Aegis-Flow:Multi-Agent Cloud Security Orchestrator
Architected an Autonomous Security Orchestration System (AEGIS-FLOW) using LangGraph, implementing a multi-agent workflow for automated AWS infrastructure auditing and remediation with a human-in-the-loop (HITL) safety gate.
Core Impact
"Achieved real-time security auditing and automated remediation for AWS environments with a Human-in-the-Loop safety gate."
Architecture Breakdown
Engineered a multi-agent LLM pipeline (LangGraph + Claude Sonnet) that autonomously detects and remediates AWS security misconfigurations with a 100% remediation success rate and post-fix verification pass rate across all completed scans.
Reduced Mean Time to Detect (TTD) to 34s and Mean Time to Remediate (MTTR) to 83s by designing a stateful agent graph with parallel audit nodes, structured tool dispatch, and a recursion limit guard to prevent runaway inference loops.
Optimized LLM inference cost to ~$0.04 per full security scan (27K tokens) by engineering targeted system prompts, filtering tool arguments against live MCP schema to eliminate hallucinated parameters, and adding explicit tool-to-action mappings to prevent the model from inventing non-existent function names.
Implemented a human-in-the-loop safety gate within the agent loop that intercepts the model's remediation plan pre-execution, surfaces structured approval context to operators, and resumes the agent via stdin signal — achieving 100% operator-reviewed fix rate with zero unauthorized automated changes.
Systems Analysis Concluded