Inspiration Students drown in scattered notes, so we built Edify—to unify your courseware into a single, source-grounded path to mastery.
Inspiration
Every semester, we noticed a recurring bottleneck in how we study: key insights are fragmentedly scattered across hours of unindexed lecture recordings, dense textbook chapters, and disjointed slide decks. When turning to standard AI tools, we constantly hit a wall—either they hallucinated outside knowledge or provided generic explanations completely detached from our actual course syllabus. We realized that true academic acceleration requires an intelligent system that treats official coursework as the absolute, verifiable source of truth. That core realization is what drove us to engineer Edify.
What it does
Edify is a comprehensive, production-ready multimodal study platform built for Track D (Personalized Tutoring & Adaptive Learning) at the IIT Mandi hackathon. It seamlessly ingests raw, unstructured course assets without requiring manual preprocessing and transforms them into an interactive cognitive ecosystem:
- Source-Grounded Chat & Citations: Every single explanation, summary, and response is strictly tied to the curriculum, generating precise citations that instantly open up the exact textbook page, slide number, or video timestamp.
- Adaptive Assessments: Dynamically generates rigorous quizzes—including MCQs, short-answer questions, and numerical problem sets—tailored to a student-chosen scope and difficulty level.
- Intelligent Knowledge Tracing: Powered by robust probabilistic learning models, Edify tracks per-topic mastery estimates and automatically updates them after every session to generate deep diagnostic post-assessment reports on student misconceptions.
How we built it
We designed Edify from the ground up with a rigorous, enterprise-grade systems architecture:
- Multimodal Ingestion Pipeline: Utilizing advanced Vision-Language Models (VLMs) to parse textbook figures, charts, and slide decks, alongside robust speech-to-text transcript processing for lecture videos.
- Vector Knowledge Base & Graph: Structuring ingested content into a clean, hierarchical web of topics, subtopics, and strict prerequisite dependencies.
- Evaluation Rigor: Benchmarking our retrieval-augmented generation (RAG) pipeline using industry-standard frameworks to guarantee high faithfulness and context precision. Mathematically, context precision is defined as: $$ \text{Context Precision} = \frac{\sum_{k=1}^{N} (\text{Precision} \times \text{v}k)}{\sum{k=1}^{N} \text{v}_k} $$
Challenges we ran into
Engineering a seamless multi-format extraction engine presented major hurdles—specifically aligning precise video timestamps accurately with textbook page numbers and slide indices without drifting. Furthermore, tuning our retrieval and generation guardrails to gracefully decline out-of-scope or off-material queries (rather than allowing the model to hallucinate assumptions) required extensive prompt-chain optimization and rigorous threshold testing.
Accomplishments that we're proud of
- Successfully architecting and deploying an end-to-end multimodal pipeline that bridges video timestamps, slide decks, and heavy textbooks into a single unified interface.
- Implementing absolute source-grounding mechanisms that completely eradicate unverified AI hallucinations.
- Delivering a sophisticated per-topic mastery tracking and adaptive quiz engine that gives students clear, actionable insights into their learning gaps.
What we learned
We learned that building effective educational technology requires placing strict boundaries around AI. Students don't just want an "answer"—they need verifiable, traceable proof of where a concept originates in their coursework, paired with targeted feedback that genuinely moves the needle on exam performance.
What's next for Edify
- Expanding support for mixed-language content and regional Indian-language interactions (such as English lectures paired with native Hindi explanations).
- Implementing fully audio-driven, conversational tutoring sessions.
- Integrating an automated, interactive visual course flow map to display topic prerequisites graphically.
Built With
- adaptive-learning
- bayesian-knowledge-tracing-(bkt)
- computer-vision
- css3
- fastapi
- google-gemini-api
- html5
- item-response-theory-(irt)
- javascript
- multimodal-ai
- numpy
- pydantic
- pymupdf
- pytest
- python
- ragas
- rest-api
- retrieval-augmented-generation-(rag)
- scikit-learn
- three.js
- uvicorn
- vector-embeddings
- web-speech-api
- webgl
Log in or sign up for Devpost to join the conversation.