Applications
Handwritten Exam Grader
Handwriting to text to rubric-based grading suggestions, with a human in the loop.
This prototype takes scanned handwritten exam scripts through text recognition, a rubric and AI-suggested marks that a person reviews. It is a prototype, and student material stays private.
The problem
Marking handwritten scripts is slow, and consistency suffers over a large batch. The first obstacle is not grading but reading: scripts must become text before any rubric can be applied.
What I built
The Streamlit prototype is about 3,917 lines and runs in eight steps from handwritten script to report: text recognition with Azure Document Intelligence, a rubric, grading suggestions from a language model, and reports. The model provider can be switched between Azure OpenAI and a local Llama 3.1 through Ollama. The grading output is a suggestion for a person to review, not a final mark. A product version was started with FastAPI, SQLAlchemy, Next.js and Microsoft Entra sign-in.
This is not a Claude or small-model build. The default grader is Azure OpenAI, with a local Llama 3.1 option.
How it performed
I have no accuracy figure, such as agreement with human markers, and no deployment outcome is recorded. Student scripts are confidential and I do not use them here.
What I learned
Keep a person between the model and the mark. Recognition errors in handwriting pass silently into grading, so a reviewer needs to see both the read text and the suggested score. The unanswered question is how often the machine and a careful human marker agree. Until that is measured, this is a demonstration of the workflow, not evidence it works.