Dispatches on shipping AI into production, for better or worse.
How we add agents and retrieval into production systems with zero downtime, one workflow at a time.
The evaluation checklist that decides whether a model is actually ready, beyond a good demo.
Why self-hosted inference is cheaper than it looks once you account for what you're actually running.
Agents look done at 80%. The last 20% — reliability, guardrails, and recovery — is the actual project.