intelligent-audit-system

Why we included this project

Most agent frameworks stop at a chat demo, but AuditPilot targets a harder use case: turning audit fieldwork into a defensible, repeatable deliverable. Evidence retrieval is grounded in cited sources, tool calls are gated by RBAC and tenant isolation before they run, and each run is recorded as a traceable episode with failure attribution and checkpoints for human review. Its evaluation harness scores task results, execution traces, and evidence citations in layers, and a failed key assertion blocks release outright, so improvements only get reused after passing regression review and human approval. For teams building agents in other regulated fields like compliance or risk, it's a practical reference for keeping human oversight and explainability in the loop.