



Day 5 is long, so here is the shape of it before we start.
The ML problem is already solved. The notebook in experiments/ tried a bunch of models on the spam dataset and found one that works. Nothing today improves that model. The whole day is about turning that one working solution into a project that runs itself.
Each step exists because the step before it left a problem behind:
Components → a notebook cannot run on its own, so the code becomes five small .py files.
DVC pipeline → running five files by hand is easy to get wrong, so one command runs them in order.
params.yaml → tuning a number still means editing code, so the numbers move into one settings file.
dvclive → after a few runs you forget which settings gave which score, so each run records itself.
AWS S3 → the data still sits on one laptop, so it moves to the cloud.
We start with logging, because we cannot write those five files properly without it.