• 

• 

• 

• 

• 

• 

• 

• 

StudyMind AI — Source-Grounded Multimodal Tutor & Adaptive Learning Engine

Every answer cited. Every struggle traced to its root. Every concept mastered with calibrated confidence.

---

💡 Inspiration

Current AI study tools and generic LLMs frequently hallucinate facts, guess answers to un-uploaded topics, and offer superficial explanations without knowing why a student is stuck. When students study complex undergraduate courses—like Distributed Systems or Advanced Algorithms—a generic chatbot fails because:

1. It quotes external internet knowledge rather than the professor’s exact slides and lecture recordings.
2. It treats mistakes as isolated failures rather than symptoms of missing foundational prerequisites.
3. It cannot detect when two course readings contradict each other or audit if the whole syllabus is covered.

We built StudyMind AI to be a strictly source-grounded academic tutor that bridges textbook PDFs, PowerPoint slides, and lecture videos into an interconnected Knowledge DAG governed by Bayesian Knowledge Tracing (BKT) and Self-RAG claim-level entailment.

---

🚀 What It Does & 6 Key Architectural Innovations

StudyMind AI transforms raw lecture notes, video recordings, and syllabus documents into an adaptive learning operating system:

1. 🔍 "Why Am I Stuck?" Root-Cause Prerequisite Tracing

Instead of just re-explaining a failed concept, StudyMind AI walks backward across the course's prerequisite Directed Acyclic Graph (DAG) and traces error patterns to the deepest unmastered root cause.

• Measured Impact: In our empirical simulation benchmark, targeted root-cause remediation reduced average student time-to-mastery from 8.40 ± 0.42 sessions down to 3.07 ± 0.25 sessions (a 2.73× acceleration), achieving 100% Top-1 root-cause diagnosis accuracy.

2. ⚖️ Source Conflict Detector

When professors update lecture slides but assign legacy textbook chapters, contradictions arise. StudyMind AI automatically audits ingested sources for cross-document contradictions (differing definitions, conflicting complexity bounds, or outdated protocol defaults) and surfaces them with exact dual citations so students aren't misled.

• Measured Accuracy: 1.000 Precision, 1.000 Recall, 1.000 F1 on injected course contradictions.

3. 🎓 Interactive Teach-Back Mode with Claim-by-Claim Entailment

To ensure true active recall, students can teach concepts back to the AI. StudyMind AI breaks down the student's explanation into atomic claims, runs NLI entailment against the source documents, identifies flawed statements and missing key points, and directly adjusts the student's Bayesian Knowledge Tracing profile ($P(L)$) based on genuine understanding.

4. 📋 Automated Syllabus Coverage Audit & Strict Refusal

StudyMind AI parses official syllabus PDFs, extracts unit-by-unit learning objectives, and audits course uploads for coverage depth (Covered, Thin, or Missing). If a student asks about a concept not covered in their uploaded materials, the tutor strictly refuses rather than hallucinating external facts.

• Audited Performance: 96% coverage mapped across real academic curricula with 1.000 Refusal Recall on out-of-scope topics.

5. 🎯 Confidence-Calibrated Mastery & Distractor Misconception Profiling

Every assessment question pairs multiple-choice answers with metacognitive confidence tags (Sure, Not sure, Guess).

• Calibrated BKT discounts lucky guesses and heavily penalizes high-confidence misconceptions.
• Real-time tracking calculates the student's Brier Score ($0.0 \le BS \le 1.0$), reducing state prediction error (20.27% RMSE reduction over standard BKT from $0.3501$ down to $0.2791$).

6. ⏱️ Lecture-to-Slide Multimodal Dynamic Alignment

Aligns unstructured lecture video transcripts with slide decks using monotonic Dynamic Programming (DP forward pass). Students can click "Watch this part of lecture" from any slide, or view "Slides for this moment" during video playback.

---

🛠️ How We Built It

• Frontend: High-performance React 19 + TypeScript + Vite styled with customized TailwindCSS, featuring interactive Force-Directed Knowledge DAGs, audio-visual study desks, and real-time Socratic chat interfaces.
• Backend: FastAPI modular micro-router architecture with Pydantic schemas and thread-safe persistent atomic JSON caching.
• Multimodal Pipeline:
  • Video & Audio: OpenCV keyframe extraction with subtitle parsing and speech-to-text chunking with exact timestamp boundaries.
  • Document Layout: PDF, DOCX, and PPTX structural chunking preserving table hierarchies, formulas, and visual diagrams.
• Adaptive Intelligence:
  • Custom implementation of Bayesian Knowledge Tracing (Corbett & Anderson) extended with metacognitive Brier score weighting.
  • Self-RAG Guardrail: Sentence-level claim decomposition and cosine-similarity verification against retrieved chunks before any response is rendered.
• Database & Auth:

Built With

• fastapi
• groq
• opencv
• postgresql
• python
• rag
• react
• supabase
• tailwindcss
• typescript
• vite
• whisper