AI Agent · Software Systems · Backend Engineering

🎓 2027 Graduate Applicant — Open to graduate research opportunities

Information Security undergraduate at Hainan University.

Research interests: AI Agents · Coding Agents · LLM Systems · Intelligent Software Engineering · Agent Security

Currently looking for potential graduate advisors and research opportunities for Fall 2027.
目前正在联系 2027 级推免导师，欢迎 AI Agent、软件系统、智能软件工程及相关交叉方向的老师交流。

I build AI agent systems, backend services, and evaluation infrastructure, with a focus on controlled execution, tool calling, verification, and reliable LLM application engineering.

Coding Agent Runtime & Evaluation Harness for controlled, repository-level software repair.

• Built a bounded Agent Runtime with multi-turn tool calling, state management, execution budgets, timeout/cancellation, and typed termination.
• Added isolated Git workspaces, constrained repository tools, independent verification, execution traces, and evaluation artifacts.
• Evaluated a frozen HAL SWE-bench Verified Mini fixed-50 configuration with the official evaluator: 35/50 officially resolved, 50/50 valid executions, with zero Harness-invalid or grading-invalid runs.
• Built a local read-only Runtime & Evaluation Console for inspecting real execution traces, verification results, artifacts, and benchmark analysis.

View AgentPatchCheck →

A follow-up empirical audit derived from APC's frozen fixed-50 runs, examining whether typed runtime termination can substitute for evaluator-derived resolution.

The mechanical proxy finished → resolved / non-finished → unresolved disagreed with the official evaluator on 12/45 binary-outcome runs (≈26.7%), with errors in both directions. The analysis supports separating runtime telemetry from evaluator verdicts while retaining termination reasons as diagnostic signals.

• LangChain4j — #6116: fix(agentic): finalize failed executions after async sibling completion
• Spring AI Alibaba — #4911: fix(graph): notify lifecycle listener on streaming errors
• Spring AI Alibaba — #4925: fix(graph): preserve streaming node id in error callbacks
• MaxKB — #5280: fix: sanitize exported chat HTML content

Backend: Java · Spring Boot · Spring AI · Node.js

AI / Agent: Agent Runtime · Tool Calling · Agent Evaluation · RAG · MCP · Text2SQL

Data & Infrastructure: MySQL · PostgreSQL · Redis · Docker · Git

Frontend / Tooling: TypeScript · React · Vite