Homepage hero showing source-grounded learning and platform features

Inspiration

Generic generative AI chatbots often fail in educational settings: they produce convincing hallucinations, cite non-existent sources, and fail to track whether a student is actually retaining concepts. We built AI Study Companion for the IIT Mandi Multimodal Hackathon 2026 (Track D: Personalized Tutoring & Adaptive Learning) to transform AI tutoring into a grounded, measurable, and adaptive learning experience that teachers and students can completely trust.

What it does

AI Study Companion is a source-grounded intelligent tutoring platform that personalizes learning pace while eliminating misinformation:

100% Source-Grounded Tutoring: Every response is backed by inline citations (``). Clicking any chip immediately jumps to the exact textbook page, video timestamp, or lecture slide.
Honest Abstention: If a question falls outside syllabus boundaries or uploaded materials, the tutor safely abstains rather than inventing answers.
Bayesian Knowledge Tracing (BKT): Updates the posterior mastery probability $P(L_t)$ in real time after every practice question to trace true conceptual comprehension: $$P(L_{t+1}) = P(L_t \mid \text{Action}) + (1 - P(L_t \mid \text{Action})) \cdot P(T)$$
Adaptive Quiz Engine: Dynamically selects practice problems tailored to learner ability, calibrating around a ~70% expected success rate.
Interactive Prerequisite Knowledge Map: Visualizes concept dependencies with color-coded nodes based on BKT mastery levels (High, Moderate, Low, Untested).
Active Revision Hub: Features spaced-repetition flashcards with flip interactions and synthetic audio briefs.
Evaluation & Benchmarks: Integrates RAGAS evaluation metrics measuring Faithfulness, Answer Relevancy, Context Precision, and Refusal Recall.

How we built it

Frontend & Routing: Built with Next.js (App Router) and TypeScript (strict mode) for a fast, responsive user interface.
State & Architecture: Managed via Zustand stores for student session state, BKT progression, and quiz interactions, with Zod schema validations for data contracts.
Styling & UI: Styled using Tailwind CSS, enhanced with Lucide React icons and Framer Motion for smooth card transitions and node visualizations.
Data Visualization: Built dynamic mastery and performance charts using Recharts.
Dual-Mode Backend: Architected to run out of the box on realistic mock data on Vercel with clean toggles to connect to a production FastAPI backend via environment variables.

Challenges we faced

Multimodal Alignment: Ensuring inline citation chips properly resolved across varied media formats (e.g., deep-linking to an exact timestamp in a video vs. a specific page in a PDF).
Client-Side BKT State Management: Implementing Bayesian Knowledge Tracing transitions cleanly on each quiz answer without causing latency or unnecessary UI re-renders.
Strict Grounding Enforcement: Designing UI patterns that clearly differentiate verified cited content from unsupported off-syllabus queries.

Accomplishments that we're proud of

Seamlessly deploying a production-ready application to Vercel that functions smoothly with interactive UI states.
Creating a clean, intuitive knowledge map that lets students visually diagnose prerequisite gaps before taking advanced quizzes.
Achieving strong alignment with the RAGAS framework principles for educational evaluation.

What we learned

The mathematical nuances of tuning BKT priors, transition rates , slips , and guesses to accurately mirror genuine human learning.
How to structure modern Next.js applications to toggle seamlessly between local mock testing and live API endpoints.

What's next for AI Study Companion

Direct ingestion pipelines for full course syllabi, lecture slides, and recorded lectures.
Socratic voice-to-voice interactive tutoring sessions.
Teacher dashboards offering aggregate mastery analytics across entire classrooms.