Professor Me | Devpost

Inspiration

Every study app out there makes the AI the teacher. But the best way to actually learn something is to explain it out loud to someone clueless about the topic and watch where you get stuck. It's also know as the Feynman technique, and nobody's built it because it requires having an active listener that actually engage with what you're teaching, something you might not always have. We saw the problem there, and built Professor Me to help.

What it does

First you pick any topic you want to study. You can upload notes or choose a topic on your own. Next you choose your student - Sam or Elena - and their curiousity level. You explain the topic out loud, and they ask for example when you are vague, get stuck when you hand-wave, and ask "wait, why?" just like a real student would.

After your teaching session, you get a gap map to explain what you nailed, where you could improve, and what topics you missed. You can also visit your profile to check account analytics and revisit old sessions.

How we built it

One model can't be dumb and strategic at the same time. So we split it: an Evaluator that secretly knows the truth and grades you, and a Persona that knows nothing and only reacts. A plain non-AI function sits in between deciding probe vs. dig deeper vs. move on, so we could unit test it instead of guessing.

Integrating voice was one of the parts we were scared of. We used Chrome to listen to speak, ElevenLabs to talk back, both of which streams so the student starts talking mid-sentence instead of waiting for the whole reply.

Challenges we ran into

Models paraphrase all the time. This is not ideal a big part of our project is quoting the user verbatim when reviewing their session. So every quote gets checked word-for-word against your real transcript, and gets swapped out if it doesn't match.
Vercel can't hold a websocket open, so our original plan broke about halfway through. We ended up moving the ElevenLabs connection into the browser instead, with a key that only does text-to-speech, in case someone pokes around in devtools.
Our student was also way too agreeable at first, basically being a complete yes-man and repeating your sentence back at you. We spent a lot of hours honing the ai so that it would be agreeable but also push further if necessary.

Accomplishments that we're proud of

Four of us built this in parallel with almost no merge conflicts, mostly because we forced ourselves to agree on data shapes in hour one and never touched them again. We created a detailed design doc with tasks, checkpoints, and milestones for each member to refer to and that made collaborating a breeze.

What's next for Professor Me

Add more personas with different personalities
Allow Professor Me to remember past chats for use in future chats
Expand on notes upload feature: allow for user to select specific topics to focus on