Project 01

FilingLens: Evidence-Grounded RAG Agent

A financial analysis agent for U.S. public companies, designed to help users understand a company faster. Users can ask natural-language questions about fundamentals, cash flow, valuation, risks, or business models, and receive evidence-backed analytical answers with citations and clear limitations.

Stack LangGraph, FastAPI, ChromaDB, DuckDB, vLLM, Pytest

Corpus 980 SEC filings, 88K+ chunks

Data 150K+ financial facts

System

Retrieval plus tools before generation

Maps user questions into target companies, analysis intent, required evidence, and tool calls before answer generation.
Combines ChromaDB vector search with DuckDB financial facts and price/event data.
Keeps numerical reasoning in deterministic Python tools with dependency tracing and provenance tracking.

Reliability

Answer checks for evidence and boundaries

Checks citation validity, numeric grounding, and missing-evidence disclosure at runtime.
Separates financial analysis from investment advice with explicit non-advisory boundaries.
Designed for local-LLM compatibility rather than depending on one hosted model provider.

Workflow

How a query moves through the agent

Understand

Extract company, intent, evidence needs, and tool plan.

Retrieve

Search filing chunks, facts, prices, and events.

Compute

Run deterministic financial metric tools outside free-form generation.

Validate

Check grounding, citations, missing evidence, and advisory boundaries.

Demo Preview

Interactive console direction

Analyst query

Analyze NVIDIA's cash flow quality, valuation boundary, and major risks.

Evidence-grounded response

The agent should return structured analysis with cited facts, tool-derived metrics, missing-evidence notes, and non-advisory boundaries.