• 

  This is what the extension looks like.

• 

  It can translate the text to languages to help the lecturers.

• 

  It can help write learning objectives from a text

• 

  It can help create lecture outlines from a webpage

• 

  It can create quizes such as MCQs

• 

  It can also create short question from the webpages

About the Project — Lecture Co-Pilot

Inspiration

Teachers told us the same thing: “I’ve got the content, but turning it into objectives, outlines, notes, and quizzes eats my prep time.” Chrome’s built-in, on-device AI (Gemini Nano) promised private, fast summarization and translation right in the browser, so we set out to turn any web page into ready-to-teach assets with one click.

What we built

Lecture Co-Pilot is a Chrome extension that:

• Reads a page’s rendered text and produces key-point summaries via the Summarizer API (type, length, format).
• (Optional) Uses the Prompt API for structured JSON outputs—objectives, outline, quiz, and notes—when enabled for extensions.
• Translates summaries/notes with the Translator API; pairs nicely with Language Detector for auto source-language. All inference happens on device after a one-time model download, keeping class material local.

How we built it

1. MV3 plumbing: a content script extracts innerText; the popup orchestrates actions; an offscreen page is available for any DOM-dependent cleanup so the service worker stays light.
2. Messaging: popup ↔ content script via tabs.sendMessage; internal hops via runtime.sendMessage.
3. Summarizer flow: availability() → create({ monitor }) to display download progress on first run → summarize() with type/length/format.
4. Prompt flow (optional): create a model session with initial prompts and request JSON-schema-constrained output for teaching artifacts.
5. Translate flow: Translator.create() → translate() for bilingual outputs; Language Detector can auto-detect source language.

What we learned

• On-device UX matters: the first-run model download needs clear progress UI; Chrome exposes a monitor callback on create() for that.
• Rendered text > raw HTML for summarization; it naturally drops navigation/boilerplate and improves quality.
• API status differs: Summarizer/Translator/Language Detector are stable in Chrome 138+, while Prompt/Writer/Rewriter/Proofreader may be origin-trial or early-preview—so we made those modules optional

Challenges

• Model readiness & device variability: handling unavailable/downloadable states, showing progress, and keeping the UI responsive.
• Injection/messaging pitfalls: avoiding the classic “receiving end does not exist” by guaranteeing content-script presence before messaging.
• Scope control: staying within input budgets; chunking long pages without losing coherence.

A tiny bit of math 🧮

To stay within the model’s effective input window, we chunk long pages before summarizing. If (L) is the character length and the chunk size is (c) (≈ 3.5–4k chars), the number of chunks is [ n=\left\lceil \frac{L}{c} \right\rceil, ] then we summarize each chunk and run a final key-points pass over the concatenated summaries. (This mirrors the docs’ guidance to keep inputs short for best results.)

What’s next

• Schema-first Prompt sessions for richer quizzes (MCQ distractors, short-answer rubrics).
• Detector + Translator for fully automatic bilingual outputs.
• Better extraction on tricky sites (shadow DOM, virtualized lists) and PDF viewers.

Built With

• chrome-extensions-(manifest-v3)
• css
• dompurify
• html
• javascript-(es-modules)
• language-detector-api
• marked-(markdown)
• offscreen-api
• optional)-vite/rollup-bundler
• prompt-api-(extensions)
• runtime-messaging-api
• scripting-api
• service-workers
• storage-api-(chrome.storage.session)
• summarizer-api-(built-in-ai)
• tabs-api
• translator-api