On the wine project, the notebook is the product. Distributions, boxplots after outlier removal, SMOTE class counts, and model reports all stay in the file. That is how I remember what I actually did — not what I wish I had done.
Print the shapes
1,599 rows in. 974 after IQR. 2,478 after SMOTE. 1,982 / 496 train–test. Those numbers are the spine of the story. Without them, “93% accuracy” is just a vibe.
Keep failed models visible
SVM at 64.52% is still in the notebook. Leaving weaker results in place makes the Random Forest win credible. Deleting losers turns research into marketing.
Commit the notebook
A clean repo with data + notebook beats a polished slide deck with no provenance. If someone clones the project, they should be able to re-read the experiment.