Michelangelo Machine Learning Platform

Uber Engineering's 2017 account of Michelangelo, its internal machine learning platform, shows how shared data infrastructure supports live predictions. For meal delivery estimates, models used inputs such as a restaurant's average preparation time over the last seven days or the last hour. Michelangelo precomputed these values and stored them for low-latency retrieval when an order needed a prediction. Its shared feature store let teams reuse named inputs, and Uber said that applying the same data preparation and transformation logic at training and prediction time was meant to keep a model's inputs consistent between development and production. Uber described the feature store as a way to reduce duplicated work across teams. Platform account.