Angelo - Sell Project

Angelo — self-hosted Kubernetes SRE control plane When something breaks in Kubernetes, most of the outage is spent working out what happened. Angelo watches your clusters, turns firing Prometheus alerts into incidents automatically, gathers the evidence, ranks the likely root causes, and proposes a fix from a fixed catalog of safe actions. Nothing touches the cluster unless its policy allows it. Features Incident loop: firing alerts open incidents automatically, then evidence is collected, root causes are ranked, a fix is proposed, and a report is generated as Markdown, HTML or JSON. Policy-gated remediation: 6 catalog actions (rollout restart, scale, delete pod, cordon, drain, Argo CD sync) run in one of 3 modes: dry-run (the default), human approval, or autonomous. A model can never write its own commands. Root-cause ranking: rule-based by default, with optional LLM input from OpenAI, Claude, Gemini or a local Ollama model, and memory of similar past incidents (Qdrant). Console (Next.js): operations dashboard, live service topology graph, infrastructure host map, Prometheus metrics explorer, monitors and alerts, incidents, evidence, timeline and actions, with dark and light themes. Integrations: Prometheus, Loki, Elasticsearch, OpenTelemetry, SigNoz, Pixie, Trivy, Falco, Argo CD, Redis, RabbitMQ and Kafka. Security: JWT login with role-based access (viewer, operator, approver, admin), HMAC-signed actions, a namespace allowlist and an audit trail. Startup refuses default secrets in production. Tech stack: Python 3.12, FastAPI, LangGraph, SQLAlchemy/Alembic, PostgreSQL, Redis, Qdrant, Next.js 15, TypeScript, Docker, Helm, Kustomize. Deployment: Docker Compose for local use, a script that builds a real 3-node kubeadm cluster, and a Helm chart for EKS, GKE, AKS or self-hosted clusters. CI runs lint, tests, Helm/Kustomize validation and a Trivy security scan. Status: working MVP, demonstrated on a real 3-node cluster (see screenshots). No paying users. Remaining work a buyer should know about: Root-cause ranking is mostly keyword rules; it needs a minimum-confidence gate before autonomous actions. Node actions (cordon/drain) aren't covered by the namespace allowlist. Self-approval of actions isn't blocked. The flow-ingest token needs to be required in production. The repo includes a second, experimental React console that isn't wired into the build. Licensing: the code was published on GitHub under Apache-2.0 for about three months (0 stars, 0 forks) and is now private. The sale transfers copyright ownership, the repo, the name and the docs. What's included: full source code, Helm chart, Kustomize overlays, Docker Compose setup, CI workflows, the cluster provisioning scripts, docs and screenshots.