It has lately develop into possible to run private digital assistants on telephones and different private units. On this paper, we describe a design for a pure language understanding system that runs on-device. Compared to a server-based assistant, this method is extra personal, extra dependable, quicker, extra expressive, and extra correct. We describe what led to key decisions about structure and applied sciences. For instance, some approaches within the dialog techniques literature are troublesome to keep up over time in a deployment setting. We hope that sharing learnings from our sensible experiences might assist inform future work within the analysis neighborhood.