On-device AI is moving from concept to production. Modern phones and consumer devices now ship with powerful NPUs, GPUs and accelerators, making local inference practical and fast. The key question is no longer whether AI should run on-device, but how to intelligently decide what runs locally, what gets routed to the cloud and how to build coordinated hybrid intelligence systems. In this talk, I’ll share a practical framework for model selection under real constraints like memory, latency, energy, privacy and updates. I’ll also cover proven routing patterns using semantic routing, embeddings, classifiers, rules and lightweight decision layers to handle cloud escalation. A simple Android demo app shows local AI managing fast, private and low-cost tasks while intelligently escalating complex requests to the cloud. Attendees will leave with a clear decision framework and reusable patterns for building real-world hybrid AI products.