Implementing AI Toxicity Detection in Mobile Apps

We have encountered projects where hate speech remained undetected due to language-specific cultural nuances. Spam is easy to catch with patterns, but toxic content is unique—written by a real person and often grammatically correct. Detection is complicated by letter substitutions (leetspeak) and veiled insults. We have built a turnkey solution that handles this challenge.

Why Standard Toxicity Detection Models Fail for Russian-Language Apps

General toxicity models like unitary/toxic-bert work well on English Reddit datasets. In a Russian-language app, they produce false positives on culturally specific words and miss masked profanity with letter substitution—a common filter evasion tactic in the CIS audience. The same issue applies to Ukrainian and Belarusian text. Another pitfall is synchronous model inference on message send. The user clicks “send” and waits 800 ms—UX broken. Detection must be either asynchronous post-processing or fast enough to be imperceptible.

Multi-Level Architecture for Toxicity Detection: On-Device + Server

We use two classification levels. First is on-device and fast: a regex + dictionary with 2,000 obvious toxic patterns, including leetspeak variants. It processes in <5 ms, requires no network, and catches 60–65% of toxic messages with minimal false positives. Second is server-side ML: a fine-tuned model on a Russian-language dataset (RuToxic or similar from Hugging Face). It is called asynchronously after message display—if triggered, the message is hidden and replaced with a placeholder. The on-device filter is 100x faster than the server (5 ms vs. 200 ms) and handles up to 65% of toxic messages without network load.

// Android: optimistic send + async toxicity check
fun sendMessage(text: String) {
    val tempMessage = Message(text = text, status = MessageStatus.PENDING_REVIEW)
    chatAdapter.addMessage(tempMessage)  // show immediately

    viewModelScope.launch {
        val result = toxicityRepository.classify(text)
        if (result.isToxic && result.confidence > 0.78f) {
            chatAdapter.updateMessageStatus(tempMessage.id, MessageStatus.HIDDEN)
            showToxicityNotice()
        } else {
            chatAdapter.updateMessageStatus(tempMessage.id, MessageStatus.VISIBLE)
        }
    }

    messageApi.send(tempMessage)
}

This optimistic UI + post-publication check eliminates latency. The user sees the message instantly while the check runs in parallel.

Multilingual Support via xlm-roberta-base

For apps with audiences in multiple countries, we use xlm-roberta-base fine-tuned on a mixed dataset. The model in ONNX format is deployed behind a FastAPI endpoint. Important: inference must be batched under high traffic—onnxruntime supports dynamic batching, providing ~4x throughput compared to sequential processing.

iOS: Core ML for Pre-Filter

On iOS, the pre-filter is conveniently implemented via Core ML with a Text Classifier converted using coremltools:

let request = NLModel(mlModel: toxicityModel.model)
let prediction = request.predictedLabel(for: text) ?? "safe"
let confidence = request.predictedLabelHypotheses(for: text, maximumCount: 2)

if prediction == "toxic", let score = confidence["toxic"], score > 0.9 {
    return .block
}

NaturalLanguage.framework with a custom NLModel is the cleanest path for iOS, requiring no third-party dependencies in the build.

How to Fine-Tune a Toxicity Detection Model for Your Dataset

First, we collect historical user reports and annotate them via Label Studio or Toloka—up to 10,000 examples. Then we fine-tune a base model (e.g., xlm-roberta-base) on domain-specific data. Next, deploy the inference API and integrate into mobile clients. The final step is tuning thresholds based on precision/recall tradeoff to match product requirements. Monitoring includes the share of auto-blocked messages and false positive rate from user complaints.

Here is a comparison of the two levels:

How to Set Trigger Thresholds: Step-by-Step Guide

1. Collect an annotated dataset of 1,000–2,000 messages labeled by toxicity categories.
2. Build an ROC curve for each category on a held-out set.
3. Choose an auto-block threshold and a human-review threshold based on desired false positive rate (typically 1–5%).
4. Set thresholds in the system configuration and start monitoring.

Apple Core ML documentation: "Core ML models can be updated on device without sending data to the server, preserving user privacy and reducing latency."

What the AI Toxicity Detection Work Includes (Deliverables)

• Dataset collection and annotation (up to 10,000 examples)
• Model fine-tuning (BERT-based or XLM-RoBERTa)
• Inference API deployment (FastAPI + ONNX)
• iOS (Core ML) and Android (ML Kit or custom) integration
• Threshold tuning and monitoring
• Operations documentation
• Access to the model and API
• Training for your team and ongoing support

Timeline Estimates

Basic integration of a ready multilingual model: 4–6 days. Fine-tuning on your own dataset and deployment: an additional 2–3 weeks. Full system with categorization, human review queue, and feedback loop: 4–6 weeks.

Our Experience and Advantages

With over 5 years in the mobile app market and a proven track record, we have implemented toxicity detection for three projects with audiences exceeding 1 million users each. Our certified models and guaranteed performance ensure savings on manual moderation reach 60–80%, and the solution pays for itself within 2–3 months. A pilot project starts from $2,500. Clients report saving an average of $10,000 per month after deployment.

Company metrics: 5+ years of experience, 3+ successful implementations, each handling 1M+ users.

Get a free assessment of your project—contact us for a consultation. Order a pilot project to test the solution on your data with a 30-day money-back guarantee.

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.