Legal Document Agent & Fraud Detection Engine

Overview

This repository contains an Agentic AI system designed for the financial sector. It automates the extraction of critical data from legal documents (Subpoenas, IRS requests, Court Orders) and performs real-time fraud detection by identifying duplicate identities across internal databases.

Key Concepts

Generative AI vs. Agentic AI

• Generative AI: Focuses on content creation. In this project, it is used to summarize documents and extract entities (SSN, Names, Case Nos) using Large Language Models (LLMs).
• Agentic AI: Focuses on autonomous action. Unlike a standard chatbot, this system acts as an "agent" that can:
  1. Plan: Determine which extraction steps are needed based on document type.
  2. Tool Use: Query a SQL/Aurora database to find duplicates.
  3. Reason: Decide if a record matches a known fraud pattern without human intervention.

Technical Stack

• Model: Claude 3.7 Sonnet (via AWS Bedrock)
• Orchestration: Boto3 / LangChain
• Database: Mock SQL Database (simulating Aurora/RDS)
• Infrastructure: Python 3.11+, AWS SDK

Workflow

1. Ingestion: User uploads a PDF/Image of a legal document.
2. Extraction: The Bedrock Agent identifies SSN, Account Numbers, and Requestor info.
3. Verification: The Agent triggers a database lookup.
4. Logic: If the customer is tied to multiple account numbers under different aliases, a "High Risk" flag is generated.