DevSwarm is my final year project: an agentic DevOps/SRE platform where a swarm of five agents (Deploy, Monitor, Security, Cost, Incident Response) is coordinated by a LangGraph supervisor.
Each agent uses real tools through MCP servers (GitHub, Docker, Kubernetes, Prometheus, security scanners).
When agents conflict, say Deploy vs Security, the orchestrator pulls similar past decisions from Qdrant‑backed RAG memory before deciding to proceed, block or escalate.
Stack is FastAPI, LangGraph, MCP SDK, Qdrant and a Next.js dashboard, and it extends the pattern I used in CodeSense.
I’m building it fully free of cost: local kind/minikube, self‑hosted Prometheus and Grafana, Trivy and OWASP ZAP, free‑tier LLMs (Groq/Gemini) and local BAAI/bge-large-en embeddings.
Cost is simulated with Kubecost, not live billing, and I plan to say that openly in the report and viva.
The 16‑week schedule includes three demo scenarios covering deployment, security conflict handling, and cost‑driven scaling.