Deploying models to mobile devices is only the beginning. As AI moves closer to users, on-device inference becomes a critical architectural decision that directly impacts user experience, privacy, reliability, and cost. However, production-grade ML on mobile is not just about model accuracy or performance, it requires experimentation, observability, resilience, and operational control. In this session, we will explore the operational layer needed to run on-device AI safely and at scale. You will learn why controlled rollouts, feature flags, and remote configuration are essential for model experimentation. We will cover strategies for monitoring model health in production, adapting models dynamically to device performance tiers, shipping updates independently of app releases, designing rollback mechanisms, and ensuring reliable behavior in offline or unstable network conditions.