Containerize a Python inference service and set up GPU-aware deployment on AWS

Right now our ML engineer deploys the model server by SSHing into a GPU instance and running tmux. It works until it does not, and she would rather be training models. The service itself is a FastAPI app in Python that serves an image classification model for quality control in food production lines. Scope is tightly defined: a production-grade Dockerfile (multi-stage, CUDA base image, dependency pinning), a docker-compose setup for local development that our team can run on CPU, deployment onto our existing AWS account with autoscaling between one and three GPU instances based on queue depth, and health checks that actually reflect model readiness rather than just process liveness. Small, well-bounded project - we know exactly what we want and will not expand scope mid-flight. Handover is a one-hour walkthrough call plus written deployment notes.