AI Recommendation System for Mobile Apps: Implementation Guide

The Problem: Slow or Irrelevant Recommendations

Recently, an e-commerce client approached us: their iOS app showed recommendations with a 2-second delay — users scrolled past before they loaded. The conversion rate in the recommendation block was 1.2%. We moved the final reranking to the device using CoreML — response time dropped to 50 ms, conversion rose to 4.1%. Savings on server resources: 40% (about $3,500/month) due to reduced requests. The cold start for new users was solved with an onboarding quiz (2 preference questions) and a popularity-based fallback. After 10 sessions, personalized recommendations worked.

We build AI recommendation systems not as black boxes but as pipelines: collecting behavioral events, feeding them into ML models, ranking, and embedding into the UI without performance loss. Our team has 7+ years of experience and 15+ projects for iOS and Android. A hybrid architecture outperforms pure server-side: CTR is 2–3 times higher with the same data.

How to Choose Architecture: On-Device or Server?

On-device reranking cuts latency by 3–5 times and saves up to 60% server resources (up to $4,000/month). Testing shows a hybrid approach improves CTR 2–3 times over pure server-side.

Why Event Collection Is the Foundation of Quality?

A recommendation system is only as good as its data. On mobile, you must log at minimum:

• item_view — object view (with dwell time, not just impression)
• item_click — tap/click on object
• item_purchase / item_save — conversion action
• item_skip — scrolled past (important negative signal)

// Android: batched event logger
class RecoEventLogger(private val api: RecoApi) {
    private val buffer = mutableListOf<RecoEvent>()
    private val flushInterval = 30_000L // 30 seconds

    fun log(event: RecoEvent) {
        buffer.add(event.copy(timestamp = System.currentTimeMillis()))
        if (buffer.size >= 20) flush() // or by timer
    }

    private fun flush() {
        if (buffer.isEmpty()) return
        val batch = buffer.toList()
        buffer.clear()
        viewModelScope.launch(Dispatchers.IO) {
            runCatching { api.sendEvents(batch) }
            // On error — write to Room for retry
        }
    }
}

Important: dwell time is often a missed signal. Track when a card enters the viewport (RecyclerView.OnScrollListener or LazyList.onVisibleItemsChanged) and when it leaves. A view under 2 seconds is probably a scroll-through. In one project, adding dwell time increased CTR by 18%.

How On-Device CoreML/TFLite Reranking Works?

If the server returns top-200 candidates, final ranking can happen on device. This eliminates an extra network request on every screen open.

On iOS with CoreML:

// Load model (bundled or via Core ML Model Deployment)
let model = try MLModel(contentsOf: modelURL)
let input = RerankerInput(
    userVector: userEmbedding,        // Float32 array 64d
    itemVectors: itemEmbeddings,      // [Float32 array 64d]
    sessionFeatures: sessionContext   // last 10 actions
)
let output = try model.prediction(from: input)
let scores = output.featureValue(for: "scores")?.multiArrayValue

TensorFlow Lite on Android uses Interpreter with ByteBuffer input. For models >10 MB, use GPU delegate (GpuDelegate) — acceleration of 3–8x on flagships.

Updating the model without an app release: on iOS — Core ML Model Deployment via CloudKit or custom CDN with MLModel.compileModel(at:). On Android — Firebase ML with RemoteModel or direct .tflite download into filesDir with hash verification.

Steps to Implement a Recommendation System

1. Data & Event Audit — check which events are already logged, add missing ones (dwell time, skip).
2. Architecture Selection — decide what lives on server vs. on device.
3. Develop Event Tracker — with batching, retry mechanism, Room storage for offline.
4. Server Model — collaborative filtering or a ready service (Amazon Personalize, Google Recommendations AI).
5. Client Model Integration — CoreML/TFLite, reranking candidates.
6. UI Components — adaptive blocks with lazy loading.
7. A/B Testing — Firebase Remote Config, Amplitude Experiment.
8. Documentation & 6-Month Guarantee.

What On-Device Reranking Delivers (Comparison)

How to Handle Cold Start?

First 5–10 sessions lack data for personalization. Standard approach — hybrid:

1. Onboarding quiz (2–3 preference questions) gives initial profile.
2. Popularity-based recommendations as fallback.
3. Implicit feedback from first interactions quickly shifts profile.

Avoid showing “recommendations for you” until minimal history is collected — it’s fair to the user and keeps metric quality.

Which Quality Metrics to Track?

Click-through rate (CTR) and conversion are basic. But for mobile UX, also track “recommendation blindness”: if the block is ignored, it’s worse than low CTR. A/B testing via Firebase Remote Config or Amplitude Experiment is mandatory when changing algorithms. Minimum sample for statistical significance: 1000+ unique users per variant.

What’s Included (Deliverables)

• Technical documentation: event tracker architecture, data model.
• Source code of event tracker with batching and retry (Swift/Kotlin).
• Integration of server recommendation model (or custom).
• In-app UI recommendation component with lazy loading.
• A/B testing and metric monitoring setup.
• Team training on system usage.
• 6 months of technical support.

Timeline Guidelines

Integration of a ready server recommendation service with event tracker — 2–3 weeks. Hybrid system with on-device reranking, custom events, and A/B testing — 6–10 weeks. Cost is determined individually.

Contact us for a free audit of your app — we will assess your architecture and propose the optimal solution. Request implementation and get a consultation on model selection.

Machine Learning in Mobile Apps: CoreML, TFLite, and On-Device Models

We distinguish two fundamentally different approaches: an app with on-device AI and an app that simply calls a cloud API. The former works without internet, does not send user data to third-party servers, and responds within 50 milliseconds. The latter depends on network latency and pricing plans. Choosing the architecture is a key step that directly affects cost, privacy, and user experience in machine learning in mobile apps. Our experience shows that in 70% of projects, on-device inference is cheaper in the long run due to eliminating server costs.

How to Choose Between CoreML and TFLite for On-Device Inference?

CoreML — Apple's native framework for running ML models on device. Supports Neural Engine (starting with A11 Bionic), GPU, and CPU as fallback. Models are converted to .mlmodel format via coremltools from PyTorch, ONNX, or TensorFlow. Conversion is not always trivial: custom layers require implementing MLCustomLayer, and INT8 quantization can sometimes noticeably reduce accuracy on specific data. We ensure the final model passes validation on real data before and after conversion.

TensorFlow Lite — cross-platform alternative for Android and Flutter. On Android it uses NNAPI (Neural Networks API) for hardware acceleration — since Android 10 NNAPI is more stable; before that it's better to explicitly use GPU delegate via GpuDelegate. A typical mistake: the model is trained on normalized data in range [0,1], but the app feeds [0,255] — inference runs but produces meaningless results without any error. We include an automatic input data validation module in the SDK.

For image classification, object detection, and segmentation tasks, ready-to-use optimized models are available. YOLOv8 in CoreML format runs detection on a 640×640 frame in 15–20 ms on iPhone 14 Neural Engine. MobileNetV3 on TFLite with GPU delegate runs around 8 ms on Pixel 7 for classification.

What If You Need Text Generation On-Device?

Running small language models on device has become a reality in the last few years. Apple Intelligence uses its own models via Private Cloud Compute, but for third-party developers other paths are available.

llama.cpp with Metal backend on iOS is a working approach for phi-3-mini (3.8B parameters, 4-bit quantization, ~2.3 GB). Inference: 15–25 tokens/second on iPhone 15 Pro. For integration in Swift, use the Swift Package llama.swift or a wrapper via C interface llama.h. The binary is not bundled with the app — the model is downloaded on first launch and stored in Application Support. Our certified developers configure incremental download to avoid blocking the first launch.

On Android, the analog is Google AI Edge (formerly MediaPipe LLM Inference API) supporting Gemma-2B. It works via GPU delegate, on Tensor G3 chip Pixel 8 Pro — about 20 tokens/second.

Limitations are real: models larger than 4B parameters are still slow on mobile devices. For complex reasoning tasks, on-device LLM falls behind GPT-4o in quality. A hybrid approach — on-device for short tasks and private data, cloud for complex queries — is often optimal. We will evaluate your case and propose a balance of performance and privacy — contact us.

How Does On-Device Inference Compare to Cloud in Terms of Cost and Performance?

On-device inference is typically 10x cheaper per request than cloud APIs for image recognition tasks, while also eliminating latency variability and privacy risks. The table below summarizes the trade-offs.

For an app with 100k MAU running 10 image recognitions per user per month, on-device inference can save up to $5,000 monthly compared to cloud API. Get a free consultation on your ML architecture today.

Integrating OpenAI API and Other Cloud Models

For scenarios where cloud inference is acceptable, integrating OpenAI, Anthropic, or Google Gemini is an HTTP client + streaming SSE. In Swift, AsyncThrowingStream is convenient for streaming responses. In Kotlin, use Flow.

Critically: API keys must never be stored in the app bundle. Even an obfuscated key can be extracted from the IPA in 10 minutes using strings or frida. Correct architecture: mobile app → your own backend → OpenAI API. The backend controls rate limiting, logs requests, and protects the key.

What Is Included in the Work (Deliverables)

• Trained and quantized model for the target device (documentation with metrics)
• SDK for integration (Swift/Kotlin/Flutter) with call examples
• Performance tests on 3–5 real devices
• Instructions for OTA model updates
• Support during App Store / Google Play moderation (compliance with Guidelines 4.2, 5.1)
• 2 weeks of technical support after release

Typical Project Pipeline

1. Task analysis — measure latency, privacy, size, supported devices.
2. Model prototyping — in Python, evaluate accuracy on target data.
3. Conversion and quantization — for CoreML/TFLite with validation.
4. Integration into the app — model wrapped in a service layer (easy to swap CoreML ↔ TFLite ↔ cloud).
5. Testing — on real devices, measure FPS, RAM, battery.
6. Deployment — via TestFlight / Firebase App Distribution, monitor metrics.

Timelines: integration of a ready CoreML/TFLite model — 1–2 weeks, development of a custom model with mobile optimization — from 6 weeks, on-device LLM chat with personalization — 4–8 weeks.

Why We Take on Complex Cases?

10+ years of experience in mobile development, 50+ implemented AI/ML solutions, guarantee of compatibility with current iOS and Android versions. All projects undergo code review and load testing. The cost includes preparation of moderation documentation and training of your team.

Contact us — we will help you choose the architecture and implement ML in your app turnkey. Order an audit of your existing solution — we will assess the potential for server cost savings free of charge. In some projects, savings can reach significant amounts per month.