80 million users, nine billion radio stations, 75 billion thumbs. That's a lot of data points. Pandora provides an evolving set of sequential items, and needs to react in just a few milliseconds for each listener. There's also a variety of factors (musicological, social, geographical, generational) playing a critical role in deciding what music to play to a user. This presentation will show how a dynamic ensemble learning system that combines curational data and machine learning models provides a truly personalized experience. This approach allows P to switch from a lean back experience (exploitation) to a more exploration mode to discover new music tailored to each users individual tastes.