jigsaw Agile Community Rules Classification (SFT track)
- 3rd Prize — China Postgraduate Mathematical Contest in Modeling (2025)