Keep humans in the decision loop
Assisted evaluation is separate from final judgement. Moderation makes disagreements visible and preserves a review trail.
Research2025–26
Drishti.
Make every mark traceable.
A human-in-the-loop marking platform connecting answer-sheet intake, AI-assisted evaluation, moderation, and QR-verifiable records.
- MY CONTRIBUTION
- Research writing & capture-system architecture
- CONTEXT
- Team Shiny · IIT Patna
- STATUS
- Software demo · research in progress
WORKFLOW / DRISHTI
- 01Capture + intake
- 02Assisted evaluation
- 03Human moderation
- 04Audit trail
Make every mark traceable.
01THE PROBLEM
What needed solving?
Answer evaluation needs readable capture, correct script-to-record mapping, and a way to review differences between human and AI assessments.
02MY CONTRIBUTION
Where I came in.
My work covers the research manuscript, literature review, cost modelling, and an ESP32-S3 CAM capture-rig architecture. The optical design considers lens calibration, lighting, and capture quality before evaluation.
Team repository03KEY DECISIONS
The thinking behind it.
Start with capture quality
Unreadable input undermines every later step. The capture design considers optical budget, calibration, and dual-LED illumination.
Separate modelling from validation
Analytical estimates guide the prototype but require empirical evaluation before operational claims.
04OUTCOME & NEXT STEPS
Built. Learned.
Still improving.
Research manuscript in preparation, with a documented capture architecture and evaluation workflow. Team and upstream credit are preserved.
Where it stands today
The software works with uploaded or browser-captured answer sheets. The repository documents the physical camera provider as a planned, uncertified integration; a separate team post describes a hackathon capture-system build. The manuscript is in preparation, and measured grading accuracy and hardware performance remain to be validated.
What comes next
Evaluate capture quality and review outcomes using an appropriate, permissioned test dataset.