Effective Date: June 6, 2026

Your privacy is the absolute foundation of our application. GemTranslate ("we," "our," or "the App") is developed as a local-first, privacy-centric AI tool. This Privacy Policy describes how we handle your data, or more accurately, how we ensure your data never leaves your device.

1. Ultimate Privacy Principle: 100% On-Device Processing

GemTranslate operates under a strict architecture: No cloud processing, no remote telemetry, no background analytics, and no remote data logging.

• No Server Storage: We do not own, rent, or operate any external servers to process or store your inputs.
• Local Computing: All AI model execution, deep learning inference, and data transformations are handled 100% locally on your physical iOS device using Apple Silicon GPU (Metal) hardware acceleration.

2. Types of Data and How They Are Handled

Because all features run completely locally, the following types of data are strictly ephemeral or kept securely within your local device sandbox:

• Text and Translation Inputs: Any text you type, paste, or translate is processed locally in temporary volatile memory and is never uploaded anywhere.
• Voice and Audio Recordings: Audio clips recorded for the voice interaction features (sampled locally as PCM/WAV files) are kept entirely on-device. They are used exclusively to process your instant AI command and are never transmitted to any third-party voice-recognition servers.
• Images and Vision Inputs: When you upload images (up to 9 concurrent inputs) for multimodal analysis, the visual files are processed purely via your local on-device vision model weights and are never indexed, tracked, or leaked to remote entities.
• Chat Contexts: Local conversations are managed using iOS native SwiftData structures. This database is stored under your application's protected Application Support directory and is explicitly excluded from iCloud backups to ensure absolute zero cloud trace.

3. Network Connections & Third-Party Outbounds

GemTranslate does not build personal profiles or track behaviors. The application initiates network communication under only two explicitly defined scenarios:

1. Initial Model Download (User-Initiated): Upon the first launch, the App downloads the required underlying AI model weights (~2.3 GB main model + ~812 MB multi-modal extension pack) directly from the official Hugging Face repository mirrors. No user-identifiable data or personal attributes are exposed during this file request.
2. Apple System Crash Reporting: If the application crashes, basic diagnostics are captured and transmitted via Apple's official TestFlight or App Store diagnostic pipelines. This infrastructure is fully managed by iOS system frameworks under Apple's standard developer privacy agreements; our code integrates no third-party telemetry toolkits, trackers, or marketing SDKs.