Experiment with Azure Machine Learning: Review

Waiting two minutes for the sandbox to spin up is a big chunk of a one-hour module. Once inside, the experience is pure point-and-click. You're handed a pre-loaded dataset and a script, and the main task is just finding the right buttons in the Azure ML Studio UI to launch an AutoML job.

This feels less like learning and more like following a recipe without knowing why you're adding each ingredient. The AutoML section is the biggest offender. It finds the 'best' model, a VotingEnsemble, but doesn't really dig into what that is, why it was chosen over a simple logistic regression, or what any of the metrics mean beyond 'higher is better.' It's a black box inside a black box. For someone trying to actually understand the process, this is just frustrating.

The second half on using MLflow to track experiments in a notebook was slightly better. At least there's code. Seeing `mlflow.start_run()` and `mlflow.log_metric()` calls actually connect to a run history in the UI was a useful 'a-ha' moment. It makes the concept of experiment tracking tangible.

But the path ends right there. It doesn't connect to anything. It doesn't build on the model or deploy it. It’s a self-contained demo. I left knowing what buttons to press in Azure to run AutoML, but I don't feel any more prepared for a real project or a technical interview.