Project Name: LearnSphere AI — Multimodal Personalized Tutor

Track: Track D: Personalized Tutoring and Adaptive Learning | Multimodal AI Hackathon 2026

Inspiration: Students often struggle to study across separate textbook PDFs, presentation slides, lecture videos, and handwritten notes. When they ask generic AI chatbots for help, the chatbots often hallucinate, invent fake references, and do not know what the student has actually understood. We built LearnSphere AI to solve this problem. It is an evidence-grounded AI learning companion that answers questions using verifiable course citations and tracks student mastery using real mathematics.

What It Does:

Multimodal Materials Workspace: Combines textbook PDFs, slide decks, video transcripts with timestamps, and study notes while keeping exact page numbers, slide numbers, and timestamps.
Evidence-Grounded AI Tutor: Answers student questions with clickable citation badges like Slide 5 or Page 14. Clicking a badge opens the exact source excerpt so students can verify the answer.
Honest Refusal Safeguard: If a question is not covered in the course materials, the tutor honestly refuses to answer rather than guessing or making things up.
Adaptive Quiz Studio: Generates practice questions calibrated to the student skill level (Easy, Medium, Hard) and checks numerical answers deterministically.
Bayesian Knowledge Tracing: Uses the Corbett and Anderson mathematical model to calculate the student mastery probability after every response, accounting for lucky guesses and accidental slips.
Knowledge Graph and Study Planner: Visualizes prerequisite concepts and schedules personalized revision time focusing on the student weakest topics.

How We Built It:

Frontend: Built with React, Vite, TypeScript, and Tailwind CSS.
Backend: Built with Python FastAPI and local storage.
Learner Model: Full implementation of Bayesian Knowledge Tracing update equations.
Search and Retrieval: Fast TF-IDF passage retrieval with confidence-based refusal thresholds.
Evaluation: Built-in benchmark lab measuring Faithfulness at 95 percent, Relevancy at 92 percent, and Refusal Honesty at 100 percent.

Challenges We Faced:

Synchronizing unstructured textbook paragraphs, slide bullets, and video timestamps into clean searchable chunks.
Implementing the Bayesian Knowledge Tracing equations to ensure stable mastery updates after correct and incorrect answers.
Setting the right retrieval threshold so the AI tutor answers valid questions while refusing out-of-domain trick queries.

What We Learned:

Source-grounded citations build genuine trust with students and prevent hallucination.
Probabilistic learner models like Bayesian Knowledge Tracing provide much better personalized guidance than simple test percentages.
Clean local fallbacks ensure an application is 100 percent reliable and ready for demo without needing paid external API keys.

What is Next for LearnSphere AI:

Direct diagram and handwriting recognition using mobile photo uploads.
Translation support for Indian languages like Hindi, Telugu, and Tamil while preserving original English citations.
Voice-enabled audio tutoring companion for hands-free study sessions.