Persona

Main feel/look
User speaking
Scorecard
AI speech analytics
Revision PDF
Session history

Inspiration

We've all done it. Studied for hours, felt completely ready, then the moment someone asks us to explain it...nothing. The words don't come out.

That's not a studying problem. That's a how-you-study problem. You were reading, not learning. Your brain recognized the words on the page, but it never actually built the pathways needed to recall and explain them under pressure. We wanted to fix that.

What it does

Persona puts you in a real voice conversation with an AI study partner named Sophia. You speak. She listens. She pushes back with "wait, can you explain that part again?" or "can you give me a real-world example?", exactly like a curious friend who just doesn't quite get it yet.

There's also a Quizzer mode where you upload your notes and Persona fires rapid oral flashcard questions at you, evaluating your answers in real time.

At the end of every session you get a full scorecard, your fluency score, your weak spots, your strong points, and a set of custom revision questions generated specifically around what you got wrong. You can download the whole thing as a PDF to study from before your next exam.

Features

You talk, it listens, and it pushes back.

The core of Persona is a real voice conversation. You aren't typing answers into a box or swiping through cards, you're speaking out loud, being heard, and getting a response in seconds. It feels surprisingly close to actually talking to someone.

Two modes, depending on what you need.

If you're trying to deeply understand a topic, Teach-It mode puts you in conversation with Sophia, an AI peer who genuinely wants you to explain things to her. She'll ask "wait, can you give me an example?" or "I don't quite understand that part" until the concept actually clicks. If you just need to drill your notes before an exam, Quizzer mode uploads your material and fires questions at you, direct questions, true/false, fill-in-the-blanks, one after another, fast.

It catches the things you don't even notice.

When you hesitate, say "um" or "uh", or start trailing off mid-sentence, Persona tracks it. It records how many filler words crept into your speech and calculates a hesitation density score. That number matters because it separates "I know this and I'm explaining it confidently" from "I kind of half-know this and I'm winging it."

Your notes become your quiz.

You drag in a PDF or paste your notes, and Persona automatically pulls out the key concepts your session will revolve around. No setup, no manual input.

You can see your progress in real time.

As the session goes on, a live concept checklist tracks what you've nailed and what you're still shaky on. Green means you explained it clearly. Yellow means you stumbled. It updates as you speak.

At the end, you get a real scorecard.

Not just a number. A full breakdown, how clear were your explanations, how deep did you go, how fluent did you sound, what were your strongest moments, and where did your understanding clearly fall apart. Then, based on exactly where you struggled, Persona generates a set of targeted revision questions, specific to your session, your weak spots, that you can download as a PDF and study from before the exam.

The whole thing downloads as a PDF.

One click. Your full scorecard, analytics, and revision questions, saved as a clean, formatted document you can refer to whenever.

How we built it

We needed the conversation to feel instant, otherwise the whole illusion falls apart. So we chose Groq as our inference backbone - it's genuinely the fastest way to run Llama 3 and Whisper today.

The voice pipeline is: your mic → Whisper (via Groq) for transcription → Llama 3 for the AI response → ElevenLabs streaming TTS for Sophia's voice. The whole loop is tuned to be as tight as possible.

The frontend is Next.js 15 with Framer Motion for the mic orb animations, and Tailwind CSS v4 for styling. The PDF export uses html-to-image and jsPDF.

Challenges we ran into

Getting Whisper to keep filler words like "um" and "uh" was surprisingly hard. Most transcription models are trained to clean those up. But for us, those words are data, they tell us where a student hesitated or lost confidence. We had to engineer a custom transcription prompt to force the model to preserve them.

Accomplishments that we're proud of

Mainly the fact that we got the whole thing working, and the fact that this is going to be useful for all of us in our exam prep. The fluency analytics are also something we're proud of. Happy that we added a deterministic scoring engine that mathematically penalises high hesitation density, so the feedback isn't just vibes, it's based on real speech metrics.

What we learned

We learned that prompting a conversational LLM is completely different to prompting one for structured output. Keeping Sophia short, focused, and genuinely curious, without lecturing, required a lot of careful iteration.

What's next for Persona

We want Persona to plug directly into your Notion notes or university course page so there's zero friction getting started. Longer term, vocal stress analysis, going beyond filler words to measure how confident your voice actually sounds. Would love to make this a tool for all students around the world.