QuickMind

Smarter browsing, powered by on-device AI.

Comment

💡 The Genesis of QuickMind: On-Device AI for Effortless Productivity

Inspiration

The project was born from the friction inherent in using modern cloud-based AI. Productivity is constantly disrupted by the multi-step process required: copying text, opening a new tab, pasting, prompting, and waiting for a distant server response. More critically, this process compromises user privacy by sending sensitive data to the cloud.

My goal was to create a productivity tool that was instant, private, and prompt-free. The solution was to leverage Google's built-in on-device AI (Gemini Nano), shifting processing locally to deliver immediate results while ensuring zero data leaves the device.

What it does

QuickMind is a Chrome Extension that transforms any highlighted text on the web into a condensed summary or a real-time translation—all with a single right-click.

  • Instant Summarization: Summarizes articles and long texts.
  • Multilingual Translation: Provides fast, on-device translation between languages.
  • Privacy Guaranteed: All functions run 100% locally using Chrome's native AI APIs.

How we built it

QuickMind is a standard Chrome Extension structured with:

  1. A Service Worker (background.js): Handles all communication with the underlying AI APIs.
  2. A Content Script (content.js): Manages the user interface, injecting the result dialog directly onto the active webpage.

The core functionality relies on two experimental native Chrome APIs: the Summarizer and the Translator. The architecture enforces a strict on-device only policy, with no cloud fallback.

Challenges we ran into

Category Challenge Solution Implemented
Instability Risk: Service Worker Lifecycle The greatest instability risk was the "long-running operation" of await Summarizer.create(). Since initialization can take several seconds, Chrome's lifecycle manager often killed the Service Worker before the model was ready. Implemented a defensive initialization pattern using global state flags and a polling promise. This ensures only one creation process runs at a time, protecting the worker from concurrent calls and premature death.
Token Limit & Long Context The initial summarizer failed on long articles because the on-device API has limited input tokens ($\approx 6,144$ tokens, or $\approx 20,000$ characters for Gemini Nano). Developed the Summary of Summaries technique (a recursive Map-Reduce algorithm) to chunk the text and progressively reduce it, enabling summarization of full-length articles.
Service Worker and Dynamic State Switching between target languages created persistent errors. The Service Worker architecture caused the Translator object's state to be unstable, complicating instance management. The fix involved careful state tracking and safely resetting the instance to null before creating a new language pair object, working around the API's lack of a dispose() method.

Accomplishments that we're proud of

  • Overcoming Token Limits: Successfully implemented the Summary of Summaries Map-Reduce technique, a complex architectural fix that enables the extension to handle long documents where the underlying API would normally fail.
  • Lifecycle Stability: Engineered a robust, defensive pattern to prevent the Service Worker from being killed during model initialization, ensuring high reliability for the core features.
  • Guaranteed Privacy: Delivering a fully functional, AI-powered tool with the fundamental guarantee that no user data ever leaves their device.
  • Translator Speed: Achieving near-instantaneous translation, showcasing the immense speed advantage of the on-device model over cloud alternatives.

What we learned

  1. On-Device Constraints: We learned that while on-device AI is fast and private, it imposes strict constraints (token limits, hardware requirements) that necessitate creative algorithmic solutions like Map-Reduce.
  2. API Nuances: Mastering the Chrome Service Worker lifecycle is critical for long asynchronous tasks. We had to build custom solutions to manage object state and prevent worker termination.
  3. Performance Trade-offs: Our custom Map-Reduce Summarizer is slower than Chrome's built-in version due to the overhead, but our Translator is exceptionally fast. This highlighted the differences in model optimization within the on-device ecosystem.
  4. Value Proposition: Even as Chrome adopts similar native features, QuickMind's value remains in providing a customizable, prompt-free, and privacy-guaranteed user experience.

What's next for QuickMind

  • 🔍 Add more features: Implement more features like proofreading, writing assistance, etc.
  • 🗂️ Local History: Introduce a feature to save a local history of processed results for easy reference.
  • 🌈 UI Refinements: Continuous improvements to the user interface for even smoother and more intuitive interaction.
  • 💬 Expanded Support: Add more language options and customizable tone settings for output.

Built With

Share this project:

Updates