MLOps · API Deployment
VisionServe: CIFAR-10 Inference API
A Dockerised FastAPI service demonstrating production-aware model serving with reproducible training, artifact tracking, and confidence-calibrated predictions.
FIG // 02.ACIFAR-10 classes served behind a single inference endpoint
CLI command spans preprocessing, training, metrics, and artifact export
Container ships CPU-friendly PyTorch wheels for easy portability
Problem
Most image-classification tutorials stop at a trained model in a notebook. Shipping that model as something another service can actually call — with reproducible training, versioned artifacts, and a container that runs the same way on a laptop as it does in CI — is a different problem, and the one this project was built to demonstrate end-to-end.
Approach
VisionServe wraps a CIFAR-10 classifier in a FastAPI service, with the surrounding tooling treated as first-class rather than an afterthought:
- A single CLI handles the full lifecycle: preprocessing, training, metrics export, and artifact versioning, so a run is reproducible from one command instead of a chain of ad-hoc scripts.
- Predictions are confidence-calibrated rather than raw softmax outputs, so downstream consumers get a more honest signal about how much to trust a given classification.
- A confusion matrix and metrics summary are generated automatically per run, giving a fast QA check before anything ships.
Architecture
The service is packaged as a Docker image built on CPU-friendly PyTorch wheels — deliberately avoiding a CUDA base image so the container stays portable across environments that don’t have a GPU available, which is the common case for lightweight inference services.
Result
The result is a template for production-aware model serving: train reproducibly, version the artifact, containerize it, and expose it over a documented API — the same shape of work needed to take any classifier from notebook to service.