AI diagnostic agent
DevOps Agent
An agent that watches a Kubernetes cluster continuously, alerts only on a real anomaly and explains the failure by citing company documentation.
Two uses: answering a question by citing the indexed documentation, or diagnosing a real problem by chaining its own tools (documentation search, kubectl, reading manifests). The model that writes is interchangeable, local or through an API; documentation search stays local either way.
What I built
- Captures cluster events through watch and filters out noise: thresholds and deduplication.
- On a real anomaly, writes a report: the cause and the fix, citing indexed runbooks, architecture decisions and post mortems.
- Hybrid search over the documentation (bge-m3 embeddings and BM25), with local models through Ollama or the Claude API.
Highlights
- 5,000 healthy events, 0 alerts.
- Strictly read-only, enforced in code: it suggests, it never applies.