Inspiration
Every study app makes quiz questions on the whole topic, not each individual branch. Normally that's not too much of an issue, but when it comes to students who have no idea what anything is, they will struggle to even answer those questions. Eventually, I got so tired of taking practice tests of content that I didn't even understand, which led me to think, "What if I broke the topic down into parts, and studied each of them individually." That's what Grove does.
What it does
Grove firsts lets you put in your own notes. Links, PDFs, Files, videos, almost anything can be put. Next, it creates a tree-like structure, with the main topic in the middle, and subtopics surrounding it. If you are still confused on those subtopics, you can create subtopics of the subtopics. Once you are ready studying a specific subtopic (how many times have I said subtopic), you can take a 5 question quiz. If you get 5/5, you complete it, if you get it wrong, you have to restart with new questions. When all the subtopics are complete, a final quiz on everything is ready for you to take.
How I built it
I first made files turn into text, so that the AI could read it. Then I made the AI build a topic tree and write quizzes. The problem with that was that it was completely untrustworthy, so I then made AI show it's sources. After that it was just troubleshooting errors,
Challenges I ran into
First, proving a quote is actually real was a challenge. YouTube was also difficult, since it no longer gives caption data to programs, so if the captions don't exist, I made it so that it requests it through the player interface and transcribe the audio. Finally, two bugs made the app look frozen while every server test passed. Both were in the browser, and we only found them by screenshotting the real page instead of trusting the code.
Accomplishments that I'm proud of
One thing that I'm proud of is that when uploading a massive playlist of videos, it surprisingly held up and actually worked. Another thing I'm proud of is that the fact-checking actually does its job. If the AI writes a question it cannot back up with a real quote from your material, that question gets thrown out, and if none of them pass, Grove tells you instead of showing you a quiz it made up. I tested this by making the checker reject everything on purpose, just to be sure it would fail successfully.
What I learned
I learned that generating questions with AI is easy, but making sure it's accurate is the hard part. I also learned a lot about how text search works. Finding exact words in a text is not enough to prove a quote is real. Small words like "not" or a changed word order can completely flip the meaning, even if the words look the same. Building a search tool from scratch means teaching a computer to understand what sentences actually mean, rather than just matching up letters.
What's next for Grove
To start, we want to make accounts, so that you can get quizzed on your phone, computer, and on any device. Also, I'll add more subtopic layers for larger amounts of material. In addition, I plan to make shared spaces so that a classroom or group of friends can work on it together.
Built With
- css
- express.js
- groq
- html
- javascript
- llama
- node.js
- whisper

Log in or sign up for Devpost to join the conversation.