Mobile app for daily personal use where the conversational engine is at the heart of the experience. The client wanted a product with a real narrative personality — not a generic ChatGPT wrapper — coupled with real-time BLE hardware integration.
Combine two LLMs with distinct roles (an in-house fine-tuned narrative model + an instruction-tuned one for structured actions) without blowing up latency or API costs. And make this engine dialog with a custom BLE sensor that doesn't forgive any delay (sub-100ms required on certain events).
Cascading LLM pipeline: lightweight router on the Flutter side decides on the fly which model to invoke based on intent. Semantic cache on recurring prompts (40% reduction in API calls). On the hardware side, separate native Swift + Kotlin plugin exposing an async BLE channel to Flutter via an optimized method channel — we wrote JS for the React Native bridge earlier, dropped it for native Flutter after benchmarks.
In production since 2025, with continuous iteration alongside the client. The narrative engine stays light enough to run with a controlled envelope even at peak usage.