Bubblesheet AI

livePython, YOLOv8, Flask

Grades multiple choice exams from ordinary scans. A YOLOv8 model reads every filled bubble, checks it against the answer key, and returns per student scores as a CSV.

0.536
mAP@50
6
Bubble classes
4
Interfaces

The problem and the build

Grading multiple choice exams by hand is slow and error prone, and classic OMR scanners need dedicated hardware and rigid pre-printed sheets. Bubblesheet AI replaces both with a trained object detector that reads plain scanned PDFs or images. Multi page PDFs are split into pages, a fine-tuned YOLOv8n model classifies each question's marked bubble as a, b, c, d, e, or none, and the results are scored against the key. It ships four ways: a CLI, a Flask upload app, a FastAPI JSON service, and a Gradio demo. The 6 MB weights live in a GitHub Release and download on first run.

Key insight

Publish honest numbers. The first model reaches 0.536 mAP@50 on a small dataset, and the README says so, along with where it fails: blank bubbles and the none class cause most errors. Knowing the weak spot is what tells you which data to collect next.

Stack

  • Python
  • YOLOv8
  • Flask
  • FastAPI
  • Gradio

More case studies

Need something like this in production?

Fixed price, 4 to 6 weeks. I reply to every inquiry within 24 hours.

Reply within 24 hours