The problem
Marking handwritten exam scripts is slow, repetitive and inconsistent between markers. The interesting difficulty is not reading the handwriting — it is deciding whether an answer is correct when it is phrased differently from the model answer.
What it does
The pipeline segments scanned answer sheets, recognises the handwriting, and scores each response against a reference using language models rather than string matching. Vision handles extraction; language handles judgement.
The part that mattered
Separating those two stages. Recognition errors and grading errors have different causes and different fixes, and keeping them apart makes both measurable.