Kaggle Gemma 4 Good Hackathon submission · Fully offline · No accounts · No cloud

DreamLit is a native Android app that combines AI-powered dream journaling with a personal lucid dreaming coach — all running entirely on your device. Gemma 4 decodes your dreams, identifies Jungian archetypes, tracks recurring symbols, and coaches you toward lucidity. Your dreams, emotions, and inner life never leave your phone.

InferenceEngine (factory)
    ├── AICoreEngine      ← Android AICore / ML Kit GenAI (Pixel 8+, S24+)
    └── LiteRTEngine      ← LiteRT-LM fallback (Android 8+, 4 GB+ RAM)

VisualEngine              ← MediaPipe on-device Stable Diffusion

Feature layer (engine-agnostic):
    DreamInterpreter  · LucidCoach  · PatternAnalyzer

All text generation engines implement the same LLMEngine interface, so every feature works identically on both AI paths.

Full architecture diagram: see ARCHITECTURE.md

Dream visualization (Stable Diffusion) requires an additional ~1.5 GB download, deferred until the user first requests it.

• Android Studio Meerkat (2024.3) or later
• JDK 17+
• Android SDK with API 35 (targetSdk)

git clone https://github.com/JJRPF/dreamlit.git
cd dreamlit

Open the project in Android Studio. It will sync Gradle automatically.

AICore requires enrollment in the beta program:

1. Join the aicore-experimental Google Group
2. On your Pixel 8+/S24+, opt into the Android AICore beta via Google Play Store (search "Android AICore", tap "Join beta")
3. Wait for the Gemma model to download in the background (usually 1–2 hours on Wi-Fi)
4. The app detects AICore automatically — no code changes needed

Note: If AICore is unavailable (wrong device, model not yet downloaded), the app automatically falls back to LiteRT-LM.

The LiteRT model (gemma-4-e2b.task, ~2.6 GB) is downloaded at first launch via the in-app download screen. To skip the wait during development, place the model file manually:

adb push gemma-4-e2b.task /data/data/com.dreamlit/files/gemma-4-e2b.task

Obtain the model file from Google AI Edge Model Explorer.

# Debug build
./gradlew assembleDebug

# Install on connected device
./gradlew installDebug

# Release build (requires signing config in app/build.gradle.kts)
./gradlew assembleRelease

For voice capture without internet:

1. On the device: Settings → General Management → Language → Text-to-Speech (varies by OEM)
2. Enable offline speech recognition for your language
3. Or: Settings → Accessibility → Live Transcribe → download offline model

app/src/main/java/com/dreamlit/
├── ai/
│   ├── LLMEngine.kt            # Common interface (generate, isReady, close)
│   ├── AICoreEngine.kt         # ML Kit GenAI Prompt API
│   ├── LiteRTEngine.kt         # LiteRT-LM fallback
│   ├── InferenceEngine.kt      # Factory: AICore → LiteRT → NoOp
│   ├── VisualEngine.kt         # MediaPipe on-device Stable Diffusion
│   ├── DreamInterpreter.kt     # Transcript → structured JSON analysis
│   ├── LucidCoach.kt           # Dream sign extraction + technique recommendation
│   └── PatternAnalyzer.kt      # Batched cross-entry weekly insight
├── data/
│   ├── JournalEntry.kt         # Room entity
│   ├── JournalDao.kt           # All DB queries
│   ├── AppDatabase.kt          # Room database singleton
│   └── CsvManager.kt           # Import/export
├── hardware/
│   ├── AudioRecorder.kt        # MediaRecorder M4A wrapper
│   └── VoiceToText.kt          # SpeechRecognizer Flow wrapper
├── notifications/
│   ├── RealityCheckWorker.kt   # WorkManager periodic reality checks
│   ├── WbtbAlarmReceiver.kt    # AlarmManager WBTB alarm
│   └── MorningCheckInWorker.kt # Daily morning dream prompt
└── ui/
    ├── DreamLitNavHost.kt      # Bottom nav + NavHost
    ├── theme/                  # Midnight indigo dark theme
    ├── capture/                # Voice recorder + text input + backdater
    ├── dreams/                 # Dream list + AI detail screen
    ├── lucid/                  # Intentions, technique guides, dream signs
    ├── insights/               # Heatmap, charts, achievements
    └── profile/                # User context, CSV, engine status

• No network permissions — the app cannot make internet requests
• No analytics, no tracking, no accounts
• All AI inference runs locally via Android AICore or LiteRT-LM
• All data stored in Room (SQLite) on-device only
• Audio recordings saved to app-private storage, never shared
• Full data export/import via CSV at any time

The Kaggle notebook at kaggle_notebook.ipynb documents the dual-path inference design, on-device image generation pipeline, privacy architecture, and "for good" impact narrative for the Gemma 4 Good Hackathon.

MIT License — see LICENSE for details.

Built for the Kaggle Gemma 4 Good Hackathon · April 2026