On-Device AI: Why Your Content Never Leaves Your Phone

The centerpiece of Remain Faithful's privacy model is on-device AI classification. Here's what that means in practice, and why it matters.

The ReplayKit Sandbox

iOS's ReplayKit framework creates a separate extension process for screen recording. This process is sandboxed: it cannot make network requests, cannot access your files, and cannot communicate with the outside world. Its only channel is a shared app group container that connects it to the main RF app.

This architectural constraint means your screen content is physically incapable of being transmitted over the network from the broadcast extension. That's not a policy commitment. It's a technical guarantee built into how iOS works.

Three Layers of On-Device Classification

Layer 1: URL Blocklist + Regex

When a browser is detected, visible URLs are checked against a local blocklist of known adult domains. Visible text is pattern-matched against regex rules for explicit content categories. This is fast, deterministic, and 100% local.

Layer 2: Apple Vision + SensitiveContentAnalysis

Apple provides two relevant frameworks: Vision OCR (which extracts text from screen frames) and SensitiveContentAnalysis (which detects nudity and explicit images). Both run on the device's Neural Engine, the dedicated AI chip in modern iPhones. No server involved.

Layer 3: Local Keyword Classifier

Our open-source keyword classifier assigns weighted scores across content categories based on the OCR output. The weights are tuned for common explicit content patterns without requiring image analysis.

Only if all three layers are uncertain does an anonymized category query reach our cloud classifier. That query contains no screen content.

Why This Architecture Matters

A lot of accountability tools say they're "private." What that usually means is that they've made a policy commitment not to look at your data.

Policy commitments can change. Architectures are harder to change. By building on the ReplayKit sandbox and running classification on-device, we've created a system where the privacy isn't dependent on us being trustworthy. It's dependent on the physics of how iOS processes information.

We think that's a more honest form of privacy.