The book is about running models at the edge: on the device that measures, not on a server. Machine monitoring, counting, keyword spotting, cameras that send no images anywhere. Half the content is not about models at all but about framing the problem and collecting data, which is where these projects actually fail.
What I took from it
The decision tree about where execution belongs. If connectivity is intermittent, if the data is sensitive, if the answer has to arrive in milliseconds, or if shipping everything to the cloud costs more than the device, the answer changes. In a warehouse in Quito with a weak antenna, that conversation happens before choosing a tool, not after.
The second idea I apply everywhere: data before model. The authors insist on going out and collecting data from the real site, with the real site's noise, before training anything. Translated into an ERP rollout with peripherals, it means going to look at the scale, the scanner and the printer before writing the integration, because the data that arrives in practice is never the data in the manual.
And a third, uncomfortable one: the system has to behave sensibly when the model is wrong, because it will be wrong. An unattended device has nobody to ask.
What I left out
All the hardware and tooling detail. That part has aged, and it belongs online rather than on paper. And honestly, the whole book sits to one side of what I do: I read it because peripherals turn up in nearly every implementation and I wanted to understand the boundary. What stayed was judgement, not technique.