AI Video Moderation for Mobile Apps: On-Device and Cloud Solutions

Implementing automatic video moderation in mobile apps

A user uploads a video — and you have seconds to decide whether to show it to others. Manual review doesn't scale: moderators get tired, miss violations, and a 10-minute queue kills retention. Our team of mobile engineers with experience in AI moderation has delivered 40+ projects for FinTech, EdTech, and Social platforms. We specialize in integrating on-device models and cloud APIs for automatic video content classification. For effective video moderation in mobile apps, combining on-device AI filtering with cloud-based analysis is key. Below, we break down how to avoid common mistakes and build a reliable system.

Where problems most often arise

Real-time moderation vs. post-processing

The most common architectural mistake is trying to run a model frame-by-frame on the client. CoreML on an iPhone 14 Pro can handle MobileNet v3 at 30 fps for short clips, but it kills battery and overheats the device. On Android, the situation is similar with MediaPipe: processing every frame in ImageAnalysis.Analyzer at 1080p leads to ImageProxy backlogs and crashes with java.lang.IllegalStateException: Image is already closed.

The right approach for video is not frame-by-frame analysis but selective sampling: every N frames or key scenes using AVAssetImageGenerator (iOS) / MediaMetadataRetriever.getFrameAtTime() (Android). For most moderation tasks, 1 frame per second is sufficient.

Server-side moderation via Video Intelligence API

For UGC video apps, we design the following scheme: the client uploads video to storage (S3/GCS), triggers a Cloud Function that calls Google Cloud Video Intelligence with EXPLICIT_CONTENT and OBJECT_TRACKING features. The response is JSON with timestamps and confidence scores per segment.

// Android: start upload and pass URI to backend
val uploadRef = storageRef.child("uploads/${UUID.randomUUID()}.mp4")
uploadRef.putFile(localUri)
    .addOnSuccessListener { taskSnapshot ->
        taskSnapshot.storage.downloadUrl.addOnSuccessListener { downloadUri ->
            moderationApi.submitVideo(downloadUri.toString(), onComplete = { result ->
                when (result.verdict) {
                    ModerationVerdict.SAFE -> publishVideo()
                    ModerationVerdict.UNSAFE -> rejectWithReason(result.reason)
                    ModerationVerdict.REVIEW -> sendToHumanReview()
                }
            })
        }
    }

AWS Rekognition Video is an alternative with a similar API: StartContentModeration + polling via GetContentModeration. For synchronous cases (short reels up to 30 sec), Rekognition Image applied to extracted frames works — response in 200–400 ms. Cost for processing a minute of video via Google Video Intelligence starts from $0.15 per minute.

On-device pre-filtering

Before sending to the server, it makes sense to run the first and last frames of the video through a local CoreML/TFLite model. This catches obvious NSFW on the client, saving traffic. NudeNet Lite in TFLite format is about 14 MB and delivers ~92% accuracy on NSFW benchmarks. False positives on medical content are a separate issue and require whitelist logic at the app category level.

How on-device pre-filtering works

On-device pre-filtering is applied to extracted key frames (first and last) via CoreML (iOS) or TFLite (Android). The model is NudeNet Lite, 14 MB, 92% accuracy on NSFW datasets. False positives (medical content) are handled by whitelist logic. On-device analysis is 3x faster than a server roundtrip for obvious violations and saves up to 30% of traffic. We use Swift Combine on iOS for reactive upload progress, and Kotlin Coroutines with Flow on Android for similar functionality.

Why server-side moderation is preferable for UGC

Server-side moderation via cloud APIs (Google Video Intelligence, AWS Rekognition) wins in accuracy and scalability. On-device models are limited by compute resources and support only basic classes (NSFW, violence). For complex moderation (weapon detection, contextual violations), a server-side model with larger context is required. In practice, we combine: fast on-device filter + deep cloud analysis for borderline cases.

Comparison of solutions

Typical scenarios and recommended solutions

How we build the solution

The stack depends on latency requirements and budget. For startups with low traffic — Google Video Intelligence API: pay $0.15 per minute, no need to spin up infrastructure. For high-load platforms — custom inference service based on CLIP or a custom ONNX model behind a reverse proxy with caching of already-checked video hashes (perceptual hashing via pHash prevents re-moderation of the same clip).

On the client side (iOS/Android/Flutter), we implement:

• upload progress bar with URLSession.uploadTask / okhttp3.MultipartBody
• pending state for videos in the feed ("under review")
• push notification of result via FCM/APNs

A separate case is live streaming. Video Intelligence API is not suitable here due to latency. We use streaming via WebRTC + server-side analysis of HLS segments every 2–4 seconds with a speed-optimized model (MobileViT-S in TorchScript).

What is included in the work (deliverables)?

• Architectural decision records (ADR)
• SDK integration for upload and moderation
• UI states (pending, approved, rejected)
• Testing on 100+ edge cases
• Team lead training on administration
• 6-month code warranty

Process

1. Requirements audit: content type (UGC, Stories, live), acceptable publication delay, compliance requirements (GDPR, COPPA).
2. Stack selection: on-device pre-filter + cloud moderation vs. fully server-side.
3. Development: integrate upload SDK, webhook/polling for results, status UI.
4. Testing on edge-case dataset: multilingual subtitles in frame, medical content, animation.

Time estimates

Integration with Google Video Intelligence or AWS Rekognition Video — 3–5 days. Adding on-device pre-filter on CoreML/TFLite — another 2–3 days. Full solution with live streaming and human review system — 3–4 weeks.

For more information, we can prepare a proposal tailored to your stack and load. A consultation can help discuss details.

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.