Using language models as components inside reliable systems, not as the system itself.
How constraining LLM responses to a strict schema makes AI steps testable, cheap to validate and safe to fall back from.