FinSight
LangGraph agent turning SEC filings and news into cited theses; NVIDIA GTC 2026 hackathon finalist
Project Overview
FinSight is a 7-node LangGraph pipeline: given a ticker it pulls the latest 10-K or 10-Q from SEC EDGAR, ingests Yahoo Finance and Google News RSS, and runs a ReAct loop on NVIDIA Nemotron Super 49B to output bull case, bear case, and a confidence-weighted BUY/SELL/HOLD recommendation. The agent stops when it has enough evidence; claims cite filing sections to reduce hallucination. Built for the NVIDIA GTC 2026 Agents for Impact Hackathon, where it was a finalist.
The Problem
Reading a 10-K filing takes a professional analyst 4-6 hours. Most retail investors never do it. FinSight automates the pipeline end-to-end, from raw filing to cited thesis, in under 3 minutes with only a ticker symbol.
Streamlit app
Enter a ticker to run the full agent pipeline.
7-Node Agentic Pipeline
The pipeline runs: Retriever (SEC EDGAR API + XBRL) → Scanner (headline extraction) → Planner (company-specific investigation plan) → Extractor (structured metrics, XBRL-first) → News (Yahoo Finance + Google News RSS) → Reactor (ReAct loop, max 5 iterations, self-terminating) → Thesis (BUY/SELL/HOLD with citations). LangGraph conditional edges make the reactor loop genuinely agentic: the model decides when to stop.
Pipeline architecture
End-to-end flow from filing retrieval to cited BUY/SELL/HOLD output.
Analysis in progress
Structured extraction and reasoning steps.
3-Way Tension Detection
The most powerful feature is the check_exec_vs_news tool, which detects when executive claims in the filing, current news headlines, and actual reported numbers all tell different stories. This is the primary signal professional short-sellers look for. The tool classifies 6 tension patterns from CRITICAL TENSION to ALL THREE ALIGNED and surfaces them in the thesis.
Investment thesis output
Bull and bear cases with citations.
Hackathon Result
Finalist at the NVIDIA GTC 2026 Agents for Impact Hackathon. Demonstrates NVIDIA Nemotron Super 49B's structured multi-step reasoning on a real financial analysis workflow. Live demo: nemo-finsight.streamlit.app.