How to Structure FastAPI to Gracefully Handle AI API Failures

A developer shared lessons learned after a live demo exposed how fragile FastAPI services can become when an external AI API misbehaves, such as timing out or returning malformed responses. The core recommendation is to keep router functions thin, delegating all provider-call logic to a dedicated service layer rather than scattering error handling across endpoints. A typed exception hierarchy — covering timeouts, rate limits, and malformed output — allows the app to return meaningful HTTP status codes like 504, 429, or 502 instead of generic 500 errors. The author also stresses validating AI responses with Pydantic schemas, not just incoming requests, since language models can return near-valid or structurally unexpected JSON. These patterns were distilled from multiple real-world projects and packaged into a reusable FastAPI starter template.