Fixed-layout, AI-assisted questionnaire digitization for one 34-question booklet.
The supplied questionnaire is a 12-page physical booklet scanned as six
two-page spreads. Keep reference_master.pdf one directory above this project
folder (D:\bihar-assessment\reference_master.pdf). The processor splits every
incoming six-page PDF into twelve logical pages in this order: 12|1, 2|3,
4|5, 6|7, 8|9, and 10|11. It then aligns each logical page to the
master before generating answer crops.
python -m venv .venv; .\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
$env:GEMINI_API_KEY = "..." # optional; deterministic local extraction still runs
python main.py process --input C:\scans --output C:\results
streamlit run core/review/streamlit_review.py -- --run-dir C:\resultsThe calibrated answer regions are saved once in config/pages.json; they are
not edited for individual student PDFs. The system deliberately supports only
this format. It records intermediate crops and JSON audit records in the output
directory and can resume completed PDFs.