FastAPI

FastAPI is where my Python services live when they face consumers. The system that exercises the skill hardest is a production retrieval platform serving one capability through two protocols: a single async application mounts the REST router and the MCP endpoints behind a composed lifespan, so both surfaces share the same connections, models, and retrieval core while their authentication is controlled independently — a deliberate design decision, because an internal agent runtime and an external REST consumer earn trust differently, and one switch for both would force the stricter posture on the surface that does not need it.

The rest is the discipline that makes an async service boring to operate: Pydantic models as the single contract from request validation through structured LLM output to response serialization, dependency-injected auth, an OpenAPI schema that is generated rather than hand-maintained, and lifespan-managed resources so startup and shutdown are explicit. In production the application runs on Gunicorn with a Uvicorn worker inside a container, deploys to Kubernetes as a Helm release with autoscaling and health probes, and sits behind a CI pipeline where linting and the test suite gate every image. The platform this describes is in the Related Projects panel above.