SideEye

Smart photo management with privacy at its core. SideEye uses on-device AI to scan your gallery for sensitive content and helps you delete, keep, or securely vault flagged images — all without your photos ever leaving your device.

What It Does

SideEye is an Android app that combines on-device AI content moderation with encrypted photo storage:

• Smart Scan — Scans your photo gallery using the OmniNeural-4B Vision Language Model running entirely on-device. Classifies images into configurable categories (nudity, suggestive, drugs, embarrassing) and presents flagged images in a swipeable triage interface where you can delete, keep, or vault each one.
• Private Vault — AES-256-GCM encrypted storage for sensitive images, protected by PIN (PBKDF2 hashed, 100k iterations) and optional biometric authentication with a 15-minute session timeout.
• Gallery — Full photo gallery with multi-select, batch vault operations, and Google Photos cloud backup detection that warns you before sensitive images sync.
• Complete Privacy — All AI inference happens on-device via Qualcomm NPU. No images, metadata, or analysis results are ever sent to a server.

How to Build and Run

Prerequisites

• Android Studio (or just the Android SDK with command-line tools)
• JDK 8+
• An Android device running Android 9+ (API 28)
• A Qualcomm device with NPU is required for Smart Scan (e.g. Snapdragon 8 Gen 3)

Build

git clone https://github.com/baseweight/SideEye.git
cd SideEye
./gradlew assembleDebug

Install

adb install -r app/build/outputs/apk/debug/app-debug.apk

First Run

1. Launch SideEye — the onboarding flow will guide you through:
   • Downloading the OmniNeural-4B model (~2.4GB, WiFi recommended)
   • Granting storage permissions
   • Optionally setting up a vault PIN
   • Choosing which content categories to scan for
2. The app opens to your photo gallery
3. Tap the overflow menu and select Smart Scan to analyze your photos
4. Flagged images appear as swipeable cards — swipe to delete, keep, or vault

Where and Why Nexa SDK Is Used

Nexa SDK (ai.nexa:core:0.0.20) is the AI inference engine that powers SideEye's Smart Scan feature. It provides the runtime for executing the OmniNeural-4B Vision Language Model on Qualcomm's Neural Processing Unit (NPU).

Where

• ImageAnalyzer.kt (app/src/main/java/ai/baseweight/sideeye/data/ai/ImageAnalyzer.kt) — The core integration point. Uses NexaSdk.getInstance().init() to initialize the SDK, then builds a VlmWrapper with plugin_id = "npu" to run inference on the NPU. Each image is resized to 448x448, passed to the VLM with a classification prompt, and the streamed response is parsed for content categories.

• ModelDownloader.kt (app/src/main/java/ai/baseweight/sideeye/data/ai/ModelDownloader.kt) — Downloads the OmniNeural-4B model files from Nexa's S3 model hub (nexa-model-hub-bucket.s3.us-west-1.amazonaws.com) with a HuggingFace fallback (NexaAI/OmniNeural-4B-mobile).

Why

SideEye needs to run a 4-billion parameter Vision Language Model entirely on-device to maintain its privacy guarantee — photos never leave the phone. Nexa SDK was chosen because:

1. NPU acceleration — The npu plugin leverages Qualcomm's dedicated AI hardware for fast, battery-efficient inference instead of burning through CPU/GPU
2. VLM support — Nexa SDK provides first-class support for Vision Language Models through VlmWrapper, handling the multimodal input pipeline (image + text prompt) and streaming token generation
3. Model ecosystem — The OmniNeural-4B model is available through Nexa's model hub, pre-optimized for mobile NPU execution

Inference Flow

Gallery Image → Resize to 448x448 → NexaSdk VlmWrapper (NPU plugin)
    → OmniNeural-4B classifies content → Parse "Category: N" response
    → Flag image if category matches user settings

The model is reinitialized every 4 inferences to maintain NPU stability, and all inference runs on Dispatchers.IO to keep the UI responsive.

Project Structure

app/src/main/java/ai/baseweight/sideeye/
├── MainActivity.kt              # Entry point and navigation graph
├── data/
│   ├── ai/
│   │   ├── ImageAnalyzer.kt     # VLM inference via Nexa SDK
│   │   ├── ModelDownloader.kt   # Model download with S3/HuggingFace fallback
│   │   └── FlagCategory.kt     # Content classification categories
│   ├── security/
│   │   ├── SecurityManager.kt   # Encrypted preferences
│   │   ├── VaultPinManager.kt   # PBKDF2 PIN hashing
│   │   └── BiometricAuthManager.kt
│   └── vault/
│       └── VaultRepository.kt   # Encrypted file storage
└── ui/
    ├── gallery/                 # Photo grid and image viewer
    ├── moderation/              # Smart scan settings and swipe UI
    ├── vault/                   # Vault auth, browser, and settings
    ├── onboarding/              # Setup flow (model download, permissions, PIN)
    ├── about/                   # App info and legal links
    ├── splash/                  # Launch screen
    └── theme/                   # Material3 theming

License

Copyright 2026 Baseweight. All rights reserved.