flutter_gemma started as a thin wrapper over MediaPipe for running Google Gemma on Android and iOS. Today it's a full-fledged platform for on-device AI: 6 platforms (Android, iOS, Web, macOS, Windows, Linux), 2 inference engines, multimodal support, function calling for local agents, and on-device RAG with vector search. The plugin has been featured by Google AI for Developers.

In this talk, we'll walk through the evolution via key engineering decisions. How adding Web platform forced us to rethink file handling and led to sealed classes instead of string URLs. Why we had to spin up a gRPC server in Kotlin with a bundled JVM for desktop — and how to automate the build through Xcode build phases. How the Strategy pattern allowed us to add a second inference engine (LiteRT-LM) without rewriting existing code, and how the Adapter pattern helped reuse the MediaPipe implementation. Why we needed chunk buffering for engines with fundamentally different APIs.

We'll dive deep into on-device RAG: how to build a SQLite VectorStore that works identically on mobile and in the browser via WASM. I'll show real bugs and their fixes — from Web hot restart crashes to iOS Simulator limitations with vision models.

Every architectural decision in flutter_gemma is an answer to a specific problem. Minimum theory, maximum code, diagrams, and stories of "how it broke and why it looks like this now".