AutoGrading

AI-assisted English answer grading prototype.

This repository is a public demo version of a local tool built to reduce repetitive work in English answer correction. It uses AI to prepare OCR text, draft scoring comments, and write the result back onto a PDF. The final judgment is always left to a human grader.

What It Does

answer PDF -> OCR draft -> human check -> AI grading draft -> PDF annotation

• Extracts handwritten answer text from PDFs with Gemini.
• Drafts scores and feedback with Claude from a structured master JSON.
• Writes scores and comments back to the answer PDF with PyMuPDF.
• Provides a coordinate picker for mapping PDF answer areas.

Important Notes

• This is a learning/prototype repository, not a production grading service.
• The included data is a small demo set. Real exam materials, real answer sheets, and private grading records are intentionally excluded.
• AI output should be reviewed by a human before being used.

Setup

python -m venv .venv
.\.venv\Scripts\activate
pip install -r requirements.txt
copy .env.example .env
copy config.example.json config.json

Then set the API keys in .env:

GOOGLE_API_KEY=...
ANTHROPIC_API_KEY=...

Usage

Create these local folders when needed:

inputs/
step1_texts/
step2_json/
step3_final/
done/

Run OCR:

python step1_mark_and_text_v2.py

Review and edit the generated step1_texts/*_draft.txt, then run grading and PDF output:

python step2_and3_combined.py

To register coordinates for a new answer sheet layout:

python coordinate_picker.py

Repository Shape

data/
  coord_db/       demo coordinate data
  masters/        demo scoring master JSON
  prompts/        grading system prompt
  rubric_txts/    demo rubric text
step1_mark_and_text_v2.py
step2_and3_combined.py
coordinate_picker.py
preprocessing.py

Runtime folders such as inputs/, done/, step1_texts/, and step3_final/ are ignored by Git.