GradeMIND

GradeMIND is a handwritten answer-sheet grader, built for a national hackathon.

Marking is a trust problem. A mark has to be explainable and repeatable, and a language model's numeric output is neither.

The model reads. Arithmetic decides. Every mark traces to a criterion ID, an evidence span and deterministic arithmetic, and never to an LLM numeric output.

This is the same deterministic core and LLM language layer split used across PRYSM and AHAL AI, applied to grading.

Noisy handwritten scans go through three OCR engines and Gemini Vision. Scores are computed by code.

• Three OCR engines fused by confidence voting, with Gemini Vision for noisy handwritten scans
• Question segmentation and annotated-PDF output tracing every mark to an evidence span
• Deterministic scorer with reproducibility tests
• Atomic job-state persistence with resume and cache reuse
• Human-override preservation

• Keeping numeric authority out of the model while still using it to read.
• A production-readiness audit exposed fabricated agent results and a score-desync bug. A six-phase remediation plan now gates real student data behind human review.

Built for a national hackathon. Next step: CBSE-grade production.

3

OCR engines fused

Verify against real state, not summaries. The audit only mattered because it checked what the agents had actually produced.