While some AI problems can be solved with end-to-end deep learning models that go from raw inputs to outputs, practitioners (including our customers!) find that such "mega models" are, on their own, not enough to build production-ready AI applications. In practice, it’s critical that AI engineers can inspect, test, and refactor the modular components of their applications, as they would with any piece of infrastructure or software.

In this talk, we’ll introduce a data-centric approach to AI engineering that highlights the advantages of modular components, fine-grained evaluation, and rapid iteration through programmatic labeling. We'll discuss the practical trade-offs of incrementally building and testing pipelines composed of models, preprocessing steps, and business logic. Along the way, we’ll share examples of these principles in practice through real-world case studies.