Edge-native intelligence. Zero signal required.

ZeroSignal is an offline-first RAG AI assistant for Android that runs entirely on-device using NexaSDK. It combines Llama 3.2 3B inference with local vector search over downloadable knowledge packs — delivering expert-level answers in emergency medicine, wilderness survival, vehicle repair, and more, with zero cloud dependency.

• Offline AI Chat — Ask natural-language questions and get grounded, context-aware answers powered by on-device Llama 3.2 3B inference
• Local RAG Pipeline — Questions are vectorized and matched against pre-embedded knowledge chunks using ObjectBox's HNSW vector index; top matches are injected into the LLM prompt for factual, hallucination-resistant responses
• Pack Store — Browse and download curated knowledge packs (emergency field medicine, wilderness survival, Yosemite navigation, vehicle mechanics, daily life essentials) from a FastAPI backend; downloads are cursor-paginated and resumable
• NPU-Accelerated — Prioritizes Qualcomm NPU acceleration via NexaSDK for fast inference on Snapdragon devices, with automatic CPU/GPU fallback

ZS-android/                          ZS-web/
├── app/          (Compose UI)        ├── server/    (FastAPI read-only API)
├── core/domain/  (Models, UseCases)  └── pipeline/  (Scrape → Chunk → Embed → Upload)
├── core/ai/      (NexaSDK Engine)
└── core/data/    (ObjectBox, Networking)

Android — 4-module Clean Architecture with Hilt DI, Jetpack Compose, and Material3 Backend — FastAPI serving pre-computed packs from Qdrant Cloud (deployed on Vercel) Pipeline — Local Python tool that scrapes web sources, chunks text, summarizes with LLM, generates embeddings (all-MiniLM-L6-v2), and uploads to Qdrant

• Android Studio Hedgehog (2023.1.1) or later
• JDK 17
• A physical Snapdragon Android device (min SDK 26 / Android 8.0)
• ~6GB free storage on device (for model files)

1. Clone the repository

   git clone https://github.com/CC-ZeroSignal-AI/ZS-android.git
   cd ZS-android

2. Open in Android Studio

   • File → Open → select the ZS-android directory
   • Wait for Gradle sync to complete

3. Connect a Snapdragon device

   • Enable USB debugging on your device
   • Connect via USB and verify with adb devices

4. Build and run

   ./gradlew installDebug

   Or use the Run button in Android Studio targeting your connected device.

5. First launch — Setup

   • The app will detect your device hardware
   • It downloads the Llama 3.2 3B model (~5.5GB total: NPU + GGUF fallback) from HuggingFace
   • Progress is shown in real-time; wait for "Ready"

6. Use the app

   • Chat tab — Ask questions; the app retrieves relevant context from local knowledge packs and streams an AI-generated answer
   • Pack Store tab — Browse and download additional knowledge packs from the ZeroSignal server

No setup needed — the Pack Store backend is already deployed at https://zerosignal-web.vercel.app and the app comes pre-configured to use it. Knowledge packs are available out of the box.

NexaSDK is the core inference engine powering all on-device AI in ZeroSignal. It is integrated in the :core:ai module:

1. NPU-first inference — NexaSDK enables direct execution on Qualcomm's Neural Processing Unit, delivering faster and more power-efficient inference than CPU-only approaches. ZeroSignal uses a dual-runtime strategy: it attempts NPU loading first (NexaAI/Llama3.2-3B-NPU-Turbo-NPU-mobile), then falls back to CPU/GPU via GGUF (bartowski/Llama-3.2-3B-Instruct-GGUF).

2. Streaming generation — NexaSDK's generateStreamFlow() API enables real-time token streaming to the Compose UI, creating a responsive chat experience identical to cloud-based assistants.

3. Chat template support — applyChatTemplate() correctly formats multi-turn conversations with Llama 3.2's expected format, including system prompts with RAG context injection.

4. Lightweight integration — A single Maven dependency (ai.nexa.sdk:nexa:0.0.22) replaces what would otherwise require bundling and managing ONNX Runtime, custom JNI bridges, or other heavyweight inference frameworks.

User Question
    ↓
ObjectBox HNSW vector search → Top-K relevant chunks
    ↓
PromptBuilder assembles Llama 3.2 chat template with context
    ↓
NexaSdkEngine.streamCompletion()
    ├── applyChatTemplate()
    └── generateStreamFlow(maxTokens=2048)
    ↓
Streaming tokens → Compose UI

Five sample packs ship for this hackathon, but the architecture is designed to scale to any domain with new packs addable on demand:

• On-device PDF ingestion — Drop any PDF into ZeroSignal and have it chunked, embedded, and searchable locally — turning personal documents into queryable offline knowledge
• Agentic image search — A dedicated agent that indexes and searches photos on your phone's gallery using natural-language descriptions, powered by on-device vision inference via NexaSDK
• Task agent — Chain multiple on-device tools together with an agentic orchestration layer — all running offline
• Expandable Pack Store — Community and on-demand pack creation for any domain, from medical references to trade manuals

Built by engineers who believe the future of AI isn't bigger servers — it's smarter devices.

ZeroSignal = AI that works even when the signal doesn't.