← rishabh kumar

Blog

How production AI systems fail, usually quietly, and the gates that catch it. Drawn from systems I actually run.

Jun 2026
A wrong character in a URL killed our LLM tracing for 4 weeks
The health check was green the whole time. A story about why existence checks aren't arrival checks.
5 minOpen source
Ship-Safe Harness: test, eval, and deploy gates for AI codebases
A free checklist and generator skills that add the missing gates around your codebase, taken from the harness I run in production.
repo ↗