VisionAI

Inspiration

The inspiration came from the need for a truly offline AI-powered study tool that respects user privacy while providing advanced learning capabilities. With students often lacking reliable internet access and growing privacy concerns, we wanted to create a personal AI tutor that runs entirely on local hardware without compromising functionality.

What it does

AI FlashCards transforms any PDF into an interactive learning experience using local AI models. It combines advanced PDF annotation (7 types including highlights, notes, shapes) with dual AI models - gpt-oss:20b for text understanding and llava:13b for visual content analysis. The AI can see your annotations and prioritizes content you've highlighted, generating smart flashcards that test understanding rather than memorization. Everything runs 100% offline after initial setup.

How we built it

We built it using Spring Boot (backend) with Spring AI for Ollama integration, JavaFX for the desktop UI, and Apache PDFBox for PDF processing. The architecture uses optimized prompt engineering (2000-4000 char windows) to ensure fast local inference. We implemented continuous PDF scrolling with real-time annotation layers and integrated two AI models through Ollama's REST API for comprehensive text and visual analysis.

Challenges we ran into

JavaFX Threading: Crashes due to improper thread management between Spring Boot and JavaFX
Memory Management: Handling large AI models (40GB + 8GB) efficiently on consumer hardware
FXML Injection: Field naming mismatches causing null pointer exceptions in flashcard generation
Zoom Calculations: Incorrect math causing Fit Width/Page buttons to compound zoom levels
Native Access Warnings: JVM restrictions requiring specific flags for JavaFX on newer Java versions

Accomplishments that we're proud of

Created a fully functional offline AI study assistant combining two large language models
Implemented 7 different PDF annotation types with real-time AI awareness
Achieved 2-5 second AI response times with intelligent prompt optimization
Built continuous PDF scrolling with smooth performance
Made AI-powered education accessible without internet connectivity

What we learned

Local AI models can provide powerful functionality without cloud dependencies
Prompt engineering is crucial for performance - reducing context from 10K to 2-4K chars dramatically improves speed
JavaFX and Spring Boot require careful thread management to work together
Multi-modal AI (text + vision) significantly enhances educational applications
User annotations provide valuable context that AI can leverage for personalization