Brief #22 · Machine Learning
ML Model Deployment (MLOps)
Hard⏱ 20-28 hSoon
Turn a model that lives in a Jupyter notebook into a real service: a reproducible training pipeline, a model registry, an API, drift monitoring and a safe rollout.
Outlines are free · building with tracking needs Pro
MThe briefFrom the team
Our data scientist built a delivery-time prediction model that's great in her notebook. It has never shipped because nobody knows how to retrain it, serve it, or tell when it goes stale. Build the path from notebook to production so the next model ships in days, not never.
What you'll build
- 1Move the notebook into a training pipeline with fixed seeds and pinned data versions, so the same inputs always give the same model.
- 2Register each trained model with its metrics, data version and code commit, and promote models only if they beat the current one.
- 3Serve the model behind an API with input validation, and share the feature code between training and serving.
- 4Deploy a new model in shadow mode first: it predicts on live traffic but nobody sees its answers, and you compare it with the live model.
- 5Monitor input drift and prediction quality, and alert when the model needs retraining.
Step-by-step walkthrough
Walkthrough coming soon
This project's step-by-step walkthrough, with real outputs at every step, is still being written. You can read the brief and try it on your own meanwhile; tracking opens when the walkthrough is ready.
Done when
- Running the pipeline twice on the same data produces the same model metrics.
- Any prediction can be traced back to the model version, data version and commit that produced it.
- A shadow comparison report for a new model before it goes live.
- An injected data shift (for example, all distances doubled) fires the drift alert.
Stretch goals
- Automatic retraining when drift is detected, with a human approval step before promotion.
- Canary rollout: send 5% of traffic to the new model and roll back automatically if errors rise.