Operations | Monitoring | ITSM | DevOps | Cloud

You shipped your AI app. But do you have all the answers? Do you actually know which model ran, how many tokens it consumed, or why it stopped? This is what LLM observability gives you, and most AI engineers are skipping it entirely. I built an SOS detection app and used OpenTelemetry to get full visibility into every single call. Token usage, model version, finish reason, and cost per call all in one place, standardised across any provider. Check out the OpenTelemetry GenAI docs in the link below; there is a lot more you can track than you think.

Your logs showed 500 errors. The traces showed the dependency graph. Neither showed the actual bug, a DEL control character getting appended to the query string. This is how I found it. In this video I walk through Speedscale BYOC (bring your own cloud): capture real production traffic, store it in your own Elasticsearch cluster inside your VPC, pull it down locally with a single script, and reproduce the exact bug using proxymock. The data never leaves your environment.

Lightrun's Gidi Freud demonstrates how your AI coding agent can now investigate and fix production errors, autonomously. Watch how Cursor, guided by Lightrun's Error Remediation skill, picks up a Sentry error, instruments the live service with a runtime snapshot, captures real evidence, and opens a validated PR for approval.