ML Engineering & MLOps

Pick a public tabular dataset and show, with a proper split, how much a model actually adds.

1. Frame the problem and the metric a business would care about.
2. Build a baseline with no ML at all, then a simple model, then a better one.
3. Use a pipeline so preprocessing is fitted on training data only.
4. Report the gain over the baseline with a confidence interval.

Turn the baseline project into a package.

1. Move code out of the notebook into modules with unit tests for data validation and features.
2. Drive runs from configuration, not edited code.
3. Track every run with MLflow, including the data version.
4. Prove it: a fresh clone and one command reproduce your best run.

Serve your reproducible model behind an API.

1. Wrap it in a small web API with input validation and a health endpoint.
2. Containerise it and run it locally with compose.
3. Deploy it to any free cloud tier.
4. Load-test it and report latency at p50 and p95.

Add monitoring to your model service.

1. Log inputs and predictions from the live service.
2. Simulate drift by shifting one feature's distribution in replayed traffic.
3. Detect it with a drift report and an alert threshold you can defend.
4. Write the retraining runbook: who acts, on what signal, and how you roll back.

Write the document a team would review before building.

1. State requirements: scale, latency, freshness and the business metric.
2. Design data and features, including how labels are collected.
3. Choose training and serving patterns, with the alternatives you rejected and why.
4. List failure modes and how each would be detected.