The AI Agents that investigate your incidents - on your infrastructure.

Running three backends (Ollama local, Anthropic, OpenAI) with different latency and cost profiles required production patterns - not just retry logic. Circuit breakers open per model after three consecutive failures and retry in half-open state after 30s. A self-healing daemon polls Ollama every 20s and auto-restarts failed models. Priority queue ensures CRITICAL alerts skip ahead of INFO. Sliding-window rate limiter caps spend at $50/day by default, configurable via env var. Kill switch via POST /pause for incident response.

LLMs are overconfident. We learned this after shipping the first version with a pure LLM self-rating and seeing 95% claimed confidence on RCAs that were wrong. The current formula blends the LLM's rating (40%) with actual signal coverage - how many of the six RCA steps returned useful data (60%) - and then applies a bonus/penalty based on whether the service has a history of accurate RCAs. This is the single most important reliability improvement we shipped.

We didn't want a RAG-over-logs approach - that's just search. Episodic memory stores structured episodes and matches new incidents against past ones using weighted similarity. When two or more similar episodes exist, we synthesise a strategy and inject it into the LLM prompt before the new investigation runs. This is the compounding-learning loop.