Grade Sheet OCR — Stage 1 Probe

Quick test to see if a local vision model can read handwritten Portuguese grade sheets.

Setup (one time)

1. Install Ollama — download from https://ollama.com/download, drag to Applications, launch it.

2. Pull a vision model:

   ollama pull minicpm-v

   Alternative (larger, often more accurate, slower):

   ollama pull llama3.2-vision

   If you switch model, update MODEL at the top of test_ocr.py.

3. Install poppler (system dep for PDF rendering):

   brew install poppler

4. Install Python deps:

   pip install pdf2image ollama

Run

python test_ocr.py "/path/to/turma.pdf"

The script will:

• Render each page of the PDF to an image
• Send each image to the local vision model
• Print the raw model response and a parsed JSON summary at the end

What to look for

This is a probe, not the final app. Read the output and judge:

• Are the names mostly correct?
• Are the grades close to the truth (compare to the original PDF visually)?
• How wrong is "wrong" — off by 1 digit, or completely hallucinated?

If accuracy is acceptable → continue to Stage 2 (Excel output, multi-page tabs). If accuracy is poor → try llama3.2-vision, or pivot to cloud API.

Files

• test_ocr.py — the probe script
• README.md — this file