Vela

Private, on-device AI assistant for Android. Powered by Gemma 4 — no data leaves your device.

What is Vela

Vela is a fully local voice assistant: hold the mic (or use the system assist gesture), speak in any language, and Gemma 4 — running on the phone's GPU — either answers directly or invokes a tool (timer, calendar, music, contacts, flashlight, volume, app launch, web search). Tool calls run on-device against Android system APIs. The audio goes straight into the model; there is no separate cloud speech-to-text step. No telemetry. No analytics. No accounts.

Privacy

Vela is private by design, not by promise. The reasons are technical and inspectable in this repo:

• Inference runs locally. The model file (gemma-4-E2B-it.litertlm) is downloaded once on first launch from the public HuggingFace litert-community repo and lives in the app's external files dir. All token generation happens inside the process via LiteRT-LM on the device's GPU.
• No analytics or crash SDKs. No Firebase, Crashlytics, Sentry, Mixpanel, Amplitude, Google Analytics, Bugsnag, or any other telemetry library. You can verify this in android/app/build.gradle.kts — the dependency list is short and every entry is either AndroidX, Compose, Hilt, kotlinx, Room, DataStore, Timber (local logging only), or LiteRT-LM.
• Only one outbound network request. The INTERNET permission is used solely by data/local/ModelDownloader.kt to fetch the model from HuggingFace on first run. No other code path opens a socket. The web-search tool fires an Intent.ACTION_WEB_SEARCH which is handled by your default browser — Vela itself does not contact a search engine.
• No remote logs. Logging uses Timber, which writes to logcat only. Nothing is uploaded anywhere.
• Auditable source. GPL-3.0-or-later. Read the code, build it yourself, change it.

Hardware requirements

• Recommended phone: Samsung Galaxy S22 or equivalent — 8 GB RAM, modern Snapdragon/Tensor SoC, GPU with OpenCL support.
• Android version: 8.0+ (API 26).
• Storage: ~3 GB free for the Gemma 4 weights (downloaded on first launch).
• GPU: OpenCL is preferred; the runtime falls back to CPU if Backend.GPU() initialisation fails, but inference latency on CPU-only devices is significantly worse and not the supported configuration.

Lower-end / older devices may install and run, but the assistant experience (especially first-token latency and audio inference) is not guaranteed.

Building

cd android
./gradlew assembleDebug

Java version

This project requires JDK 21. The Kotlin toolchain is pinned to 21 via kotlin { jvmToolchain(21) } in app/build.gradle.kts, and the Foojay resolver in settings.gradle.kts will auto-download a JDK 21 if one isn't installed. If you have JDK 25 installed system-wide and Gradle still picks it up, point JAVA_HOME at an installed JDK 21 — for example, the one bundled with Android Studio:

export JAVA_HOME=/Applications/Android\ Studio.app/Contents/jbr/Contents/Home

JDK 25 is currently incompatible with the Kotlin/Gradle versions used here.

Model download

The APK does not bundle the model weights. On first launch the app downloads the Gemma 4 E2B .litertlm file into the app's external files directory and verifies completion via a sentinel marker:

• Model: gemma-4-E2B-it.litertlm
• Source: https://huggingface.co/litert-community/gemma-4-E2B-it-litert-lm
• Size: ~2.5 GB (Wi-Fi recommended)

ModelDownloader.kt reports per-byte progress to a Flow consumed by ChatScreen (which renders a download UI), supports HTTP byte-range resume on partial downloads, and surfaces errors back to the UI.

Supported features

Tools registered in data/tool/ and exposed to Gemma 4 via @Tool / @ToolParam annotations:

• Timer — set countdown timers via the system clock app (AlarmClock.ACTION_SET_TIMER).
• Alarm — set alarms via AlarmClock.ACTION_SET_ALARM.
• Calendar (read) — query events from CalendarContract.Events for a given time range.
• Calendar (write) — create new events.
• Contacts — look up phone numbers / names via ContactsContract.
• Music — play, pause, skip, previous, now-playing. Targets Poweramp explicitly when installed (its API documents onPlayFromSearch from third-party callers, unlike Spotify which gates that path to Google Assistant). Falls back to whichever app handles MEDIA_PLAY_FROM_SEARCH.
• Device — get battery status.
• Flashlight — toggle the camera torch via CameraManager.setTorchMode.
• Volume — adjust media stream volume.
• App launch — start any installed app by package via PackageManager.getLaunchIntentForPackage.
• Web search — fire Intent.ACTION_WEB_SEARCH so the system browser handles the actual query.

Vela also registers as a VoiceInteractionService so the user can set it as the device's default Assistant — long-pressing home (or using the assist gesture) brings up the Vela overlay.

Known limitations

• Spotify auto-play is gated. Spotify's onPlayFromSearch only honours requests from Google Assistant; third-party callers (including Vela) can open Spotify but cannot make it auto-play a specific artist. If Poweramp is installed it is preferred; otherwise music control uses MediaSession transport keys (pause / next / previous) which work across all apps.
• KV cache auto-resets every 20 turns. Empirically Gemma 4 E2B's persistent dialogue cache starts producing 1-token nonsense after enough turns. Vela now auto-drops the conversation at turn 20 and emits a subtle Snackbar ("Conversation reset to keep responses fresh."). The "+" button in the top bar resets manually any time.
• Lower-end devices not guaranteed. First-token latency depends heavily on GPU support; CPU-only fallback is functional but slow enough to feel broken on phones without OpenCL.
• Phase 6 (always-on listening) is not shipped. Push-to-talk and the system assist gesture are the two entry points.
• Vision input is out of scope for v1.

Contributing

See CONTRIBUTING.md for the fork-and-PR workflow, the recipe for adding a new tool, code style, and how to file a bug report.

License

GPL-3.0-or-later. See LICENSE for the full text.