Connecting the LLM Was the Easy Part

We wired an agent into every production alert: OpenTelemetry into SigNoz, tools over MCP, a model we host ourselves — no SaaS, no vendor API. It rarely found a root cause, and the alerts were to blame, not the model: a rotation paging monthly, restarts that were counter resets, gaps between rules. Now it grades every alert — cutting what stays green, drafting what incidents proved missing, flagging what nobody can explain. You will leave knowing which rules to delete and which you lack.