My goal is to make things work better and smoother so they fade into the background. Developers and users shouldn't have to worry about whether something will work, it just should. My research focuses on making deep learning on mobile devices work. Specifically, how can we make it so we can use deep learning on mobile devices so that it is fast, reliable, and seamless to both the end-user and the developer.I've approached this by considering this problem from the core of the constraints -- mobile devices are inherently resource-starved and have an extremely heterogeneous operating environment, while deep learning models are large and resource-hungry. By considering inference requests from both sides, both how to optimize individual requests and the impact of aggregated requests, I've shown it's possible to bound latency, increase accuracy and improve resource utilization.