AI-Powered Mobile App for Car Repair Cost Estimation from Photos

A car owner photographs a dent with a smartphone and wants to know how much the repair will cost — without visiting the service center. Traditional assessment requires an expert inspection and takes an hour, while calculation errors lead to disputes with the insurance company. Our solution reduces assessment time to seconds and minimizes human error. The task is not just to detect the defect but to calculate the real cost considering region, car make, and current parts prices. We develop such a pipeline turnkey: from detection to an estimate in PDF format. Contact us — we will assess your project timeline to MVP in 2–3 weeks.

How the assessment pipeline works

Cost estimation is not a direct neural network output. It is a multi-step deterministic calculation:

Photo → Damage Detection → [type, severity, area, localization]
                                          ↓
                             Repair normative database
                             (labor-hours per operation type)
                                          ↓
                             Regional rates
                             (cost per labor-hour)
                                          ↓
                             Parts cost
                             (OEM vs aftermarket)
                                          ↓
                             Final estimate [min, expected, max]

The neural network is used only in the first step — everything else is database calculations. We use YOLOv8 for detection, fine-tuned on a dataset of 50,000+ damage photos.

Normative repair databases

In Russia, the de facto standard is Audatex (now Solera) and GT Motive. They contain labor-hours for each operation for each vehicle: replacing a front fender on a Toyota Camry XV70 — 2.4 n/h, straightening — 1.8 n/h, painting — 3.2 n/h. API access is through a partnership agreement.

For an MVP without expensive licenses, we use open databases such as Mitchell1, AllData (USA), or our own database built from public STO price lists via scraping and normalization. Our experience: 8+ years in mobile development, 30+ projects for auto insurance. We guarantee accuracy within 10–15% deviation and offer a 3-month post-launch support period to fine-tune results.

// iOS: cost calculation request
struct DamageCostRequest: Codable {
    let vehicleInfo: VehicleInfo           // make, model, year
    let damageDetections: [DamageDetection]
    let repairLocation: RepairLocation     // city, country
    let repairType: RepairType             // dealer, certified, independent
}

struct VehicleInfo: Codable {
    let make: String
    let model: String
    let year: Int
    let bodyType: BodyType
    let vin: String?
}

// Response contains breakdown by line items
struct DamageCostEstimate: Codable {
    let lineItems: [CostLineItem]
    let laborCost: MoneyAmount
    let partsCost: MoneyAmount
    let paintCost: MoneyAmount
    let totalMin: MoneyAmount
    let totalExpected: MoneyAmount
    let totalMax: MoneyAmount
    let currency: String
    let validUntilDate: Date           // rates change
    let disclaimer: String
}

struct CostLineItem: Codable {
    let description: String            // "Replace front bumper"
    let damageType: DamageType
    let panelLocation: PanelLocation   // front_bumper, hood, etc.
    let laborHours: Double
    let laborCostPerHour: MoneyAmount
    let partsCost: MoneyAmount?
    let repairVsReplaceRecommendation: RepairOption
}

Why repair vs replace is a key cost factor

For each damage, the system recommends: repair (straightening + painting) or part replacement. Simplified rule:

func recommendRepairOption(
    damage: DamageDetection,
    panel: PanelInfo
) -> RepairOption {
    let damageAreaRatio = damage.maskAreaPx / panel.totalAreaPx

    // If damage > 40% of part area → replace, not repair
    if damageAreaRatio > 0.4 { return .replace }

    // Structural damage → replace only
    if damage.severity == .structural { return .replace }

    // If repair costs more than 70% of new part → replace
    let repairEstimate = calculateRepairCost(damage, panel)
    let partPrice = panel.partPrice.aftermarket
    if repairEstimate > partPrice * 0.7 { return .replace }

    return .repair
}

AI assessment is 5 times faster than manual expertise and eliminates human error in calculating labor costs. Economies of scale — up to 60% budget savings on damage assessment. A typical project reduces costs by several hundred thousand rubles. For example, a fleet operator with 10,000 claims per year can save over $500,000 annually.

What if the car is not in the normative database?

If the model or body is not found in Audatex/GT Motive, the system uses the closest analog with a correction factor. For rare cars, we incorporate calculation based on damage photos and expert assessment. Alternatively, the client can upload their own rates via CSV or API. The architecture is flexible enough to adapt to any make.

Comparison of assessment methods

Comparison of implementation timelines

UI: clear output for non-technical audience

An insurance agent or car owner does not read "ASDA-norm 2.4 n/h". The interface shows a clear breakdown:

@Composable
fun CostEstimateScreen(estimate: DamageCostEstimate) {
    Column(modifier = Modifier.padding(16.dp)) {

        // Total amount — large and first
        TotalCostBanner(
            min = estimate.totalMin,
            expected = estimate.totalExpected,
            max = estimate.totalMax
        )

        Spacer(Modifier.height(24.dp))

        // Breakdown by expense items
        SectionHeader("Expense Breakdown")
        CostBreakdownBar(
            labor = estimate.laborCost,
            parts = estimate.partsCost,
            paint = estimate.paintCost
        )

        Spacer(Modifier.height(16.dp))

        // Damages and work items
        SectionHeader("Damages and Works")
        estimate.lineItems.forEach { item ->
            DamageLineItemCard(item = item)
        }

        // Disclaimer — mandatory
        DisclaimerText(text = estimate.disclaimer)
    }
}

The min–expected–max range is more honest than a precise figure. A precise figure creates false expectations and leads to conflicts when the actual service bill arrives.

How is assessment accuracy ensured?

Accuracy is achieved through a combination of three factors: detection quality (YOLOv8 fine-tuned on 50k+ photos), normative database timeliness (monthly updates), and regional coefficients. We also use currency exchange and inflation data for adjustments. On average, deviation does not exceed 15%, and with a full normative database, 10%. Contact us — we will demonstrate accuracy on your data. We certify that our system meets industry standards and provide a performance guarantee.

What is included in the work

• API documentation (Swagger/OpenAPI)
• Mobile SDKs for iOS (Swift) and Android (Kotlin)
• Test environment with 1000+ test photos
• Client team training (2 days)
• 3 months post-launch support
• Access to our cloud instance for alpha testing
• Compliance certificates for data security

Timeline and cost guidelines

Backend with damage detection (YOLOv8) + cost calculation using a fixed regional database + mobile client — 2–3 weeks. Full system with Audatex/GT Motive integration, regional rates, current parts prices, PDF report export and signature — 1–3 months. Contact us for a consultation on your project. Budget is calculated individually based on requirements. Typical project cost starts at $50,000 for MVP and scales up.

Typical mistakes and how to avoid them

• Relying solely on AI without a normative database. This yields 50–70% accuracy and unreliable estimates. Always combine AI detection with a deterministic cost calculation.
• Ignoring regional rate variations. Using national averages can misprice repairs by 30% or more. Always apply local labor and parts cost multipliers.
• Forgetting about repair vs replace logic. Without it, the system may recommend expensive repairs when replacement is cheaper. Our rule-based engine handles this automatically.

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.