Describe the lifecycle of an LLM application in production.

Design (use case, success criteria, risks) → build (prompts, retrieval, tools, orchestration) → evaluate offline (golden sets, safety tests) → deploy through CI/CD with eval gates and staged rollout → operate (observability, cost tracking, incident response) → monitor quality and drift online → collect feedback and failures → improve (prompts, data, retrieval, models, fine-tuning) → re-evaluate and redeploy. It's a continuous loop, and external changes (provider model updates, new documents, new user behaviour) restart it even when your code hasn't changed.