Financial Document Analyzer

• 

  Upload as many financial docs as you want for OCR pipeline

• 

  Ask questions in natural language

• 

  Generate a loan application report

Inspiration

While preparing for an interview with a fintech startup, I wanted to deeply understand their space. I discovered that small business loan underwriting is still heavily manual, relying on tedious document review. I thought: what if I could use LLMs and retrieval-augmented generation (RAG) to accelerate that process? That question led me to build this prototype system to assist underwriters in evaluating SBA loan applications.

What it does

This tool allows a loan underwriter to upload financial documents like tax returns, bank statements, and payroll reports. It extracts text from the documents using OCR, stores it in a vector database, and enables intelligent question-answering and report generation using a Claude-powered RAG pipeline. The user can also generate a structured underwriter evaluation draft, editable before export.

How I built it

Backend:

• FastAPI
• OCR: Tesseract + pdf2image for PDF-to-image conversion
• LLM: Claude via the Anthropic API
• Vector DB: ChromaDB with default embedding model (all-MiniLM-L6-v2)

Frontend:

• Next.js

Architecture:

• PDF Uploads -> OCR w/ tesseract converts into text
• Text is stored as embeddings in ChromaDB

Challenges I ran into

• Understanding what a real underwriter evaluation report looks like
• Managing long input contexts and keeping the system fast and responsive

Accomplishments that I'm proud of

• works as intended!

What I learned

• OCR-RAG-LLM integrations!

What's next for Financial Document MCP-RAG System

• chunking up the text instead of writing it all completely
• custom embedding model instead of default embedding, to support semantic search instead of just normal embedding
• having the LLM be able to highlight places in the pdf

Built With

• chromadb
• claude
• fastapi
• next.js
• ocr
• python
• rag
• tesseract
• typescript