Aura

Inspiration

What it does

How we built it

Inspiration

Aura started with a problem we face while studying: the answer is probably somewhere in our lecture slides or PDFs, but finding it—and actually understanding it—can take a lot of time.

We often turn to AI for help, but a general chatbot doesn’t necessarily know what our professor covered or which explanation matches our course. Even when an answer sounds convincing, we still have to check it against our notes.

We wanted something that could explain our own study materials and show us where the explanation came from. That became Aura.

What it does

Aura lets students upload their course materials and ask questions about them in everyday language. It provides explanations with citations so students can go back to the relevant page or slide and check the source.

Students can choose beginner or advanced explanations, ask follow-up questions, and request summaries of longer materials.

We also wanted Aura to be honest about what it knows. When an answer goes beyond the uploaded material, it should make that clear. When a question needs more context, it should ask for clarification.

How we built it

We built Aura with a FastAPI backend, a Vite-based frontend, and PostgreSQL with pgvector for searching document content.

When a student uploads a file, Aura extracts the text, breaks it into smaller sections, and keeps track of where those sections came from. We use Qwen3 embeddings to help find passages related to the student’s question.

For this hackathon, we integrated NVIDIA Nemotron through Nebius Token Factory to generate explanations and help evaluate whether the retrieved material contains enough information to answer the question.

We also used LibreOffice to convert older PowerPoint files, because real lecture materials don’t always come in convenient formats.

Challenges we ran into

One frustrating issue was Aura saying that a topic wasn’t in the material when we could see it right there in the slides. Fixing this meant looking beyond the final answer and checking how the content was extracted, retrieved, and evaluated.

Document processing was another challenge. Some presentations took several minutes to embed on the CPU, which made uploading feel slow. We investigated the bottleneck and configured GPU acceleration for embeddings.

Long summaries exposed a different problem. Retrieving a few relevant passages can work for a specific question, but it doesn’t necessarily give the model enough context to summarize an entire presentation. We had to rethink how Aura handled those broader requests.

Throughout development, we kept coming back to the same question: Can a student understand this answer and check why they should trust it?

What we learned

We learned that a useful AI tutor depends on much more than the model. If document extraction misses something or retrieval brings back the wrong passages, even a strong model can struggle to give a useful answer.

We also learned that finding relevant content and finding enough content are two different things. That distinction became especially important for summaries and questions that span several slides.

Working with Nemotron on Nebius Token Factory gave us practical experience using a model for both explanations and structured decisions inside a larger application.

What we're proud of

We’re proud of bringing the pieces together into a working flow: upload course material, ask a question, get an explanation, and follow the citations back to the source.

Testing Aura with actual lecture slides helped us see where it worked and where it still needed improvement. Seeing it produce a useful explanation grounded in those materials made the work feel worthwhile.

What's next

We want to make uploads faster, improve retrieval and citation accuracy, and make long-document summaries and follow-up conversations more reliable.

We also plan to expand personalized practice and support more types of learning material.

Ultimately, we want Aura to be something students can keep beside them while studying—a place to ask “Why does this happen?” or “Can you explain that more simply?” without losing the connection to what they’re learning.

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Aura

Built With

• cuda
• embedding
• factory
• fastapi
• javascript
• libreoffice
• nebius
• nemotron
• nvidia
• pgvector
• postgresql
• python
• pytorch
• qwen3
• rag
• sentence
• token
• transformers
• vite