Box fuses llama.cpp, LiteRT, NPU routing

The interesting part here is not just "offline AI on Android" but the routing layer: Box is trying to make heterogeneous mobile silicon usable in one app instead of betting on a single backend.

• –Hybrid dispatch across LiteRT, llama.cpp, and native speech/image stacks is a practical way to squeeze more out of Snapdragon and Pixel-class devices than a one-runtime approach
• –The feature set is unusually broad for a phone app: voice-to-voice chat, camera Q&A, local document ingestion, custom GGUF import, and Stable Diffusion generation all without cloud dependencies
• –On-device memory and persistence are likely the real limiting factors now, which matches the maintainer’s own framing more than raw compute throughput
• –The security posture is stronger than most local-AI demos: hard offline mode, encrypted history, biometric lock, and prompt sanitization make the privacy story credible
• –This is most relevant as infrastructure for local-first mobile AI builders, not as a consumer assistant replacement for cloud chatbots