On-Device AI — Privacy by Architecture

Why On-Device?

Most sales intelligence tools send your meeting audio, transcripts, and deal data to cloud servers for processing. Nutan takes the opposite approach: AI runs on your device. Your data never leaves your computer.

This isn't just a privacy feature — it's an architectural decision that means:

• No audio uploaded — Meeting audio is processed locally and discarded after transcription.
• No transcripts on servers — Transcripts live in your encrypted local database.
• No deal data exposure — Pipeline intelligence is computed on-device.
• Works offline — Core AI features work without an internet connection.
• No per-seat cloud AI costs — You're using your own hardware, not paying for cloud inference per call.

The AI Models

Nutan runs two AI models locally:

Speech-to-Text Model

An on-device speech-to-text model runs on your Mac for real-time meeting transcription:

• Supports dozens of languages
• Processes audio in real time as your meeting progresses
• Uses hardware acceleration on modern Macs for best performance
• Audio is transcribed in memory — raw audio is never written to disk

Language Intelligence Model

An on-device language model runs locally for all intelligence tasks:

• Meeting summarization
• Action item extraction
• Signal detection and deal analysis
• Coaching suggestions during calls
• Chat responses
• Morning briefings
• Pattern detection across meetings

Uses hardware acceleration on modern Macs for fast inference.

Model Management

First-Time Download

AI models are downloaded automatically on first launch. This is a one-time download:

• Models are stored locally on your Mac.
• Download size varies by model — check your internet connection for the initial download.
• Nutan works without the models during download, but AI features are unavailable until complete.

Model Status

Check the status of your AI models in Settings. You'll see:

• Download progress for pending models
• Model version information
• Whether hardware acceleration is active

Updates

Model updates are delivered with app updates. When a new model version is available, it downloads in the background and activates on next launch.

Performance

Newer Macs (Recommended)

On newer Macs, Nutan uses hardware acceleration for AI inference. This means:

• Real-time transcription with minimal system impact
• Fast meeting summaries (generated in seconds, not minutes)
• Responsive chat with low latency
• Coaching suggestions appear within moments of a trigger

Older Macs

Nutan still works on older Macs — AI runs without hardware acceleration. Expect:

• Slightly delayed transcription (may lag a few seconds behind the conversation)
• Longer summary generation times
• Higher system load during meetings

For the best experience, we recommend a Mac from the last four years.

What Still Uses the Cloud

While AI inference is local, a few features require internet:

Meeting capture, transcription, coaching, deal analysis, chat, and all AI intelligence features work entirely offline after initial setup.