NovaLearn AI

Inspiration

Students often study using a combination of PDFs, lecture slides, textbooks, images, videos, and notes. Finding relevant information across these resources and knowing what to study next can be difficult. We wanted to build an AI study companion that does more than provide generic answers — one that understands the student's learning material and adapts to the student's progress.

This inspired us to build NovaLearn AI, a personalized learning platform that combines multimodal content processing, source-grounded tutoring, adaptive assessments, and learner progress tracking.

What it does

NovaLearn AI turns learning resources into an interactive and personalized study experience.

Students can upload course materials such as PDFs, PowerPoint presentations, images, videos, and transcripts. The system processes the material, extracts content with its source information, organizes it into topics and concepts, and makes it searchable.

The AI tutor answers questions using retrieved content from the student's course material and provides source references. When the uploaded material does not support an answer, the system can identify that the information is not covered instead of simply guessing.

NovaLearn AI also generates quizzes, mock exams, and diagnostic assessments. Questions are associated with topics, difficulty, answer keys, explanations, and source references. The system tracks performance and uses a Bayesian Knowledge Tracing model to estimate topic-level mastery.

Based on the learner's performance, the platform can recommend topics to revisit, targeted practice, revision, and future assessments.

How we built it

We built NovaLearn AI as a full-stack web application with a Python backend and modern web frontend.

The content pipeline follows:

Upload → Validation → Inspection → Text/OCR extraction → Chunking → Embeddings → Indexing → Topic organization → Retrieval

The system supports multimodal learning resources including PDFs, PowerPoint files, images, videos, and transcripts. PDF and image content can be processed using OCR, while video workflows can use timestamped transcripts.

For retrieval, we implemented hybrid search using PostgreSQL full-text search and pgvector similarity search. Retrieved passages are then used by the source-grounded tutor.

The learning system includes verified question generation, deterministic grading, topic-level performance tracking, and Bayesian Knowledge Tracing for estimating learner mastery.

We also implemented evaluation workflows for retrieval, grounded answers, refusal of unsupported questions, and simulated personalization.

Challenges we ran into

One of our biggest challenges was connecting different types of learning resources while preserving their original source information. A useful AI answer is not enough for an educational system — students also need to know where the information came from.

We also had to make assessment generation reliable. Generated questions need correct answers, meaningful difficulty, appropriate distractors, and valid source references. We therefore added verification and deterministic checks before questions are offered to students.

Another challenge was personalization. Instead of assuming that every student starts with the same knowledge level, we implemented topic-level mastery estimates and used those estimates to influence future learning recommendations.

Accomplishments that we're proud of

We are proud of building and deploying a working end-to-end prototype rather than only a chatbot.

NovaLearn AI combines multimodal content ingestion, source-grounded tutoring, adaptive assessments, learner modeling, personalized recommendations, revision support, analytics, and evaluation tools in one platform.

We also focused on transparency: learner mastery values are presented as estimates, recommendations store the reasoning behind them, and tutor citations point back to the underlying learning material.

What we learned

Through this project, we learned that building an educational AI system requires more than connecting an LLM to a chat interface.

We learned about multimodal document processing, OCR, embeddings, vector search, hybrid retrieval, retrieval-augmented generation, source attribution, assessment verification, Bayesian Knowledge Tracing, and evaluation of AI systems.

Most importantly, we learned that personalization and reliability must be designed into the system rather than added after the AI is built.

What's next for NovaLearn AI

Our next goal is to improve the learner model using larger amounts of real learning interaction data and stronger evaluation datasets.

We plan to improve multilingual learning support, richer video understanding, personalized study schedules, spaced repetition, and visual learning resources.

We also plan to perform a carefully reviewed benchmark evaluation using course-specific questions and source labels, allowing us to measure retrieval quality, grounding, refusal accuracy, and personalization more rigorously.

Built With

• adaptive
• ai
• bcrypt
• bkt
• docker
• fastapi
• ffmpeg
• gemini
• grok
• hybrid
• javascript
• jwt
• ml
• ocr
• pgvector
• postgresql
• python
• rag
• react
• search
• tailwind
• tesseract
• typescript
• vector
• vite