WebVision – A Multi-Model AI Approach to Privacy-Preserving Web Accessibility

Abstract-This paper adheres to web accessibility through a privacy-centric, AI-powered approach via an extension in Chrome. The extension implements a multi-model architecture that combines Google Chrome’s built-in AI capabilities using Gemini Nano with a JavaScript library transformer.js to process machine learning (ML) models directly in the browser that is web content run locally on users’ devices. Unlike existing solutions that rely on cloud processing or limited built-in browser features which might hinder the user’s privacy, our extension prioritizes user privacy by performing the computational/processing tasks on-device while providing comprehensive accessibility features. System also has voice commands for hands free navigation, generates summaries based on prompts and utilizes moondream(AI model) to provide detailed descriptions of images present in the web-content. Performance metrics indicate that the local processing approach maintains robust functionality while preserving user privacy.Our user testing shows remarkable improvements in web browsing for people with diverse accessibility needs. Users reported faster navigation, better understanding of content, and greater independence compared to traditional screen readers and similar tools. Our approach of processing information locally on users’ devices maintains strong performance while protecting privacy. This research advances accessible technology by showing how AI models can be integrated into browser extensions to make the web more inclusive without compromising privacy or requiring powerful computers.