AuditAI:Agentic RAG Compliance Engine
An autonomous compliance engine using LangGraph and Corrective RAG (CRAG) to audit organizational policies against the NIST CSF 2.0 framework.
Core Impact
"Reduced compliance auditing latency by 60% while maintaining 100% faithfulness to NIST CSF 2.0 via a Self-Correcting Graph Architecture."
Architecture Breakdown
Built a Corrective RAG (CRAG) agent using LangGraph with a self-healing 4-node pipeline that automatically rewrites failed queries up to 3x, achieving 100% Faithfulness and 90% Context Recall across 10 RAGAS-evaluated NIST compliance Q&A pairs.
Engineered a real-time token-streaming API with FastAPI (NDJSON/SSE) over a stateful LangGraph graph, filtering intermediate LLM grader tokens and implementing smart citation suppression — evaluated across 4 RAGAS metrics with all metrics passing ≥0.7 threshold.
Designed a semantic query router with conversation-history context (last 3 turns) that classifies intent before graph invocation, reducing unnecessary agentic calls for non-compliance queries while maintaining 76.4% Answer Relevancy on domain-specific questions.
Delivered a full RAG evaluation pipeline across 50 ground-truth NIST Q&A pairs, scoring 78.4% Context Precision, 76.4% Answer Relevancy, 90% Context Recall, and 100% Faithfulness using Gemini 2.5 Flash as judge LLM.
Systems Analysis Concluded