StudyMate

Inspiration

Students rarely study from one place. Lectures are on video, notes are in slides, and details sit in a textbook. When they turn to a general AI chatbot, the answers often don't match their course, and the bot has no idea what they already understand. We wanted a study companion that only teaches from your own material, shows its sources, and adapts to you.

What it does

StudyMate turns lecture videos, textbooks and slide decks into one source-cited knowledge base, then uses it to tutor and test you.

Multimodal ingestion: upload PDFs, slides and videos with no manual preprocessing. Text, transcripts and figures are extracted automatically.
Concept map: content is organised into topics, concepts and prerequisites, and every unit links back to its page, slide or video timestamp.
Grounded tutor chat: answers carry citations that open the exact page, slide or timestamp. If the material doesn't cover a question, StudyMate says so and clearly separates any outside knowledge.
Adaptive assessments: quizzes and mock exams (MCQ, short answer, numerical) on any scope you choose. Each question is tagged with topic, source and difficulty, verified for correctness, and never repeated.
Learner model: per-topic mastery updates after every quiz and chat, so weak topics, misconceptions and study plans reflect what you actually know.
Dashboard: a mastery heatmap, a course flow map and a post-quiz report with cited feedback.

How we built it

Ingestion: PyMuPDF for text and page images, a vision LLM to caption diagrams, and Whisper for timestamped video transcripts. Every chunk keeps its origin (page, slide or timestamp).
Knowledge graph: an LLM extracts topics, concepts and prerequisite edges, and each content unit is tagged to concepts.
Grounded RAG: hybrid retrieval (BM25 + embeddings) with reranking. Answers must cite retrieved excerpts, and a verifier pass removes unsupported claims. Low retrieval confidence triggers a refusal.
Question engine: questions are generated from source units, then verified (numerical answers by executing code, other questions by blind cross-model validation). Embedding similarity rejects near-duplicates.
Learner model: Bayesian Knowledge Tracing per concept:

$$P(L_{t+1}) = P(L_t \mid obs) + \bigl(1 - P(L_t \mid obs)\bigr)\,P(T)$$

with slip and guess parameters in the posterior update, plus an exponential forgetting term $e^{-\Delta t / s}$ for spaced review.

Evaluation: RAGAS for faithfulness, answer relevancy, context precision and recall on a team-built test set, plus simulated students comparing adaptive vs random question selection.
UI: Streamlit, SQLite and ChromaDB.

Challenges we ran into

Keeping citations precise, down to the right page or a timestamp within a few seconds, rather than just the right document.
Getting useful information out of diagrams and figures.
Making sure generated questions had correct answer keys, since a wrong key corrupts the learner model.
Deciding when the tutor should refuse versus answer from partial evidence.
Evaluating personalization honestly without real students.

Accomplishments that we're proud of

Citations that open the exact page, slide or timestamp.
Correct refusal on [X]% of off-material questions in our test set.
Faithfulness of [X] and context recall of [X] on RAGAS.
Simulated students gained [X] mastery on average with adaptive selection vs [X] with random, with a [X]% question repetition rate.

What we learned

Grounding is mostly a retrieval-quality and refusal-threshold problem, not just a prompting problem.
Verified questions matter more than many questions.
Even a simple BKT model with forgetting and prerequisite awareness makes personalization measurable.
Honest evaluation, including failure cases, tells you far more than a single headline number.