1¢, 1min, 100%: AI in University Exams (In Development)

About the project

Open-source research on AI applied to Business Administration exams, with public results, Mathpix OCR, and structured exam processing with SQLite and Prisma.

Technologies

Features

• Opensource research: results and methodology publicly available
• Models evaluated: DeepSeek API, xAI API, OpenAI API, Google AI API
• OCR system implemented with Mathpix API for exam digitization
• Automatic exam conversion program to LATEX/Markdown format
• SQLite database with Prisma for exam and results management
• Accuracy analysis: 100% success rate on multiple-choice exams
• Cost optimization: ~1 cent per complete exam
• Time optimization: resolution in less than 1 minute
• Dataset of Business Administration exams from University of Murcia
• Performance comparison between different LLM architectures
• Analysis of implications for academic assessment
• Results graphs and visualizations

Technical challenges

• Integration of Mathpix API for OCR adapted to university exam format
• Creation of custom parser to convert exams to LATEX/Markdown
• Collection and structuring of exam dataset
• Effective prompt design for different types of questions
• Results validation and quality control
• Comparative analysis between multiple AI models and architectures
• Efficient data management with SQLite and Prisma
• Interpretation of educational implications of results

Learnings

• Integration of Mathpix API for OCR in academic documents
• Development of academic document processing pipelines
• Capabilities of advanced reasoning models (OpenAI API, xAI API)
• Cost-effectiveness ratio in educational AI applications
• Research methodology design with AI
• Quantitative analysis of LLM performance
• Database management with SQLite and Prisma ORM
• Ethical and practical implications of AI in education

Screenshots

Graph 1

Graph 2

Graph 3