buzzin

Inspiration

Students get lost in lectures silently. Only about 35% of key lecture points make it into students' notes (Kiewra 1985, in Luo, Kiewra & Samuelson 2016). Attention starts to slip about 30 seconds into class, and the lapses come in ever-shorter cycles (Bunce, Flens & Neiles 2010). And 47.7% of students in large science classes never ask a single question (Nadile et al. 2021, 417 students). The professor can't see who is lost until the exam.

This also matters for students with disabilities. About 1 in 5 undergraduates report a disability (20.5%, NCES 2019–20), and 44% of students with disabilities haven't registered for accommodations (EDUCAUSE 2020). buzzin needs no diagnosis and no paperwork. Students who are deaf or hard of hearing get live captions on their own device. Students with speech impairments can ask in writing and tap "I'm lost" without having to speak up in front of the room. Students with low vision see the slides up close on their own screen. Students with ADHD or learning disabilities can catch up on the last few minutes and reread what was said. Captions help everyone: 71% of students without hearing difficulties use captions at least some of the time (Linder 2016). buzzin adds to accommodations like CART; it doesn't replace them.

What it does

A professor shares one browser tab. Every student follows along on their own device.

Follow it: the slides and live captions, on every device. Captions arrive about 1 second after the professor speaks; new slides and animated bullets arrive in about 3 seconds.
Flag it: an anonymous "I'm lost" button, counted per slide, so the professor sees where the room is getting lost while there is still time to fix it.
Fix it: Ask answers questions from this lecture's own slides and words, cited to the slide. "I'm lost" gives a quick catch-up on the last few minutes.

Students join with a QR code or class code. No student account is needed to follow along.

How we built it

Next.js 16 (App Router) on Vercel.
Supabase: Postgres with row-level security on every table, reads through SECURITY DEFINER functions, and no service-role key in the app. A Realtime broadcast channel per lecture, Storage for slide images, and Auth with email link and Google (Microsoft / Entra ID ready).
Live captions: OpenAI gpt-live-transcribe over WebRTC, straight from the professor's browser, using a short-lived client secret our server mints. The API key stays on the server, and the audio never touches it. Voice activity detection runs in the browser.
Slide capture: the professor's tab compares small thumbnails of the screen share every second, keeps a slide once it has held still for 2 seconds, and catches build animations.
Grounded AI: a vision model reads each slide once. Ask answers only from the lecture's slides and transcript, with slide citations.
MCP server: students can bring the current lecture into their own AI tools.
The pitch video's music was composed with ElevenLabs.

Challenges we ran into

Caption lag. We brought the first caption words down to 0.75 seconds after speech starts by streaming over WebRTC with minimal delay; broadcasting to students adds about 0.3 seconds.
Build animations. Bullets that appear one at a time used to be missed; the slide detector now updates the same slide for every build step.
Privacy. Designing so that audio never passes through our server and student signals carry no names.

Accomplishments that we're proud of

About 1-second captions and 3-second slides, from one browser tab, with nothing to install.
Under $3 per lecture hour for the whole room, at OpenAI's published prices (captions at $0.017 a minute, plus slide reading and Ask). Live human captioning (CART) costs $65–150 an hour, for one student.
Private by design: no student accounts, no names in signals, and audio that never touches our server.

What we learned

Grounding AI in the course matters. At Georgia Tech, an AI that answers from the course's own material got 79% of questions right, against 31% for an off-the-shelf assistant on the same questions (Jill Watson study, Taneja et al. 2024). We built Ask the same way: it answers from this professor's own slides and words.

What's next for buzzin

Two professors, one pilot. First a security and privacy review with Georgia Tech OIT, Georgia Tech sign-in and private slide storage; then a one-semester pilot with two professors, measuring caption delay, "I'm lost" use and a student survey.