FinSight was built for Inter IIT Tech Meet 13.0. The problem sounded simple but it was actually brutal. Given long financial documents, build a system that can answer questions with correct citations and not hallucinate.
We built an agentic RAG pipeline that could ingest documents, chunk them smartly, build an index, retrieve the right context, and then generate grounded answers. I worked on the pipeline design and evaluation loops. We used LangGraph to orchestrate the steps, Pathway for the data flow, and GPT 4 for reasoning. We also added LLM based evaluators to catch bad answers early and tune retrieval settings.
We ended up getting 1st place. The best part was seeing the system answer tricky finance questions and actually point to the right part of the document.