airCloset CTO Explains How to Shape Observability Data for AI Consumption

Ryan, CTO at airCloset, has outlined an observability architecture designed to make production monitoring data usable by AI systems. The core problem he identifies is that raw logs and metrics overwhelm AI context windows and cause hallucinations, mirroring issues he previously encountered with static code graphs. To address this, he divided monitoring into four distinct surfaces — application, infrastructure, CI, and LLM — each shaped into a format suited to the questions AI needs to answer. Application and infrastructure monitoring follow a standard OpenTelemetry setup routing data to Tempo, Loki, and Mimir via Grafana Cloud. CI and LLM observability layers involve more deliberate design choices, which he plans to detail further in a follow-up post covering PII handling and self-healing integrations.