From notebook to production: 4 books, starting with Andriy Burkov

The piece is built around a specific failure rather than a syllabus. The author's model passed validation and review, then drifted for three months after an upstream table definition changed; worse, the retraining procedure existed only in his head, so handover by word of mouth failed.

That experience explains the ordering he now insists on. 'Machine Learning 100+ Pages Essence' is positioned as a map rather than a dictionary — it does not go deep into implementation, but without it the warnings in the operations book lose their weight. 'Machine Learning Engineering', by the same author, barely discusses algorithms at all; it walks the lifecycle from estimation and labeling through feature design, deployment and post-serving monitoring, and its chapter on data leakage and distribution shift is the one the author says matched his own failure.

The last two books move from judgment to execution. 'Deep Learning Practical Guide' is written by practitioners who shipped AI features at Microsoft, NVIDIA and Square, and its inference-tuning material — quantization, batch size, input pipeline bottlenecks — is the part the author says paid off immediately on latency-sensitive work. 'Building Machine Learning Pipelines' closes the loop with data validation and automated orchestration, the piece of the stack that, once in place, largely ended his unexplained accuracy drops.