Answer Lens
Explainable AI grading for handwritten exam answers
A government-exam coaching institute evaluated thousands of handwritten mock-test answers by hand. Grading was slow, inconsistent between reviewers, and students got little feedback.
The challenge
- Handwriting quality varied widely across scanned answer sheets.
- Each answer had to be scored against a multi-criteria rubric with a justification a student could learn from.
- Performance had to be tracked across successive mock tests.
What we built
OCR tuned for handwriting
A pipeline combining OCR engines with a multimodal model for the pages classic OCR struggled with, plus confidence scores to route hard cases to reviewers.
Rubric-based agentic evaluation
LangGraph agents score each criterion separately and produce a written justification, then a supervisor step reconciles the criteria into a final mark.
Consistency by design
Calibration sets and evaluator prompts were validated against human graders before rollout.
Analytics for learning
A Postgres analytics layer tracks each student's criteria-level trends across tests.
What we learned
- Scoring criteria independently and reconciling afterwards was more consistent than asking for one holistic mark.
- Confidence-based routing kept human reviewers focused on genuinely hard pages.
- Students valued the justification more than the mark.
Want something like this?
Most of these projects started as a two-page brief. Send us yours.