UA AI Studio

Inspiration

Students everywhere are using AI for schoolwork, and a lot of that use ends up crossing academic integrity lines. But AI is not going away, and pretending students will not use it is not realistic. We wanted to build a suite of AI educational tools that make AI useful inside the classroom instead of working around it. The goal was to create tools that use secure course data, follow the rules of the class, and help students learn instead of just giving them the answer.

What it does

1. Classroom LLM: Professors can upload course material like syllabi, project specs, notes, readings, and other class documents. Students then get access to a class-specific LLM that answers using only that course material. It can help students understand project specs, figure out where to start, review class concepts, and ask questions without needing to wait for office hours. The model is designed to guide students instead of oversharing or giving full homework answers.

2. AI Reading: Our reading AI integrates a chatbot into a PDF viewer so students can talk with a paper or reading in the context of the class. Students can highlight specific sections and ask the AI to summarize, explain, or connect the reading back to course concepts. The goal is not to replace reading, but to help students slow down, ask better questions, and actually internalize the material.

3. AI Writing: Our writing AI reviews student writing against an assignment spec or rubric. It gives feedback on what the student is doing well, what is unclear, and where they should expand. Instead of rewriting the paper for them, it acts more like a writing coach that helps the student understand whether they are on track.

4. Forums: We also built a classroom discussion page where students can ask questions, share ideas, and learn from each other. The forum gives students a central place to discuss assignments and course material, while keeping the conversation tied to the class instead of scattered across group chats or outside platforms.

How we built it

We built a memory-aware RAG system using FastAPI, Supabase/Postgres with vector search, and Gemini. Course documents are uploaded, parsed, chunked, embedded, and stored with metadata like class, assignment, document title, visibility, and source information. When a student asks a question, the backend retrieves the most relevant chunks, passes them into the LLM, and returns both an answer and the sources used. The backend is connected to our Vite frontend, where students can interact with the classroom chatbot and course tools.

Challenges we ran into

• We ran into speed issues with our RAG pipeline. Our first model choice was too slow for a live demo, so we switched from Gemma 4 to Gemini 2.5 Flash-Lite.
• We had to figure out how to balance helpful answers with academic integrity, especially for tutor mode.
• User authentication and Supabase integration took time to coordinate across the frontend and backend.
• Setting up Lovable and connecting generated frontend work with our custom backend was harder than expected.
• We also had to make sure uploaded documents were parsed, chunked, embedded, and retrieved in a way that made the chatbot actually useful.

Accomplishments that we're proud of

• We built a working memory-aware RAG chatbot for class-specific documents.
• We created a scalable SQL database design using Supabase/Postgres and vector search.
• We connected document upload, retrieval, generation, memory, and frontend interaction into one working flow.
• We built a UI/UX experience that feels like something students and professors could actually use.
• We created multiple AI modes, including direct Q&A, tutor-style guidance, and brainstorming support.

What we learned

We learned how to build a memory-aware RAG LLM system end to end. We also learned how to work across a team, connect a FastAPI backend to a Vite frontend, use Supabase as a database layer, and make real product decisions under time pressure. A big lesson was that building an AI feature is not just about calling an LLM. The hard part is making sure the data, prompts, retrieval, memory, and user experience all work together.

What's next for UA AI Studio

We want UA AI Studio to help pioneer ethical and educational AI tools at the university level. AI is powerful, and students should be able to use it in a way that helps them learn correctly. Next, we want to improve our tools, support more document types like slides and notes, add better math and LaTeX rendering, improve professor controls, and make the system easier to use across more classes. Our goal is to keep building AI tools that make learning more personalized, more accessible, and more aligned with classroom expectations.

Built With

• fastapi
• gemma4
• python
• react
• supabase