Projects
Selected work in agents, reinforcement learning, and model systems.
Code Agent · GRPO · LoRA · vLLM
Repository-level Code Agent Post-training with GRPO
Built a repository-level retrieval, rollout, automated evaluation, and training pipeline on top of Qwen2.5-Coder-7B-Instruct. On SWE-bench Verified, the system improves pass@1 by 3.4% and achieves a patch application rate above 95%.
pass@1 +3.4%Patch application >95%
GitHub Repository ↗LangGraph · RAG · Tool Calling · FastAPI
E-commerce Customer Support and After-sales Agent
A multi-tool agent for order inquiries, refunds, logistics, and after-sales coordination, built with LangGraph workflows, RAG, PostgreSQL, and service APIs.
Task completion +14.8 ppTokens -27.6%Latency -21.3%
GitHub Repository ↗