Milan Šulc, Rossum.ai Petr Baudiš, Rossum.ai Solving real-world problems at scale often requires more than direct application of straightforward ML models. Let's journey together through architecting a complex ML pipeline and show how a challenging high-level task can be decomposed into a series of trainable sub-tasks while not compromising on a pure machine learning approach. We will demonstrate this on the problem of document information extraction that we are solving at Rossum, and look at how it can be decomposed to (still hard, but attackable) sub-tasks such as named entity (field) localisation, tables recognition, key-value detection, few-shot learning via similar document retrieval, and, of course, OCR. And perhaps we will manage to show why building an AI system capable of understanding document content is so much more than