How MathSense AI grades handwritten math
MathSense AI reads each student's handwritten work from a scan or photo, matches it to the questions on your assessment, scores each question against its point value with partial credit, checks the arithmetic, drafts feedback, and puts every score and comment in front of you to edit and approve before anything reaches a student.
Updated October 1, 2026 · Published by MathSense AI (L8O LLC)
The grading pipeline, step by step
1. It reads your assessment first
You upload the question paper as a PDF or image. MathSense AI finds each question and sub-question, reads the point value printed next to it (such as "(3)" or "[5 marks]"), records the options of multiple-choice questions, and crops any printed figure so it is shown with its question. You can correct any point value before grading.
2. It transcribes each student's handwriting
You upload one file per student: a PDF, PNG, JPG or WebP — a scan or a phone photo. The handwriting is transcribed line by line in reading order, including equations. Printed text and red teacher marks are ignored, and circled or crossed-out multiple-choice answers are recorded as such. A second pass checks the transcription against the page, faint pencil is contrast-enhanced, and a question is only marked "not attempted" after a second look confirms the space is blank.
3. It matches every answer to the right question
Students skip questions, answer out of order and continue on the back of a page. Each piece of work is paired with its question using the question labels, page continuity, sub-question parts and order on the page. When a label is ambiguous, the match is flagged for you rather than guessed.
4. It scores each question against its points, with partial credit
Each answer is judged on a 0–10 scale and converted to that question's point value, so a correct method with a slip earns part of the marks instead of zero. Once a paper is graded, its result is kept: opening it again shows the same grades.
5. It checks the arithmetic
Alongside the AI judgement, a separate rule-based check re-computes purely numeric lines — including order of operations — so a step like "2 + 3 × 4 = 20" is caught and named precisely in the feedback.
6. It drafts feedback and an annotated paper
For each student it drafts feedback that points to what was done well and where the reasoning broke, and it can produce an annotated copy of the paper with the comments placed beside the work. You can print both.
7. You review and approve every grade
Nothing reaches a student until you approve it. You can change any question's points and edit any comment, then approve the paper.
What partial credit looks like
Two worked examples of how a question is scored and what the feedback says. They show the method, not measured results.
Illustrative example — simulated student work.
Question (5 points): Solve 3(x − 2) + 4 = 19.
Student work
3x − 6 + 4 = 19
3x − 2 = 19
3x = 17
x = 17/3
Score: judged 8/10 → 4 of 5 points
Drafted feedback: You distributed the 3 correctly and combined −6 + 4 into −2. On line 3, adding 2 to both sides gives 3x = 21, not 17 — so the final answer should be x = 7. Your method is right; check the step where you move a term across.
Illustrative example — simulated student work.
Question (2 points): Evaluate the limit as x → 2 of (x² − 4) / (x − 2).
Student work
(x − 2)(x + 2) / (x − 2)
= x + 2
= 2 + 2 = 5
Score: judged 5/10 → 1 of 2 points
Arithmetic check: 2 + 2 = 5 is incorrect (2 + 2 = 4).
Drafted feedback: Factoring and cancelling (x − 2) is exactly the right move, and you correctly reduced the expression to x + 2. The last line has an arithmetic slip: 2 + 2 = 4, so the limit is 4.
What you control
- The point value of every question, before grading.
- Every score and every comment, after grading.
- What students receive and when: you approve each paper, then share the feedback as a printout, a download or an annotated copy of the paper.
Student work is used to grade it, and it is not used to train AI models.
How we test it
- Answer-to-question matching is tested on a fixed set of synthetic and real student papers on every release. On the synthetic papers, 31 of 31 answers are matched to the right question; on the real out-of-order papers, every answered question is matched correctly and none wrongly, with ambiguous labels flagged for the teacher.
- Grading time was measured on a multi-page calculus test in internal testing (July 2026): about 80–90 seconds per student paper, end to end.
- Most of our test papers are calculus, precalculus and algebra, at high-school and AP level.
These are internal tests on a small, fixed set of papers. We will publish a larger accuracy study with its method once enough graded work has been reviewed by teachers.
Where to review closely
- Very faint, smudged or cramped handwriting.
- Work that continues somewhere unexpected on the page without a label.
- Questions that depend on a diagram, and any work students draw rather than write: printed figures are shown with each question, but drawn work and proofs are not yet a tested part of grading.