Kask — Local Multimodal AI Assistant for Android

Kask is a high-performance, completely offline AI assistant for Android. It leverages the power of Qwen 3.5 and ASR models to provide Text generation, Vision (image analysis), and ASR (voice transcription) directly on your device.

Built with modern Android standards: Kotlin, Jetpack Compose, Material 3, and Hilt.

Key Features

• Fully Offline: Your data never leaves your device. All inference is performed locally using llama.cpp.
• Multimodal Intelligence powered by Qwen3.5 and Qwen3-ASR:
  • Text: Summarize, rewrite, translate, and chat.
  • Vision: "Ask about this image"—analyze and describe photos or screenshots.
  • ASR: High-accuracy voice-to-text.
• Smart Content Detection: Attach a PDF, Image, or Audio file, and Kask automatically surfaces the most relevant AI actions.
• Hardware-Aware Engine Tiers: Automatically scales from Efficient (0.8B) to Precision (4B) engine tiers based on your device's RAM and storage.
• Local Library & History: Save your AI results, rename notes, and keep track of your conversation history with local SQLite (Room) storage.
• Professional Markdown UI: High-fidelity Markdown rendering for all AI responses, supporting tables, code blocks, and math equations.

---

Architecture

Kask follows a clean, modular architecture:

com.rhuta.kask/
├── KaskApplication.kt          # Hilt entry point
├── MainActivity.kt             # Single-activity host with dynamic Keyboard/Navigation handling
│
├── data/
│   ├── db/                     # Room Database: History and Library persistence
│   ├── network/                # DownloadManager for model bootstrapping
│   └── repository/             # Single source of truth for the UI
│
├── di/                         # Hilt Modules for Database, Engine, and Network
│
├── domain/
│   ├── engine/                 # Core AI logic (AIEngine interface + LlamaCppEngine implementation)
│   ├── model/                  # Domain models: ChatMessage, EngineTier, TaskAction
│   └── util/                   # TextExtractor (PDF/Office), Media Utility
│
├── service/                    # Foreground InferenceService for background processing
│
└── ui/
    ├── home/                   # Main chat interface with smart action tray
    ├── history/                # Searchable interaction logs
    ├── library/                # Document and result management
    ├── markdown/               # Custom WebView-based high-performance Markdown renderer
    ├── settings/               # Preferences and Model Management
    ├── navigation/             # Type-safe Navigation Compose routes
    └── onboarding/             # First-run experience and model downloader

---

Tech Stack

Layer	Technology
Core AI	llama.cpp (b9789) via JNI
Models	Qwen 3.5 (0.8B, 2B, 4B) + Qwen3-ASR Q4_K_M
Language	Kotlin 2.0
UI Framework	Jetpack Compose + Material 3
DI	Hilt
Database	Room (SQLite)
Markdown	Custom CSS/JS Bridge (Highlight.js)
Image Loading	Coil
PDF Parsing	PDFBox (Android optimized)
docx Parsing	Zip + XML (Android optimized)

---

Building & Running

Requirements

• Android Studio: Ladybug or newer.
• JDK: 17+
• Device: Android 10+ (API 29).
  • Efficient Tier: 6GB+ RAM recommended.
  • Precision Tier: 16GB+ RAM recommended.

---

Deployment Highlights

• Intelligent Keyboard Handling: Uses dynamic inset consumption to ensure the chat input tray stays flush with the system keyboard, hiding the bottom navigation bar when typing for maximum focus.
• Resource Management: Implements explicit engine releasing (engine.release()) to free up RAM when starting a new thread.
• Native Prompt Templating: Utilizes the GGUF's internal chat templates for maximum model alignment and accuracy.

---

License

Apache License 2.0 © rhuta 2026