At Guiyang Big Data Development Administration Bureau, I led the architecture and development of Fangwen Legal AI, from initial design through implementation and iterative improvement. The government-facing application combines Agent Harness orchestration, modular RAG retrieval, citation-grounded reasoning, and RAGAS-based evaluation.
For Fangwen’s housing legal assistant, I implemented intent routing that combines rule-based fast paths with structured LLM classification. Its Golden Set regression tests maintained over 85% routing accuracy, while clarification and evidence-based refusal mechanisms helped reduce severe unsupported conclusions by approximately 70% in sampling-based reviews.
I also designed Fangwen’s housing regulations RAG system, using hybrid BM25 and vector retrieval with CrossEncoder reranking to ground answers in legal sources. On a Golden Set of more than 150 queries, retrieval achieved Hit Rate@5 above 85%; the system also supported privacy-preserving on-premises inference.
For Tsinghua University – Shuimu Zhifa, I contributed to a Legal AI platform and led development of a structured contract information extraction module. Grammar-constrained decoding and Pydantic validation achieved a JSON Schema validity rate above 99% in the evaluated test set. My background also includes data analysis at Guizhou Data Security Big Data Development Co., Ltd. and JD.com, and doctoral research at Southwestern University of Finance and Economics focused on AI and Law.
