Georges Byona

Network outages, costly mobile data, and areas with weak signal are still everyday challenges across African markets. Yet, most mobile AI systems continue to depend on constant cloud connections and expensive APIs. This session challenges that norm. We show how to run Google’s Gemma models directly on user devices using Flutter and flutter_gemma, enabling fast, private, and fully offline AI experiences that respond in under a second.

Drawing from real-world experience building Gemmea (1st place winner at GDG on Campus UCB Build with Gemma 2026), this interactive talk walks through the full Edge AI workflow for cross-platform mobile apps. We dive into model optimization techniques like INT4 and INT8 quantization, examine memory boundaries on mid-tier Android smartphones (3GB–6GB RAM), and explore how to integrate MediaPipe’s LLM Inference engine directly with Flutter apps.

Beyond the technical setup, we share candid lessons from what didn’t work: device thermal throttling, Out-Of-Memory (OOM) crashes on budget Android phones, UI thread lockups, and the architectural choices, like offloading inference to Flutter Isolates, that made the system reliable. By the end, attendees walk away with proven strategies, reusable code snippets, and practical design patterns to build and deploy offline-first AI apps in 2026.