TranslateOff

Translation apps want signal, and you need them exactly where signal is worst. TranslateOff removes the network from the loop entirely: the translation model runs on the phone, so after a one-time download the app never talks to anything again.

The interaction is built for two people and one device. The screen splits in half with the far side rotated 180° to face the other person, each half in its own language, each with its own mic button. Every spoken turn appears on both halves in both languages.

Core Features:

• Fully offline translation after a one-time model download — no account, no server, no conversation logs
• Two-sided split screen, far half rotated, with a volume-reactive mic button per speaker
• Every turn shown in both languages on both halves, tap any line to hear it read again
• Streaming output so text appears as the model produces it rather than after it finishes
• Optional spoken playback with explicit voice resolution, and an honest warning when the device is missing the voice data

Technical Architecture:

• Expo SDK 57 / React Native 0.86 / React 19 / TypeScript, new architecture, dev-client only (the app is native code end to end)
• A 4B quantized translation model executed on-device through llama.rn: a resident context with GPU offload and CPU fallback, a serialized completion queue, and a hand-built prompt template with explicit stop handling
• Speech-to-text through the platform recognizers with a per-tap locale, interim results, and volume events driving the mic animation
• Custom Kotlin module wrapping the OS download manager for the multi-gigabyte model transfer, so it survives the app being killed — with a persisted download handle, polling progress, and re-attach on relaunch
• Dual storage roots with byte-level verification before a downloaded model is trusted
• Streaming callback that reports the whole translation so far rather than deltas, so displayed text only ever grows and never rewrites itself mid-sentence

Engineering Highlights:

• Chose on-device inference over every cloud API by definition — privacy is the product, not a feature of it
• Evaluated and rejected a bundled speech model: several hundred megabytes more, hand-rolled voice detection, and worse accuracy than the free platform recognizers
• Separate build identities for development and preview installs so both can coexist on one device, each with its own model copy
• Kept the whole runtime dependency-free at the network layer: the only transfer that ever happens is the initial model download