Adding AI Features to Mobile Apps with TensorFlow Lite and Core ML

The smartphone in your pocket contains a dedicated Neural Processing Unit (NPU) capable of billions of AI operations per second. In 2023, on-device machine learning is accessible to every mobile developer through TensorFlow Lite (Android) and Core ML (iOS). No server costs, no latency, no data leaves the device.

Why On-Device ML?

• Privacy — user data never leaves the device
• Latency — inference in under 10ms, no network round-trip
• Offline — works without internet connectivity
• Cost — no cloud inference costs at scale

TensorFlow Lite on Android

Add the dependency, load a .tflite model from the assets folder, and run inference:

// build.gradle
implementation 'org.tensorflow:tensorflow-lite:2.12.0'
implementation 'org.tensorflow:tensorflow-lite-support:0.4.3'

// Kotlin
val model = ImageClassifier.createFromFile(context, "mobilenet_v2.tflite")
val image = TensorImage.fromBitmap(bitmap)
val results = model.classify(image)
results.forEach { classification ->
    Log.d("AI", "${classification.categories[0].label}: ${classification.categories[0].score}")
}

Core ML on iOS

Drag a .mlmodel file into Xcode. It auto-generates a Swift class with a type-safe prediction API:

import CoreML
import Vision

let model = try VNCoreMLModel(for: MobileNetV2().model)
let request = VNCoreMLRequest(model: model) { request, _ in
    guard let results = request.results as? [VNClassificationObservation],
          let top = results.first else { return }
    print("(top.identifier): (top.confidence)")
}

let handler = VNImageRequestHandler(ciImage: ciImage)
try handler.perform([request])

Model Formats and Conversion

Train your model in TensorFlow, PyTorch, or scikit-learn, then convert:

• Android: TFLiteConverter from TensorFlow → .tflite
• iOS: coremltools (Python) from TensorFlow/PyTorch → .mlmodel

Pre-Trained Models (No Training Required)

• Object detection: MobileNet SSD, EfficientDet
• Image classification: MobileNetV2, EfficientNet
• Text classification: MobileBERT
• Face detection: MediaPipe Face Mesh

For most product use cases, fine-tuning a pre-trained model on your domain data (transfer learning) outperforms training from scratch and requires a fraction of the data.

On-device AI is no longer the exclusive province of AI specialists. With TensorFlow Lite and Core ML, any mobile developer can add genuinely intelligent features to their app today.