Designed dual-parallel architecture: Claim pipeline (extraction → retrieval → judgment) and 11 Checkers execute fully concurrently, with results merged and deduplicated
Developed 16 modular data sources (Playwright browser automation + REST/JSONP API), covering encyclopedia, terminology, gazetteer, legislation, ancient texts, academic papers, and other authoritative sources
Implemented Source feature awareness: LLM automatically generates differentiated query terms based on each source's scope, skipping Chinese databases for pure foreign-language queries
Integrated PKU Law MCP service (semantic search + keyword search + precise statute lookup) for authoritative legal citation verification
Conservative judgment strategy: unfound ≠ error, achieving 90.6% precision, 1.5% FPR, 30% recall on 1483-sample test set
Unified caching mechanism (SQLite + TTL), concurrency control, error tracking, with FastAPI interface and Gradio demo