A
Source
100%
Research preview
Faithful math OCR
Your handwriting,
without the model’s corrections.
Upload a handwritten page. Proofline drafts editable LaTeX and keeps the symbols, steps, and even mistakes that are actually on the page.
This public demo uses deterministic placeholder output and keeps uploads in your browser. A live model backend is not connected.
Source ink
Faithful output
f(x) = x2 + 3x
f′(x) = 2x + 2
x = −1
The visible “+2” stays “+2.”
Transcription, not grading. You decide what is correct.
Editable by design. Review every line before export.
Purpose-built. Math fidelity is the only job.
The transcription desk
Ink in. LaTeX out. You stay in control.
Demo runtime
Placeholder output
-
1
UploadPDF or image
-
2
TranscribePage by page
-
3
ReviewEdit and export
Drop handwritten work here
PDF, PNG, JPG, or WebP · up to 20 MB
or choose a file from your device
Reading mode
B
Live
Editable LaTeX
C
Ready
Rendered proof
Illustrative output. The connected GLM model will replace this sample response.
Evaluation method
Measure transcription, not problem solving.
The public repository ships synthetic fixtures only. They test whether a model copies visible notation faithfully, keeps long derivations consistent, and preserves an intentional source error instead of silently fixing it.
Private academic records are not part of the public dataset. The original development corpus and derived raw outputs were removed. New public scorecards should be generated from the synthetic fixtures.
Read the data policyPreserve, don’t solve
The target is what the writer put on the page—not what the answer should have been.
Hold notation steady
Names, signs, indices, and inequality directions should survive a long derivation.
Make review visible
The source, LaTeX, and rendered proof stay together so errors are easy to catch.