Astronomers are facing an unprecedented era of big data, observing more phenomena than humans can possibly visually examine alone. Upcoming large-scale surveys such as the Large Synoptic Survey Telescope (LSST) will observe millions of transient alerts each night: two orders of magnitude more than any survey to date. To meet this challenge, we have developed a novel time-series classification tool, RAPID (Real-time Automated Photometric IDentification), capable of quickly classifying multi-channel, sparse, time series datasets into several astrophysical types. Using a deep recurrent neural network, we present the first method specifically designed to provide early classifications of astronomical transients, identifying transients from within a day of the initial alert, to the full lifetime of a light curve. We have begun running RAPID on the real-time Zwicky Transient Facility (ZTF) survey, and have successfully classified several transients well before peak luminosity. In this talk, I will explain the main parts of our deep learning architecture and describe our approach’s performance on simulated and real data streams.