Classify celestial bodies with MLOps best practices: a journey through local training, the cloud, serving, monitoring, labeling, and retraining.
We are the Swiss AI Center team. This guide is a practical path from notebook experiments to production, built for engineers and small teams. We selected tools that minimize friction for established workflows, with a focus on SMEs.
A step-by-step path from a notebook experiment to a reproducible, monitored, continuously-improved ML system. You will use the same tools as in production: DVC, Git, GitHub Actions, CML, Docker, BentoML, Fluent Bit, Evidently AI, and Label Studio.
Move from a notebook to clean, versioned Python scripts. Build a reproducible prepare-train-evaluate pipeline on your own machine, and use Git and DVC to track data, parameters, metrics, and plots as the model evolves.
Push the experiment to GitHub, store data in an S3 bucket with DVC, and set up a CI/CD pipeline that reproduces the run on every push. Use CML to publish parameter, metric, and plot comparisons directly in pull requests for team review.
Package the model with BentoML, expose a FastAPI endpoint, and containerize it with Docker. Push the image to a registry, wire builds and deployments into your CI/CD pipeline, and roll the model out on Kubernetes with self-hosted runners for specialized training pods.
Stream prediction logs to S3 with Fluent Bit and deploy an Evidently AI dashboard on Kubernetes to compare live data against the training reference. Open drift reports as GitHub issues so the team can decide whether to retrain or roll back by redeploying a previous container image tracked in Git and DVC.
Close the feedback loop with Label Studio. Let the model suggest labels through the FastAPI endpoint, review and correct edge cases, then merge the refined annotations and retrain with DVC. A better model makes the next round of labeling faster and more accurate.
Wrap up by deleting the cloud provider resources, GitHub repository, Personal Access Token, and local project directories. Cleaning up is part of shipping responsibly: it prevents surprise bills, removes exposed credentials, and leaves you with a reusable workflow instead of lingering infrastructure.
The guide is open source and designed to be followed hands-on. Pick a part, open your terminal, and ship a complete ML system.
Launch mission
