aerie

Inspiration

In many CS courses, we found ourselves with a ton of notes but still feeling unprepared because of a lack of practice. Your LLM of choice can generate a couple of janky multiple-choice questions, but that's not enough to get ready for a real test.

So we built a platform that not only transforms your notes into practice, but also remembers what you've learned. Like a personal tutor who sees exactly where you’re struggling, aerie works with you to target your weak spots over time.

What it does

Works from your study materials: Upload your course materials like syllabi, lecture notes, etc. -> aerie generates practice questions grounded in what you’ve actually been taught

Remembers: As you practice, aerie builds a living memory of your strengths and weaknesses, using that to surface insights and guide learning. Over time, it will update your materials to reflect your current level.

Supports: An AI conversational tutor is embedded into each question, helping you understand your mistakes in real time by generating diagrams, providing detailed explanations, and walking you through examples.

The result? Unlimited, tailored practice. Want to create a practice test focusing on your calculated weaknesses? Got it. Would you like to prioritize a topic from your notes in the study set? Done. And once you’ve chatted with aerie about what you got wrong, just keep working through variants by clicking on the “generate similar question” button until you’re confident.

How we built it

Frontend:

React + TypeScript
Tailwind CSS + shadcn/ui for styling

Backend:

Python + FastAPI
SQLAlchemy

AI:

Google Gemini for question generation, tutoring, and learning insights
Supermemory for document/user RAG and building personal student learning profiles

Challenges we ran into

Building a learning profile that's useful (not just "you answered 3 questions")
- Solution: careful integration with Supermemory's graph memory and Gemini for insight extraction.
Dealing with changing content within the question component via dynamic wrapping was a repetitive flexbox struggle.
- Solution: careful application of min-w-0, break-words, and overflow-hidden to wrapper containers

Accomplishments we're proud of

Creating a complex and effective multistep pipeline from documents and user input to frontend-displayed questions.
- We transform a prompt using Gemini’s NLP into targeted semantic queries for Supermemory to reference the most relevant user skills and notes. We then ask Gemini to brainstorm the best question types for the material and generate the full question with a preloaded answer and explanation for our frontend to display.
Crafting a cohesive user experience in a short amount of time with a refined UI, consistently themed visuals, and some fleshed-out animations.
Supporting many question types: multiple choice, checkbox, true/false, coding, fill-in-the-blank, matching, and drag-and-drop ordering.

What we learned

Supermemory's persistent memory and knowledge graphs are extremely powerful for education. By allowing us to build a profile of each student’s skills that evolves over time, each generated question becomes more useful than the last.
Scrollbar formatting just sucks in webdev. Especially if you want global scroll.

What's next for aerie

Spaced repetition: Resurface questions or topics you struggled with at optimal intervals as “needing review.”
More question types: Diagram labeling, proof construction, FRQs with autograding
Mobile support: Take your AI tutor with you wherever you go