Research: Open Source Scaffolds for On-Device AI in Android

[rainbowpuffpuff]

This is a sub-task for Issue #9. We need to find a suitable open-source Android application or SDK that can serve as a scaffold for our on-device AI processing.

Requirements:

• Must be developed or heavily supported by Google.
• Must support offline model downloading and management.
• Must be able to process local device data (images, sensors, etc.).

Initial Research Findings:

The Google AI Edge platform seems to be the most relevant ecosystem. Here are the key components to investigate further:

• Google AI Edge Gallery: An open-source Android app on GitHub that serves as a working example and scaffold for on-device AI. It uses the LiteRT (formerly TensorFlow Lite) runtime and demonstrates various use cases. This is a strong primary candidate.
• Gemini Nano: Google's most efficient model for on-device tasks, accessible via the Google AI Edge SDK. This is likely the model we would use.
• MediaPipe: A library of pre-built, customizable AI solutions for vision, text, and audio. This could accelerate development if our use case fits one of its solutions.
• Android AI Sample Catalog: A standalone app with self-contained examples of Google's AI models.

Next Steps:

1. Clone and evaluate the Google AI Edge Gallery application.
2. Investigate the Android AI Sample Catalog for relevant examples.
3. Determine the best way to integrate Gemini Nano for our specific use case.