EduAdapt AI

Inspiration

Students learn from many different sources—textbooks, PDFs, presentation slides, lecture videos, and notes. However, these resources are often disconnected, making it difficult to find the right explanation, revise weak topics, and know what to study next.

We wanted to build an AI tutor that does more than simply answer questions.

EduAdapt AI was inspired by the idea of creating a learning companion that understands a student's actual course material, explains concepts using those sources, identifies knowledge gaps, and continuously adapts to the student's learning progress.

Our vision was to turn scattered learning resources into a personalized, source-grounded learning experience.

What it does

EduAdapt AI is a multimodal AI tutoring and adaptive learning platform.

Students can provide course materials such as:

📄 Textbooks and PDFs
📊 Presentation slides
🎥 Lecture videos
🖼️ Images, diagrams, and figures

The system processes these materials and builds a structured knowledge base containing topics, subtopics, concepts, prerequisites, and source references.

Students can then ask questions through the AI tutor. Instead of relying only on general knowledge, EduAdapt AI retrieves relevant course content and provides source-grounded answers with citations to the corresponding page, slide, or video timestamp.

The platform also provides:

Adaptive quizzes based on the student's current knowledge
MCQs, short-answer, and numerical questions
Topic and difficulty tagging
Question validation and duplicate detection
Diagnostic assessments for new students
Topic-wise mastery tracking
Weak-topic and misconception detection
Personalized learning recommendations

As students interact with the system, their learner profile is updated and future questions are adapted accordingly.

How we built it

We built EduAdapt AI as a modular AI-powered web application.

Multimodal Knowledge Processing

Different learning formats are processed according to their content:

PDF → text, pages, images and figures are extracted.

PPT → slide text, diagrams and visual information are processed while preserving slide references.

Video → speech is converted into timestamped text so that students can trace explanations back to the lecture.

The extracted content is divided into meaningful chunks and enriched with metadata such as topic, concept, page number, slide number, and timestamp.

Source-Grounded AI Tutor

We use a Retrieval-Augmented Generation (RAG) pipeline.

When a student asks a question:

Question → Embedding → Relevant content retrieval → Context verification → AI response → Source citation

This allows the tutor to connect its answer back to the student's actual learning material.

Queries that are not sufficiently supported by the uploaded material are flagged instead of being presented as course-grounded answers.

Adaptive Assessment

Questions are generated from the course knowledge base and tagged with:

Topic
Difficulty
Question type
Source
Source location

Generated questions go through validation and duplicate checking before being used for assessment.

Learner Model

The system maintains topic-wise mastery for each student.

After every assessment, the student's performance is used to update their mastery profile. The next learning activity can then focus on topics where the student needs more practice.

The application is built using Python, Django, JavaScript, HTML/CSS, SQL, embeddings, vector search, and LLM-based AI components, along with multimodal processing and evaluation tools.

Challenges we ran into

One of our biggest challenges was handling multiple types of learning material in a single knowledge system.

A PDF, presentation, and lecture video contain information in very different formats. We needed to extract useful information while preserving its original location so that the AI could provide meaningful citations.

Another challenge was preventing hallucinations. A normal chatbot can answer questions using information outside the student's course. For an educational system, this can be confusing and difficult to verify. We therefore focused on grounding responses in retrieved course content and identifying unsupported queries.

We also faced challenges in:

Extracting useful information from diagrams and figures
Maintaining page, slide, and timestamp references
Generating reliable assessment questions
Preventing repetitive questions
Identifying student misconceptions
Updating mastery after every interaction
Making the learning experience genuinely adaptive rather than simply conversational
Evaluating the quality of both retrieval and generated answers

Accomplishments that we're proud of

We are proud that EduAdapt AI brings several components of an intelligent learning system together instead of building just another chatbot.

Our key accomplishments include:

Multimodal course ingestion from documents, slides, videos and visual content
Structured course knowledge base with topics, concepts and prerequisites
Source-grounded AI tutoring with page, slide and timestamp references
Unsupported-query handling to reduce ungrounded answers
Adaptive assessment generation based on course content
Question validation and duplicate detection
Topic-wise learner mastery tracking
Diagnostic assessment for new learners
Personalized recommendations based on weak topics
AI evaluation using measurable retrieval and response-quality metrics

Most importantly, we focused on creating a complete learning loop:

Learn → Ask → Assess → Analyze → Adapt → Learn Again

What we learned

Through this project, we learned that building an educational AI system requires much more than integrating an LLM.

We gained practical experience with multimodal data processing, Retrieval-Augmented Generation, embeddings, vector search, source grounding, prompt design, assessment generation, learner modeling, and AI evaluation.

We also learned the importance of maintaining traceability. For educational applications, an answer becomes much more useful when a student can immediately see where the information came from.

Another important lesson was that personalization should be measurable. Instead of simply claiming that the system is adaptive, we designed the system so that student performance can influence mastery estimates, question selection, and future recommendations.

What's next for EduAdapt AI

We want to evolve EduAdapt AI from a hackathon prototype into a more complete personalized learning platform.

Our next improvements include:

🧠 More advanced learner models such as Bayesian Knowledge Tracing
🔄 Spaced repetition and forgetting-curve-based revision
🗺️ Interactive prerequisite and course-flow maps
📚 Automatically generated revision notes and study materials
📅 Personalized exam preparation schedules
🌐 Multilingual tutoring, especially for Indian languages
🎙️ Voice-based AI tutoring
🖼️ Better understanding of complex educational diagrams
👩‍🏫 Instructor dashboards and learning analytics
📈 Long-term learning progress and mastery analysis

Our long-term goal is to make EduAdapt AI a learning companion that doesn't just answer what students ask, but understands what they know, what they don't know, and what they should learn next.