Prosise writes for engineers: little proof, plenty of code that runs. He starts with classical models, moves through neural networks, computer vision, text and audio classification, and closes with pretrained cloud services. It is a workshop book.
What I took from it
The argument I use most in proposals: on tabular data — the kind any ERP holds: invoices, customers, movements, dates — a classical model trained on that data usually beats anything that holds a conversation, costs a fraction, and can be explained in a meeting. When someone asks to predict late payment, churn or demand, that is the right answer even when it is not the one they hoped for.
The second: cross-validation and an honest wall between what is used to train and what is used to judge. The most expensive mistake I have seen in data projects is not picking the wrong model, it is measuring it with data it had already seen. The number comes out beautiful and means nothing.
And the third, about scoping: pretrained services are for the generic — reading text from an image, transcribing — and not for what is specific to a company. Knowing where that line falls avoids quoting training that is not needed and promising generic results where the specific ones matter.
What I left out
The whole final stretch, tied to one vendor's cognitive services: those catalogues have changed name and shape since. And it is worth acknowledging what the calendar did to this book: it arrived just before language models rearranged the field, so its text chapters now read as an earlier era. The classical material, by contrast, holds up entirely, and it is the part almost nobody is reading.