As the world’s leading flight tracking service, FlightAware has offered public and commercial flight-tracking services for nearly 15 years. Our newest commercial offering provides machine-learning-based flight arrival predictions based on years of flight tracking data. This talk will share some of the problems encountered and how we solved them to bring predictive arrival times to some of the world’s largest airlines. How would we massively scale up a brand-new product on a brand-new platform effectively and efficiently? How would we transition from zero cloud to the massive scale of 100,000 cores we would need for our end goal of training 3,000 models in parallel? Reducing process and cultural bottlenecks, improving machine learning efficiency, and scaling with the cloud are just some of the outcomes of the journey we have taken. Come and see what lessons we’ve learned.