Briefly

On-device notification summarizer for Android. Briefly listens to your incoming notifications, clusters them by app and conversation, and uses a local LLM (via llama.cpp) to produce a short digest — without sending anything off the device.

Internal package name is still com.notifyai (the project was renamed mid-way).

Notes on what shipping this actually looked like — capability probe design, hardware ceilings, the 1 GB → 114 MB APK story, with logcat and code refs: NOTES.md.

Features

• NotificationListenerService captures incoming notifications and stores them in Room.
• Clustering groups notifications by app / conversation thread.
• On-device summarization runs a quantized GGUF model through llama.cpp. Falls back to a heuristic summary on devices that can't sustain real-time inference.
• WorkManager runs periodic summarization in the background.
• Jetpack Compose UI with Home, History, and Insights screens.

Tech stack

• Kotlin + Jetpack Compose, Material 3
• Hilt (DI), Room (storage), WorkManager (background), DataStore (preferences)
• llama.cpp (C++ inference, built via CMake / NDK)
• Min SDK 29, target SDK 35

Project layout

app/src/main/
├── cpp/                       # JNI glue + vendored llama.cpp
│   ├── ai_chat.cpp
│   ├── CMakeLists.txt
│   └── llama.cpp/             # upstream source (vendored)
├── java/com/notifyai/
│   ├── ai/                    # engines, summarizer, clustering, model downloader
│   ├── data/                  # Room DAOs, entities, repositories
│   ├── di/                    # Hilt modules
│   ├── domain/                # models, repository interfaces, use cases
│   ├── service/               # NotificationListenerService
│   ├── ui/                    # Compose screens + view models
│   └── worker/                # WorkManager workers
└── java/com/arm/aichat/       # InferenceEngine abstraction over llama.cpp

Third-party code

A few directories contain vendored upstream sources rather than first-party code. They live in-tree because their upstreams don't ship through Maven/Gradle.

app/src/main/cpp/llama.cpp/

Vendored copy of ggerganov/llama.cpp pinned at release b9524. The NDK build compiles it from source:

1. app/build.gradle.kts points CMake at app/src/main/cpp/CMakeLists.txt.
2. That CMakeLists.txt includes llama.cpp/CMakeLists.txt, producing libllama.so and libggml*.so.
3. The JNI glue in ai_chat.cpp #includes llama.cpp headers and links against those libraries.

Pinning a specific commit guarantees reproducible builds — llama.cpp moves fast and frequently breaks its ABI. A git submodule would work equally well; the checked-in copy was chosen for simplicity.

app/src/main/cpp/vendor/

Single-header C++ dependencies used by the JNI layer. Currently just nlohmann/json (vendored at vendor/nlohmann/json.hpp) for JSON parsing in ai_chat.cpp.

com.arm.aichat package

Vendored Kotlin wrapper around llama.cpp's JNI surface, adapted from Arm's reference arm/ai-chat-android project. It sits between the app and the C++ engine:

• InferenceEngine — typed Kotlin API (loadModel, setSystemPrompt, sendUserPrompt, bench) with a StateFlow<State> lifecycle (LoadingModel → ModelReady → Generating → …).
• AiChat — one-call entry point (AiChat.getInferenceEngine(context)).
• gguf/GgufMetadataReader — reads architecture, quant type, and context length from a GGUF file so unsupported models are rejected before loading.
• internal/InferenceEngineImpl — calls into JNI, manages the native context, and surfaces generated tokens as a Flow<String>.

It's kept in its own com.arm.aichat package (rather than folded into com.notifyai) to make the upstream boundary explicit, preserve attribution, and keep an easy update path. Inside the app it has exactly one consumer: LlamaCppEngine.kt, which adapts the generic InferenceEngine to the app's LocalLlmEngine domain interface.

Build

Prerequisites:

• Android Studio (Koala or newer)
• Android NDK 29.0.13113456
• CMake 3.31.6
• A device or emulator running Android 10+ (minSdk = 29)

# Configure local SDK path
echo "sdk.dir=/path/to/Android/sdk" > local.properties

# Build a debug APK
./gradlew :app:assembleDebug

# Install on a connected device
./gradlew :app:installDebug

The first launch downloads the GGUF model (~1 GB) into the app's private files directory. Configure the URL in ModelDownloader.kt.

Permissions

Briefly needs the Notification Access special permission. After install, grant it via Settings → Notifications → Device & app notifications → Notification access → Briefly. The app routes the user there from the onboarding screen.

It also requests POST_NOTIFICATIONS, INTERNET (for the one-time model download), and runs a foreground service during inference.

Privacy

Notification content never leaves the device. The only network traffic is the initial model download from a public URL.

License

See LICENSE.