This is one of the more practical and useful articles I've read about ML in practice. Going through these calculations feels much more like solid engineering vs the hand-wavy 'use a different model' or 'have you tried model distillation' type performance discussions I see elsewhere.