I Have A Question!

Landing page hero section.
Landing page features.
Question detection, answer generation, question generation, and sleep detection settings.
Live AI notetaking, with external references.
The homepage (which resembles a desk).
Notes page.
Voice cloning page.

Inspiration

We were inspired by the Instagram reel “I have a question”. So many questions but where are the answers? So we wanted to build an app that would answer questions for you.

As students in the 21st century, we’ve realized that our attention spans have been wrecked by ENDLESS doomscrolling. This means that we’re unable to focus in class for more than 10 minutes at a time. However, as diligent and hardworking students, we also don’t want to disappoint our teachers. Hence, we realized there was a problem: How can we appear as model students to teachers, without having to pay attention in class?

The solution was simple: All we had to do was to create an AI agent to listen in class FOR us, and generate questions for us to regurgitate back to teachers when they asked questions!

What it does

Answer Generation (Main Feature)

It listens actively to your teacher talking, and detects when questions are asked. Almost instantly, answers are generated to the teacher’s questions. (If you're too lazy to talk, you can even press a button for AI to respond for you, in your voice!)

Question Generation

Generates thoughtful questions to help you seem engaged and participate actively in class.

Notes Generation

Uses AI to do live notetaking during your lessons or lectures. It also auto generates external references for you to read up more about what the teacher is talking about.

Sleep Detection

Uses AI-powered face detection to wake you up when you fall asleep in class, or when questions are being asked. (So that you can respond immediately like the model student you are of course!)

How we built it

Answer Generation

We experimented with different speech-to-text workflows such as Whisper and ElevenLabs and decided to use the ElevenLabs realtime speech transcription API. Sentences are classified as question or statement browser-side. question sentences are sent to our API for answer generation.

We made use of ElevenLab’s voice cloning API to generate answers in the user’s voice, which can then be played on speaker.

Question Generation

An LLM is given context from the lesson transcription and asked to generate a relevant and intellectual question.

Notes Generation

Generation of notes is achieved by passing chunks of the transcribed audio into Gemini which is prompted to generate notes in real-time based on a context window of the transcribed audio. Specifically, since ElevenLabs produces Partial Transcriptions for real time audio transcription, we had to wait for the output to be stable before we took that as a finalised sentence to send to Gemini for notes generation.

Sleep Detection

We used MediaPipe for sleep detection, which detected facial landmarks. We checked if the Eye Aspect Ratio met a certain threshold to determine if the user’s eyes were closed (Eye Aspect Ratio). If the user has his or her eyes closed for more than 5 seconds (indicating sleepiness), an alarm will sound when the teacher asks a question (or if the user just falls asleep).

UI/UX

For the UI, we began by designing our mascot: a white cat wearing glasses, which we lovingly drew with Clip Studio Paint. We prototyped by creating the page layouts on Google Slides. Then, we instructed Codex to reproduce the page layouts, and we made adjustments as we went. Most of the elements are created with simple CSS elements, or stock images.

(Truly the future of UI prototyping)

Challenges we ran into

Codex is stupid
Copilot is incompetent
We got hungry
Internet connection is trash
Lack of sleep
Aircon too cold
I miss my cat
Spilling food

Accomplishments that we're proud of

< 3s latency between question asked and answer generated
UI is unique and feels homey
A plethora of features to make your lesson time as easy as possible 😎
Not running out of my 5h Codex limit (WE ONLY USED 20%)

What we learned

Vibe coding is truly the future! Could not have completed this project within the given time frame without the help of my good friends Claude Opus 4.5 and OpenAI Codex.
But sometimes, you need to be niche. And vibecoding doesn’t give you niche products. Thus, in order to be niche, you need to do things YOUR WAY sometimes.
Building a “GPT Wrapper” isn’t as easy as it seems.
Sometimes, you need to start simple. Perhaps a simple google slide mockup can kickstart your project.
Our parents were wrong. Daydreaming in class can help us accomplish things! (In our case, it would be coming up with this brilliant app idea.)