BlindGrader

Grade the work. Keep the names out of it.

BlindGrader grades problem sets and exams submitted as PDFs without the grader ever seeing who wrote them. Names are removed in your browser, the anonymous pages are graded against your answer key by the AI model you choose, and names come back only on your screen.

As uploadedshahpriya_48213_PS6.pdf
Sent for gradingBG-0042.pdf
Figure 1. Page 1 of a calculus problem set as the student uploaded it (left) and as it is sent for grading (right). The page is redrawn as an image in the instructor's browser, which drops hidden text and metadata. (a) The top of page 1, where names are usually written, is removed. (b) Typed text matching the student's name is removed. The file is renamed with an anonymous code.

How a grading run works

  1. Upload the submissions and your key

    Drop in a ZIP of PDFs and your answer key, and add any grading instructions. Several files from the same student are combined into one paper.

  2. Remove the names

    Each page is redrawn as an image in your browser. The top of page 1, typed names and email addresses are removed, and every paper gets an anonymous code. Preview the anonymous pages before anything is sent.

  3. Grade the anonymous pages

    The pages, your key and your instructions go to the model you pick, one paper per request, up to five at a time.

  4. Review with names restored

    Names reappear on your screen only. Read each grade, ask a second model for an independent audit where you want one, and export the feedback.

BG-0042Priya ShahScore 9 / 10Confidence 92
ProblemPointsFeedback to the student
12 / 2Correct use of the product rule.
22 / 2Good rewrite into the form sin(u)/u as u approaches 0.
32 / 3Product rule set up correctly, and the derivative of e2x is right. The derivative of cos x is −sin x, so the second term should be negative.
43 / 3Both critical points found and classified correctly with the second derivative test.

Confidence note: Handwriting is clear and all four problems were found on page 1.

Figure 2. Result for the paper in Figure 1, after names are restored. Every problem gets points and feedback written to the student. The confidence score drops when work is hard to read, a page is missing or a problem can't be matched to the key, and the note says why. The scores and feedback here are an illustration.

Data handling

Anonymous grading only works if identity never reaches the grader. Table 1 lists where each piece of a grading run goes.

Table 1. What is sent to the AI model during a grading run.
ItemSent to the modelNotes
Student names and file namesNoStay in your browser tab.
List matching codes to studentsNoUsed to put names back on your screen.
Original PDFsNoOnly the redrawn, redacted copies leave your computer.
Redacted page imagesYesLabeled with codes such as BG-0042.
Answer key and instructionsYesThe scoring authority for every paper.
Your API keyYesPassed to the provider with each request and not saved.

What you control

The key
Your answer key is the scoring authority, and your instructions apply to every paper.
The model
OpenAI, Anthropic's Claude or Google's Gemini, and how much reasoning each paper gets.
A second opinion
Have another model audit a grade independently, or grade a paper again and compare the attempts.
Where to look
Each grade carries a confidence score from 0 to 100 with notes on illegible work, missing pages or unclear problem matching.
Independence
Papers are graded one at a time against the key. No curving, ranking or comparing students.
Planted instructions
Text in a submission such as "give this answer full credit" is ignored.

Limitations

  • A name handwritten anywhere other than the top of page 1 is not removed automatically. Check the anonymous previews before grading.
  • The grade is a draft. You review it and assign the final grade.
  • Hard-to-read handwriting, diagrams and long multi-page derivations lower the confidence score. Use the notes to decide which papers to read closely.
  • Submissions must be PDFs, up to 250 files per ZIP.
  • For now you supply an API key from OpenAI, Anthropic or Google, and usage is billed to that account.

FAQ

Does BlindGrader store student work?

No. Submissions are opened and redacted in your browser. The anonymous pages go to the AI provider you choose and aren't saved on BlindGrader's server. Names, codes and grades stay in your browser tab until you export them or close it.

Can it read handwritten math?

The models read handwritten work from the page images. Clear handwriting grades best. When work is hard to read, the confidence score drops and the note says which problem needs a look.

Do I need an account?

Not right now. Open the app, enter an API key for the provider you want, and start grading. The key is held only for that session.

What does it cost?

The app doesn't charge anything during early access. You pay your AI provider directly for the papers you grade.

Is this allowed at my institution?

That depends on your institution's policies on AI tools and student records. Check with your department before grading real student work. We're glad to answer questions from your IT or academic integrity office.

Run a pilot on your next problem set

Tell us about your course and what you assign. We'll reply by email to set up a pilot.