AI Workout Assistant for Mobile Fitness Apps

The problem: why static plans fail

A user downloads a fitness app, answers a couple of questions, and gets a template like "3×10 squats." But they have a knee injury, no barbell at home, and yesterday they trained legs. The result is pain, frustration, and app deletion. Static plans are the main reason for low retention: without context, users quit within a week. We solve this by building an AI assistant that considers real data: workout history, available equipment, physical limitations, and goals. It doesn't give generic advice—it adapts to each person in real time. In one project, integrating such an assistant boosted 30-day retention by 40% compared to static plans.

Why an AI assistant outperforms static plans

Static programs ignore context. An AI assistant, on the other hand, collects data from multiple sources: the user profile (anthropometry, goals, injuries), HealthKit (iOS) or Health Connect (Android), and completed workout history. It selects exercises three times more accurately than manual selection (based on our internal A/B tests). It also reduces overtraining risk by automatically avoiding muscle groups loaded in the last 48 hours. For over five years we have been building fitness apps and have deployed such solutions in 30+ projects. We guarantee compatibility with the latest iOS and Android versions.

How we collect data for personalization

Recommendation quality depends directly on profile completeness. The minimum required fields:

From native sources we add HealthKit (iOS) or Health Connect (Android 14+) data:

// iOS: load workouts from the last week via HealthKit
func fetchRecentWorkouts() async throws -> [HKWorkout] {
    let workoutType = HKObjectType.workoutType()
    let predicate = HKQuery.predicateForSamples(
        withStart: Calendar.current.date(byAdding: .day, value: -7, to: Date())!,
        end: Date()
    )
    let sortDescriptor = NSSortDescriptor(key: HKSampleSortIdentifierStartDate, ascending: false)

    return try await withCheckedThrowingContinuation { continuation in
        let query = HKSampleQuery(
            sampleType: workoutType,
            predicate: predicate,
            limit: 20,
            sortDescriptors: [sortDescriptor]
        ) { _, samples, error in
            if let error { continuation.resume(throwing: error); return }
            continuation.resume(returning: (samples as? [HKWorkout]) ?? [])
        }
        healthStore.execute(query)
    }
}

Apple's HealthKit documentation confirms the correctness of this data extraction approach. We also import sleep parameters and heart rate variability (HRV)—they are critical for assessing recovery.

Generating a workout plan

The prompt with full context:

func buildWorkoutPrompt(profile: UserProfile, recentWorkouts: [WorkoutSummary]) -> String {
    let workoutHistory = recentWorkouts.map {
        "\($0.date.formatted()): \($0.type), \($0.duration) min, \($0.muscleGroups.joined(separator: "+"))"
    }.joined(separator: "; ")

    return """
    Create a workout plan for today.

    User profile:
    - Goal: \(profile.goal)
    - Level: \(profile.level)
    - Available time: \(profile.availableMinutes) minutes
    - Equipment: \(profile.equipment.joined(separator: ", "))
    - Limitations: \(profile.limitations.isEmpty ? "none" : profile.limitations.joined(separator: ", "))

    Recent workouts (last 7 days): \(workoutHistory.isEmpty ? "none" : workoutHistory)

    Rules:
    - Avoid muscle groups trained in last 48 hours
    - If limitation mentions specific area (knee, back), exclude exercises for that area
    - Balance push/pull if goal is hypertrophy

    Return JSON: {
      name, totalMinutes, exercises: [{
        name, sets, reps, restSeconds, muscleGroups: [], notes, videoSearchQuery
      }]
    }
    """
}

The videoSearchQuery field is key. We use it to form a YouTube/Vimeo search query to show a demo of the exercise right in the card.

Adapting during the workout

The assistant should not stay silent after delivering the plan. Three triggers for adaptation:

1. User taps "Too hard" → LLM replaces the exercise with an easier one.
2. Time elapsed exceeds planned → offers to reduce remaining sets.
3. All exercises completed 15 minutes early → adds a bonus block.

// Android - exercise substitution
suspend fun substituteExercise(
    exercise: Exercise,
    reason: SubstitutionReason,
    availableEquipment: List<String>
): Exercise {
    val prompt = """
    The user cannot do "${exercise.name}".
    Reason: ${reason.description}
    Available equipment: ${availableEquipment.joinToString(", ")}
    Target muscles: ${exercise.muscleGroups.joinToString(", ")}

    Suggest ONE simpler/alternative exercise that targets the same muscles.
    Return JSON: {name, sets, reps, restSeconds, muscleGroups: [], notes}
    """

    val response = openAIClient.chat(
        model = "gpt-4o-mini",
        messages = listOf(Message("user", prompt)),
        responseFormat = ResponseFormat.JsonObject
    )
    return json.decodeFromString(response.content)
}

How the AI assistant determines rest days

The assistant should recommend not training when needed. If HealthKit shows a drop in HKQuantityType.heartRateVariability or sleep under 6 hours, the plan shifts to stretching or active recovery.

func shouldRecommendRestDay(healthData: HealthSnapshot) -> Bool {
    return healthData.sleepHours < 6.0 ||
           healthData.restingHeartRate > healthData.averageRestingHR * 1.1 ||
           healthData.hrvTrend == .declining
}

This is not an AI query—a simple heuristic from native data. No LLM needed here.

How to set up a profile for accurate recommendations

To get the most relevant plans, the user must enter their data correctly. For example, if the goal is weight loss and the level is advanced, the assistant will generate a high-intensity program focused on calorie burn. If the user has a knee injury, all squats and lunges will be excluded. We recommend filling all fields including available equipment—this improves selection accuracy by 50%.

What's included in development

• Architecture: HealthKit/Health Connect synchronization, profile storage, workout cache.
• Plan generation: prompt engineering, JSON response validation, LLM error handling.
• Adaptation: exercise substitution triggers, rest timer, rule-based adjustments.
• UI/UX: profile screen, workout card, video demonstrations, "Too hard" button.
• Testing: unit tests, A/B conversion test (targeting 20% retention improvement), TestFlight beta.
• Deployment: App Store and Google Play publishing, push notification setup (APNs/FCM).
• Documentation: data schema, heuristic descriptions, prompt update instructions.

In addition, we audit your current app and propose the optimal integration architecture. Reach out for a consultation—we'll show how an AI assistant can boost user engagement.

Estimated timeline

Pricing is determined individually based on integration complexity and required functionality. Our engineers hold Apple and Google certifications and have hands-on experience with HealthKit and Health Connect. Contact us to discuss your project—receive a consultation on AI assistant integration today. Get a cost estimate for your project—submit a request.

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.