AuditCopilot
An LLM-driven audit support system for fraud and anomaly detection in double-entry bookkeeping.
AuditCopilot is a research project on using large language models for audit support in double-entry bookkeeping. The system combines learned anomaly scoring, rule-based audit logic, and natural language explanations to help auditors inspect suspicious transactions more effectively.
I worked on the system as a Research Assistant at the Interactive Machine Learning Lab, DFKI, from July 2024 to March 2026. The project includes a FastAPI deployment and was validated through an end-to-end user study with domain experts.
Highlights
- Combined learned anomaly detection with rule-based audit checks.
- Generated natural language explanations for model decisions.
- Reduced false positives substantially on synthetic financial ledgers.
- Improved F1 score from 0.50 to 0.94 compared with traditional audit rules.
Topics: large language models, anomaly detection, human-in-the-loop systems, explainable AI, financial data.