Adding an LLM Narration Layer to a Self-Hosted Observability Stack

I almost made the classic AI architecture mistake. I could easily just dump raw Prometheus metrics...