Inspiration
Students learn from a lecture video, a textbook and a slide deck, but each one lives in a different place. When you get stuck, you search three tools and still don't know which source to trust. Most AI chatbots make it worse: they give confident answers without showing where the answer came from. We wanted a study companion that always shows its proof.
What it does
StudyMate merges lecture videos, textbooks and slides into one knowledge base for Track D (Personalized Tutoring & Adaptive Learning).
- Source-cited answers. Every claim carries a clickable citation: a lecture timestamp, a textbook page or a slide number. One click opens the source with the exact passage highlighted.
- Honest by design. If your materials don't cover a question, StudyMate says so instead of guessing, and shows a confidence score on every answer.
- A tutor, not just an answer box. When your mastery of a topic is low, it gives a hint and a guiding question first. A "Show full answer" button reveals the rest.
- Adaptive practice. It generates multiple-choice, short-answer and application questions from your own sources. Difficulty adapts to what you know, graded feedback cites the source, and weak topics return through spaced repetition.
- Progress dashboard. A mastery heatmap by topic, your progress over time, your weak topics, and a plan for your next study session with links back to the material.
How we built it
- Ingestion: PDFs are parsed page by page, slides slide by slide (including speaker notes), and lectures are transcribed with Whisper (faster-whisper) or taken from a timestamped transcript or a YouTube link. A vision model describes figures and diagrams. Each passage keeps its source, location and an AI-assigned topic.
- Retrieval: ChromaDB semantic search combined with BM25 keyword search, merged with reciprocal rank fusion and a similarity-based rerank.
- Tutoring and quizzes: Gemini writes cited answers, generates questions grounded in a source passage, and grades short answers against a rubric.
- Learner model: an Elo-style mastery score per topic in SQLite, with a review interval that doubles after correct answers and resets after misses.
- App: a FastAPI backend and a Next.js + Tailwind frontend.
Challenges we ran into
Keeping citations honest was the hardest part. The model can only cite numbered passages that retrieval actually returned, and low-confidence questions are refused. Getting retrieval to work across very different formats (spoken transcript, textbook prose, bullet-point slides) also took careful chunking.
Accomplishments that we're proud of
- Clickable citations that jump to the exact lecture timestamp, page or slide
- A tutor that scaffolds learning before giving the answer
- Adaptive questions that always cite their source
- A system that admits when it doesn't know
Built With
- bm25
- chromadb
- fastapi
- gemini-api
- javascript
- next.js
- pypdf
- python
- react
- sqlite
- tailwind-css
- whisper
- yt-dlp
Log in or sign up for Devpost to join the conversation.