johnsonhk88/AI-Bank-Statement-Document-Automation-By-LLM-And-Personal-Finanical-Analysis-Prediction

What it solves

This project automates the extraction, structuring, and analysis of financial data from bank statement PDFs. It eliminates the manual effort of digitizing bank statements and provides a way to perform financial trend analysis and natural language querying over personal financial records while maintaining privacy through PII redaction.

How it works

The system uses a multi-stage pipeline to process documents:

1. Extraction: It combines YOLO for layout detection, OCR, and LLMs to extract tables and text from PDFs.
2. Privacy: A secure RAG pipeline redacts Personally Identifiable Information (PII) before embedding data into vector databases like Qdrant or Chroma.
3. Orchestration: It supports three different agent frameworks (CrewAI, Deep Agents, and Hermes) to handle tasks such as parsing, redaction, and financial analysis.
4. Inference: It is designed for "Local LLM First" usage via LiteLLM, supporting tools like LM Studio and Ollama to keep data local.
5. Full-Stack Integration: A FastAPI backend and React frontend manage asynchronous document processing via Celery and Redis, storing structured data in PostgreSQL.

Who it’s for

• Individuals wanting to automate their personal financial analysis and predictions.
• Developers looking for a reference implementation of a multi-harness agent system (comparing CrewAI, Deep Agents, and Hermes).
• Users who prioritize data privacy and want to run financial LLM workflows on local hardware.

Highlights

• Multi-Framework Support: Implements the same workflow across CrewAI, Deep Agents, and Hermes for comparison.
• Privacy-Centric: Built-in PII redaction before data enters the vector database.
• Local LLM Optimized: Specifically tuned for local models (9B+ parameters) via LiteLLM.
• Complete Infrastructure: Includes a full-stack API, GPU acceleration support, and MLflow integration for tracing agent workflows.