Go beyond "add an AI feature" - understand how on-device and cloud intelligence actually work on Android.
Most Android AI tutorials give you a checklist: add this dependency, call this API, paste this prompt. This book teaches you the mental model underneath the checklist.
If you've ever wondered why NNAPI's deprecation changes how you build fallback logic, what AICore actually checks before deciding a device is eligible for Gemini Nano, how a model can be memory-mapped instead of fully loaded so the OS can evict it under pressure without crashing your feature, or why a model that passes every golden-dataset test can still quietly fail in production - this is the book that answers those questions from the inside out.
Mastering Android AI: From Fundamentals to Production takes you from the on-device-vs-cloud trade-off Chapter 1 establishes to the deepest layers of the current Android AI stack: LiteRT delegates and the post-NNAPI hardware path, ML Kit's task APIs, Gemma and other open on-device LLMs, Gemini Nano through AICore, and cloud Gemini through Firebase AI Logic - plus quantization trade-offs, thermal-adaptive performance budgets, and staged production rollouts. You won't just learn which API to call - you'll learn why a given model or delegate is right for one feature and wrong for another, so you can architect features that hold up under a real battery budget, a real device-fragmentation spread, and a real production incident, not just a demo.
Inside, you'll learn:
- How Android's AI stack fits together - LiteRT delegates and vendor NPU extensions, AICore and Gemini Nano as a platform service, and ML Kit's Play-services-backed updates, and which of the three to reach for on a given feature
- The full lifecycle of an on-device model: quantization trade-offs (FP32 to INT8/INT4), memory-mapped loading versus full residency, delegate selection across GPU/NPU/CPU, and thermal-adaptive patterns that keep a feature usable for ten minutes, not just one inference
- How cloud and on-device generation really work - Gemini's multimodal image generation versus the standalone Imagen API, streaming versus single-shot responses, and the cost-utility math that tells you when a cloud call stops being worth it
- How to architect features that combine both directions - on-device candidate generation with cloud re-ranking, offline-first assistants, and the full privacy spectrum from raw-to-cloud through federated and fully on-device
- Testing AI features when there's no single correct answer to assert against - layered golden-dataset, integration, and human-evaluation strategies that catch a silent regression before a user does
- The production lifecycle of a shipped AI feature - canary rollouts, per-user model-version bucketing, drift detection, and rollback, because a model behaving correctly in testing is the start of the relationship, not the end of it
- Privacy and responsible AI as engineering requirements, not policy language - on-device data isolation, subgroup accuracy auditing, and real consent and deletion paths for features that build a behavioral profile
- Three shippable capstone projects - an AI-powered camera app, an intelligent personal assistant, and a health-and-fitness AI companion - tying every technique to a feature that has to survive a code review
Whether you're an Android developer shipping your first AI feature, a team lead deciding where a feature should sit on the on-device-vs-cloud spectrum, or an engineer inheriting a live AI feature that needs to be made production-safe, this book gives you the mental model professional Android AI engineers use - from the API you call down to the device it runs on.
No fluff. No toy demos. Just a rigorous, complete map of how AI features get built and shipped on Android.
