What I did

• —Designed and built a multi-agent pipeline for automated financial claim verification on the FinDVer benchmark, matching cloud-only accuracy at ~50% of inference cost across 1,700 held-out claims.
• —Architected the routing system and run loop that decide whether each claim goes to a local 3B/7B model or a cloud API call.
• —Created two custom verification skills for the pipeline that further cut cloud calls by 20% while continuing to tie cloud-only accuracy.
• —Built a BM25 retriever that retrieves 75.2% of gold evidence, 5.4 points above the best published retrieval model on FinDVer.

Tools

Python · Ollama · DeepSeek API · NumPy · LaTeX