Problem
Manual grading is slow and inconsistent when submissions include both typed and handwritten responses.
NLP & ML Research · 2023Video demo
Autograder: Precision in Every Grade
Semi-automated grading system with BERT-based NLP and OCR to evaluate typed and handwritten responses.
Approach
I combined semantic similarity, OCR extraction, and a grading workflow UI to support semi-automated evaluation.
Result
The project demonstrated how NLP and OCR can reduce repetitive grading effort while still keeping human review in the loop.
Step 1 of 5: Typed or handwritten answer
Highlights
- Led a 6-person team to build an auto-grading pipeline (BERT-uncased, spaCy, PyTesseract) with a React front-end and Flask backend, enabling semi-automated grading for handwritten and typed responses.
Design
The main idea was not to remove human evaluation completely, but to shorten the repetitive parts of grading.
Stack choices
- BERT-based semantic comparison for meaning-aware answer matching.
- OCR for scanned or handwritten text extraction.
- React + Flask for an end-to-end grading flow.
Practical constraint
The system worked best as a reviewer assistant, not as a fully autonomous grader.