Ricky-7-Yan/intelligent-audit-system

What it solves

AuditPilot addresses the inefficiencies and inconsistencies of enterprise audit delivery. It replaces manual, experience-dependent processes—such as evidence collection, control testing, and risk mapping—with a governed AI Agent workflow that ensures every conclusion is traceable, verifiable, and grounded in evidence.

How it works

The system uses a bounded Agent runtime that follows a Plan-Execute-Reflect loop to handle audit tasks. It integrates an Agentic RAG system (using hybrid TF-IDF and keyword search) to retrieve evidence and an Evidence Graph to track the lineage between tasks and artifacts. High-risk or low-confidence outputs are automatically routed for human review. The entire process is gated by a multi-layer Evaluation Harness that tests task results, tool calls, and security permissions before allowing a release.

Who it’s for

It is designed for enterprise auditors and compliance teams who need to perform structured audits (such as SOX, ISO27001, or COBIT) while maintaining a strict audit trail and reducing reliance on individual auditor experience.

Highlights

• Built-in Audit Templates: Includes pre-configured templates for SOX ITGC, ERP permissions, data security, production change management, backup recovery, and third-party service audits.
• Agentic RAG: Features multi-path recall, conflict evidence identification, and missing evidence alerts to ensure grounded conclusions.
• Evaluation Harness: A 9-layer evaluation system that outputs calibration scores and confidence lower bounds to block unstable releases.
• Governed Improvement: Failures are captured as experience candidates and only reused after passing regression tests and human approval.
• MCP-style Tools: Implements a tool system with RBAC validation, tenant isolation, and circuit breakers.