Built AI Forensic Accounting Software with My Dad — CaseTrail Automates Financial Fraud Detection

CaseTrail is AI-powered forensic accounting software built by a father-son team. The tool ingests bank statements (CSV/PDF) and uses LLMs to classify transactions and flag suspicious activity. The blog post on case-trail.com walks through the technical architecture and real-world deployment.

Key Technical Details

• Data Ingestion: Parses CSV and PDF bank statements via standard libraries (e.g., pandas, pdfplumber).
• LLM Integration: Uses GPT-4 or similar to classify transactions into categories (e.g., payroll, vendor payments) and detect anomalies based on patterns.
• Anomaly Detection: Rules engine + LLM reasoning to flag potential fraud (e.g., duplicate payments, unusual vendor activity).
• Stack: Python backend (FastAPI/Flask), React frontend. The code is not open-sourced but the blog explains the core logic.

The article highlights practical challenges: handling PDF parsing inconsistencies, LLM hallucination on ambiguous transactions, and the need for human-in-the-loop reviews. They also share performance metrics — roughly 80% automation on standard cases, with manual review for edge cases.

For developers building similar vertical AI tools, the post offers a candid look at integrating LLMs with domain-specific workflows.