Grade What Was Written: A Multi-Agent System for Grading Handwritten Calculus Solutions
Supervisor Name
Mohammed M Altahrawi
Supervisor Email
mtahrawi@ucas.edu.ps
University
University College of Applied Science
Research field
Artificial Intelligence
Bio
Description
AI tools are starting to be used to grade students’ handwritten math, but they have a hidden problem. When the AI reads a solution, it sometimes “fixes” the student’s mistake without telling anyone, so a wrong answer can end up getting full marks. Recent studies show this happens often with vision-language models. A recent system called CalcTutor found that most of its grading errors came from misreading handwriting, yet nothing in its pipeline checks the reading step. In this project we will build a grading system made of five AI agents that work together: A transcriber that converts each line of the student’s handwriting into LaTeX. An auditor that reads the solution a second time and flags lines where the transcription may have changed what the student wrote. A step verifier that checks each calculation (derivatives, integrals, algebra) using SymPy. A grader that applies the instructor’s rubric. A feedback agent that writes short, useful comments for the student. We will also collect our own dataset of handwritten calculus solutions from volunteer students. Part of it will contain solutions with deliberately planted errors, so we know exactly what was written. We will compare our system with a simpler CalcTutor-style baseline and measure how closely it agrees with human instructors and how often it hides student mistakes. The system uses only open-source models, so student data stays on campus.
