Paper Blinder

Blind your papers locally. Grade them however you want.

A local web app that takes a bulk-scanned PDF of student work, redacts the name zone, splits it into individual coded submissions, and gives you a private answer key. Student data never leaves your machine.

Use it with any grading workflow — AI-assisted (Claude, ChatGPT, Gemini, whatever fits your style), human TAs, or grade them yourself with bias reduction. The blinding step is universal. What comes after is up to you.

Why?

Existing solutions couple anonymization to a specific grading engine. If that engine doesn't match your pedagogy, your students, or your standards — you're stuck.

Paper Blinder separates the two steps:

1. Blind once — redact names, split papers, generate codes
2. Grade with anything — your choice of tool, model, or human

This matters especially for:

• Teachers in high-needs districts working with handwritten papers
• Classrooms where students don't submit digitally through an LMS
• Anyone who wants to pick their own grading tool without being locked in

Features

• Drag-and-drop web UI — runs locally in your browser
• Two split modes — fixed page count, or landmark text detection
• Rotation-aware redaction — handles bulk scanner orientation quirks
• AI name detection — uses Qwen2.5-VL (local vision model via Ollama) to read handwritten names for your answer key
• Text layer fallback — if your scanner embeds searchable text, it uses that first
• Private answer key — CSV with detected names, you verify and fill in the rest
• 100% local — nothing touches the internet. FERPA-safe by design.

Setup

Required

pip install flask PyMuPDF

Optional (for AI name detection)

Install Ollama, then:

ollama pull qwen2.5vl:7b

This runs a local vision-language model that reads handwritten names off the scanned papers. Without it, the tool still works — you just fill in names on the answer key manually.

Hardware: Any machine with 8GB+ VRAM can run the 7B model. Tested on NVIDIA RTX 5080 (laptop) and RTX 5090 (desktop).

Usage

python paper_blinder.py

Opens your browser to http://localhost:5757. Then:

1. Drag in your bulk-scanned PDF
2. Pick your split mode:
   • Fixed page count: "every 3 pages is one student"
   • Landmark text: splits wherever a phrase like "Part 1" appears
3. Set the redaction zone (which edge, what percentage)
4. Hit Blind the Papers
5. Download the ZIP

You get:

blinded_papers.zip
├── BLINDED_a3f7c2.pdf    # individual papers (name redacted)
├── BLINDED_9b1e04.pdf
├── BLINDED_e7c431.pdf
├── ...
└── _ANSWER_KEY.csv       # YOUR EYES ONLY

The answer key looks like:

code	submission_number	pages	detected_name	detection_method	confirmed_name
a3f7c2	1	pp. 1-3	Maria Garcia	vision-ai
9b1e04	2	pp. 4-6	Jaime R	vision-ai
e7c431	3	pp. 7-9		none

Fill in confirmed_name where the AI got it wrong or couldn't read it. Keep this file — it's your mapping.

Grading workflow

1. At school: Scan the stack on the bulk copier → one big PDF
2. At home: Run Paper Blinder → get coded PDFs + answer key
3. Grade: Upload blinded PDFs to your AI tool of choice (or grade by hand). The grader never sees student names.
4. Map back: Use your answer key to match grades to students

Tips

• Landmark phrases: If you add something like === START === to page 1 of each test, the tool auto-detects split points — no counting pages.
• Redact percentage: Default is 15% from the top. Adjust if your name field is lower or your header is bigger.
• Scanner rotation: If the redaction hits the wrong edge, try the edge selector dropdown. Bulk scanners sometimes store pages in unexpected orientations.
• No OCR needed: The tool works without Ollama — you just fill in names yourself. The vision model is a convenience, not a requirement.

Built with

• Flask — local web server
• PyMuPDF — PDF processing
• Qwen2.5-VL via Ollama — local vision AI for handwriting recognition

License

MIT — use it, share it, adapt it.

---

Made by a first-year teacher who needed to grade 75 papers without bias and without sending student names to the cloud.