SENTINEL - Autonomous Site Reliability Engineer
An agentic-SRE system that investigates production alerts on its own with real tool calls, proposes exact remediations grounded in runbook RAG, and executes only after a human approves.
- Express
- node:sqlite
- BM25 + tf-idf RAG
- Gemini/Groq tool calling
- React
SYSTEM BUILD
What went into it
- 01
Runs a bounded agent loop with JSON-schema tools — fetch logs, pull metrics, search runbooks — that must converge on a schema-validated proposal or the incident is honestly marked failed.
- 02
Passes every remediation through a static safety analyzer with 12 hard-block rules before it reaches the human approval gate; execution is simulate-by-default with an audited strict mode.
- 03
Retrieves from a heading-weighted BM25 + tf-idf knowledge base of runbooks and post-mortems, citing documents on every proposal with a confidence score and rollback plan.
- 04
Pairs the agent with a simulated fleet of 8 services and 8 fault scenarios with ground truth, so the entire investigate → approve → execute loop demos safely offline.
MEASURED OUTPUT
What changed
Kept a human unconditionally in the loop — the agent's only write is a proposal, and safety-blocked scripts never reach the approval queue.
Backed by 38 Vitest/Supertest tests and 6 Playwright E2E specs across the API and console.