A teacher friend of mine showed me her workflow for entering grades: a stack of photographed paper sheets, a spreadsheet, and a lot of manual typing. Every term, the same repetitive work.
I asked if there was a tool that could just read the sheet and fill in the spreadsheet for her. There wasn't anything good — most "AI table extraction" tools were either too generic, too expensive, or required a level of setup that defeated the purpose.
The first version
The first version of Pixel2Excel was deliberately narrow. One job: upload an image of a grade sheet, get back an editable table. No accounts, no dashboards, no settings page.
Upload image → Gemini API extracts table → Editable grid → Export
That constraint was the point. I wanted to validate whether the core extraction was good enough to be useful before building anything else around it.
What I learned from real sheets
Real grade sheets are messier than test images. Handwriting varies, lighting is inconsistent, and column headers aren't always where you'd expect. Most of the iteration since launch has been about handling that messiness — better preprocessing, clearer error states when extraction confidence is low, and an editable interface that makes corrections fast rather than frustrating.
What's next
Right now I'm working directly with a handful of teachers, refining accuracy based on the sheets they actually use. The roadmap is intentionally short: get extraction right for the formats people actually have, then think about what comes after.
If you're a teacher dealing with this exact problem, I'd love to hear from you — reach out.