While companies often share overviews of their ML-powered applications, few reveal how their ROI considerations shape the development process. In this post, Dropbox discusses why they adopted a simpler, more interpretable model to maximize the cost impact of Cannes, which was designed to reduce the cost of their Previews feature. For context, Dropbox offers a preview feature that enables users to view a file without downloading it. This feature is powered by Riviera, a system that pre-generates and caches preview assets. To reduce the infrastructure costs of Previews, Dropbox applied ML to predict whether a preview will be used or not. In this post, Win Suen also discusses the processes whereby Dropbox A/B tested, deploys, and monitors Cannes.