Projects
Register image URLs, run character detection and single-character recognition inside your browser, and proofread the results together.
Sign in with Google to startImages are never stored
Only the image URL, its dimensions and a hash are recorded. Inference runs entirely in the browser, so images are never sent to a server.
Mistakes are kept, not deleted
A "this is not a kaō" verdict is recorded with its reason. Systematic false positives — cursive signatures, show-through from the reverse side — are exactly what a detector needs to learn from.
Corrections become training data
Every edit is recorded with who made it. Human-verified annotations become the ground truth for the next model.