AI Tutor Offline

• 

  Demo

Inspiration

Many students in remote or low-bandwidth areas lose access to timely tutoring, translation, and study aids because cloud-based AI requires reliable internet and can leak sensitive data. We wanted a fast, private alternative that runs entirely in the browser so learners can get AI help anytime — even offline — without cloud costs or data exposure.

What it does

This is a privacy-first Chrome extension that runs compact on-device models to provide instant AI chat tutoring, automatic lesson summarization, and real-time translation directly in the browser. It extracts text from webpages and PDFs, generates concise study notes, answers student questions in plain language, and works fully offline after initial setup.

How we built it

It's implemented as a Manifest V3 Chrome extension (TypeScript + React UI) with a lightweight inference runtime using WebAssembly/WebNN-compatible binaries for on-device models. We store models and assets in IndexedDB and use a service worker for background tasks and offline caching. The UI integrates an overlay chat widget and context-aware summarizer that scrapes and preprocesses page content client-side. CI builds produce compressed, quantized model blobs to balance size and latency.

Challenges we ran into

Running useful models inside the browser demanded careful trade-offs between model size, latency, and accuracy. We tackled model quantization and streaming inference to reduce memory and CPU spikes, implemented progressive loading so the extension is usable before full model download, and handled a wide variety of webpage structures for reliable text extraction. Packaging, permissions, and offline persistence across Chrome versions also required extra edge-case handling.

Accomplishments that we're proud of

• Achieved usable on-device chat and summarization in typical laptops and Chromebooks with no server round-trip.
• Built robust offline-first flows (model download, resumable caching, and fallback behaviors).
• Polished UX: context-aware chat overlay, one-click summarization, and inline translations that work without network access.
• Created an open-source repo and CI pipeline that reproducibly builds quantized model artifacts for browser inference.

What we learned

I learned that small model accuracy can be dramatically improved with smart prompting, contextual retrieval from the current page, and light post-processing heuristics. Progressive delivery (tiny core model first, larger models later) is key to a good user experience. Browser constraints force pragmatic engineering: measure memory, prefer streaming, and handle diverse real-world content formats.

What's next for it

Priorities: broaden language support and curriculum-specific prompts, add a teacher dashboard for batch distribution of materials, and optimize model quantization for faster low-end device performance. I also plan mobile-portable builds (Edge/Firefox support and a lightweight PWA), optional encrypted sync for classroom settings, and experiments with federated improvements that keep user data private.

Built With

• chrome
• eslint
• manifest
• node.js
• react
• tailwind
• typescript