Does the AI rewriting actually run on-device? The site mentions local storage but I wasn't sure if that implies local inference for the processing modes too. If it is local, curious what models you're using to handle the structured output without killing the battery.