Bubblesheet AI
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
- SwarmieFounder validation through grounded AI swarm simulation. Paste a pitch, watch 500 AI personas react like a Reddit thread, and walk away with the three real objections before you burn a single user interview.
- LearnAIA free map of all of AI, from linear algebra to agents. 14 fields, 152 steps, 162 problems. Every step is something you read, then something you build. Finish it and a star lights on your map.
- PitchcraftA pipeline that turns a CSV of local businesses into deployed prototype websites, audit PDFs, and growth proposals, ready to send. One operator, a batch of leads, a folder of finished deliverables.