SentinalSRE — 100% Local-First Autonomous SRE Engine (LangGraph + MCP) Hi everyone! I just finished a project focused on automating the on-call lifecycle without sending a single byte of production log data to the cloud. The Core Loop: I built a cyclic, multi-agent system using LangGraph that detects crashes via FastMCP, performs Root Cause Analysis (RCA) locally, and generates a verified fix. Technical Highlights: 100% Local Inference: Runs on Ollama (Llama 3) via local NVIDIA GPU—zero API costs and total data privacy. Cyclic Self-Correction: Includes a Syntax Checker node that detects "Hard Fails" (like blanking functions) and triggers an automatic re-diagnosis loop. MCP Orchestration: Integrated custom FastMCP for log ingestion and the official GitHub MCP for automated PR creation. HITL Safety: Implemented an interrupt_before gate for operator approval before any code hits the repo. Performance Metric: Reduced Mean Time to Recovery (MTTR) from a manual ~45 minutes to 288 seconds on consumer hardware GitHub: https://github.com/NANDAGOPALNG/Sentinal-SRE LangSmith Trace: https://smith.langchain.com/o/a874c146-dd28-4d8a-be47-4bdcfb74000e/projects/p/4e96ecd0-b4fc-4a62-8667-ff90b18166e8 Would love to hear your thoughts on how to scale this for distributed microservices or handle more complex edge cases in the self-correction loop!