Technical Architecture

The Engine: WhisperKit & CoreML

At the heart of CipherScript is OpenAI's Whisper model, ported to Apple's CoreML framework via WhisperKit. This allows us to tap directly into the Apple Neural Engine (ANE) on Apple Silicon chips.

• Quantization: Models are quantized to run efficiently in memory without sacrificing significant accuracy.
• Batch Processing: Audio is processed in chunks to maintain low latency and high throughput.
• Energy Efficiency: By using the ANE, transcription consumes significantly less battery than CPU or GPU-based inference.

Zero Network Entitlements

In macOS app development, "Entitlements" are permissions an app requests from the OS. Most apps request com.apple.security.network.client to access the internet.

CipherScript does not have this entitlement.

This means the macOS kernel strictly blocks any attempt by the app to open a network socket. Even if the app wanted to "phone home," the operating system would terminate the connection immediately. This provides a cryptographic guarantee of privacy that goes beyond trust.

Security Model

Since the app is sandboxed and network-gapped:

• No data can be intercepted in transit (because there is no transit).
• No data can be stored on our servers (because we have no access).
• Your transcripts exist only in the temporary memory during processing and on your local disk where you save them.