NutriTrack - On-Device Food & Macro Tracker (Google Pixel 10 Pro)

NutriTrack is an edge-first nutrition and macronutrient tracking Android application engineered specifically for the Google Pixel 10 Pro. It utilizes on-device Machine Learning trained on Google Research's Nutrition5k dataset to predict meal macronutrients from overhead photos, and Google ML Kit paired with the Open Food Facts API for rapid barcode scanning.

---

🔒 Strict On-Device & Privacy Architecture

As specified in the system design:

All computation occurs locally on the phone. The only network call leaving the application is for looking up product barcodes against the Open Food Facts API.

1. Native 16 KB Memory Page Size Compatibility (Android 15 / Android 16 / Pixel 10 Pro Tensor):
   • Built to run 100% natively on 16 KB page-size kernels without compatibility shims or system warning dialogs.
   • Migrated to Google LiteRT 2.3.2 (com.google.ai.edge.litert:litert), which natively compiles all .so binaries (libLiteRt.so, liblitert_jni.so) with 16384 (0x4000) segment alignment.
   • Configured with packaging { jniLibs { useLegacyPackaging = false } } so native shared libraries remain uncompressed in the APK and automatically aligned to 16 KB page boundaries for zero-copy memory mapping (mmap).
   • Verified via zipalign -v -c -P 16 4: All native binaries pass 16 KB page boundary verification.
2. Food Photo Prediction (Nutrition5k): Runs 100% offline on the device using an optimized TensorFlow Lite neural network running directly on the Pixel 10 Pro's Google Tensor TPU/NPU (via NNAPI / GPU delegates). Zero photo bytes or features ever leave the device.
3. Barcode Recognition: Processed 100% offline on the phone via Google ML Kit Vision Barcode Scanning.
4. Open Food Facts Lookup: Once a barcode is decoded, the app contacts https://world.openfoodfacts.org/api/v2/product/{barcode}.json with the product identifier. Scanned products are saved to a local SQLite cache so repeated scans never hit the network.
5. Operating System Network Firewall (network_security_config.xml):
   • cleartextTrafficPermitted="false"
   • Domain whitelisting strictly limited to *.openfoodfacts.org and *.openfoodfacts.net.
   • All other external network connections are rejected at the Android networking stack.
6. Local Persistence: Meals, daily diaries, macro targets, and historical metrics are stored in a private, encrypted on-device SQLite database (nutritrack_local.db).

---

🧠 Nutrition5k Machine Learning Pipeline

• Dataset: Google Research Nutrition5k (Thames et al.) comprising 5,000+ real-world food dishes with overhead RGB imagery, accurate mass (grams), and continuous nutritional ground truth.
• Model Architecture:
  • Backbone: Lightweight MobileNetV3 feature extractor.
  • Head 1: Multi-output regression head predicting continuous values for:
    • Total Caloric content (kcal)
    • Total Mass (g)
    • Protein (g)
    • Fat (g)
    • Carbohydrates (g)
  • Head 2: Softmax dish classification head across 16 canonical cafeteria meal classes.
• Physiological Consistency: Predictions enforce Atwater caloric relationship sanity: $$\text{Calories} \approx 4 \times \text{Protein} + 4 \times \text{Carbohydrates} + 9 \times \text{Fat}$$
• Hardware Acceleration:
  • Automatically activates Google Tensor GPU delegate or Tensor NPU (NNAPI) with fallback to multi-threaded CPU XNNPACK.
  • Typical inference latency on Google Pixel: ~8ms - 15ms.
• Model Asset: Packed directly inside app/src/main/assets/nutrition5k_macro_predictor.tflite (with noCompress += "tflite" enabled for zero-copy memory-mapped inference).
• Training Pipeline Script: Included in ml_pipeline/train_nutrition5k.py.

---

📦 Key Components & Directory Structure

nutritrack/
├── app/
│   ├── src/
│   │   ├── main/
│   │   │   ├── assets/
│   │   │   │   ├── nutrition5k_macro_predictor.tflite   # Trained 16-bit float TFLite model
│   │   │   │   ├── dish_classes.txt                    # Category mapping
│   │   │   │   └── model_metadata.json                 # Model input/output metadata
│   │   │   ├── kotlin/com/nutritrack/app/
│   │   │   │   ├── MainActivity.kt                     # Application entry point
│   │   │   │   ├── model/
│   │   │   │   │   ├── FoodItem.kt                     # Meal and food entity data models
│   │   │   │   │   ├── MacroTargets.kt                 # Targets & Mifflin-St Jeor TDEE formulas
│   │   │   │   │   └── PredictionResult.kt             # Prediction & Open Food Facts DTOs
│   │   │   │   ├── data/
│   │   │   │   │   ├── FoodDatabaseHelper.kt           # On-device SQLite database
│   │   │   │   │   └── FoodRepository.kt               # Reactive repository flow
│   │   │   │   ├── network/
│   │   │   │   │   └── OpenFoodFactsClient.kt          # Isolated barcode API client
│   │   │   │   ├── ml/
│   │   │   │   │   └── Nutrition5kInferenceEngine.kt   # TFLite Tensor NPU/GPU runner
│   │   │   │   ├── camera/
│   │   │   │   │   └── BarcodeAnalyzer.kt              # ML Kit real-time barcode analyzer
│   │   │   │   └── ui/
│   │   │   │       ├── NutriTrackApp.kt                # Jetpack Compose navigation shell
│   │   │   │       ├── components/
│   │   │   │       │   ├── MacroRings.kt               # Daily Calorie ring & macro bars
│   │   │   │       │   └── NutriScoreBadge.kt          # European Nutri-Score A-E badge
│   │   │   │       └── screens/
│   │   │   │           ├── DashboardScreen.kt          # Today's overview & meal logs
│   │   │   │           ├── CameraScanScreen.kt         # Dual Food AI / Barcode CameraX
│   │   │   │           ├── DiaryScreen.kt              # Day-by-day historical journal
│   │   │   │           └── GoalsScreen.kt              # TDEE calculator & security status
│   │   │   ├── res/
│   │   │   │   └── xml/
│   │   │   │       └── network_security_config.xml     # Domain whitelisting firewall
│   │   │   └── AndroidManifest.xml
│   │   └── test/
│   │       └── kotlin/com/nutritrack/app/
│   │           └── MacroTargetsTest.kt                 # Unit tests
│   └── build.gradle.kts
├── ml_pipeline/
│   └── train_nutrition5k.py                            # Nutrition5k training & export script
└── build/outputs/apk/debug/app-debug.apk               # Compiled Android APK

---

🚀 Building & Installing on Pixel 10 Pro

Prerequisites

• JDK 17+ (Microsoft OpenJDK 17 or Eclipse Temurin 17)
• Android SDK Platform 36 (compileSdk = 36)
• Google Pixel 10 Pro with USB Debugging enabled

1. Build Debug APK

$env:JAVA_HOME = "C:\Program Files\Microsoft\jdk-17.0.21.21-hotspot"
cd nutritrack
.\gradlew.bat assembleDebug

2. Run Local Unit Tests

.\gradlew.bat testDebugUnitTest

3. Install via ADB onto your Pixel 10 Pro

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