A prototype that demos well is not a product. The gap is mostly boring engineering: validation, observability, failure handling, and cost control. This checklist covers the things teams skip and regret later.
Quality
- An eval set with at least 30 cases and a recorded baseline score.
- Prompts versioned in code and reviewed like code, not edited in a dashboard.
- Input validation and output schema validation with retries on failure.
- A documented set of known failure modes and how the system handles each.
Reliability
- Timeouts and retries with backoff for every external call.
- Fallback behavior when the model or provider is unavailable.
- Rate limiting per user and per IP.
- Graceful degradation: the app remains useful when AI is down.
Performance and cost
- Caching for repeated or near-repeated requests.
- Cost tracking per request, per feature, and per user.
- Latency measurement split into time-to-first-token and total.
- Load testing at twice expected traffic.
Safety and observability
- Input and output guardrails for injection, toxicity, and PII.
- Logging of inputs and outputs with PII redaction.
- Human review sampling on a regular cadence.
- Alerts on error rate, latency spikes, and cost spikes.
Operations
- A rollback plan for prompt and model changes.
- Staging environment that mirrors production.
- Documentation of how to rotate keys and update models.
- On-call runbook for the most likely failures.
Ship when the eval score is stable and the failure modes are documented, not when the demo looks good. The demo is the easy part.