How to keep FastAPI request handlers thin by offloading heavy AI inference work to Celery workers behind a Redis broker.