Vibgyor Technologies, Oct 2024 to now
AI exam grading engine
Grades answer sheets for 3+ exam boards and replaces hours of manual marking per batch.
# AI exam grading engine
Jay Gadhiya. Vibgyor Technologies, Oct 2024 to now. Software Developer, Full-Stack & AI.
Grades answer sheets for 3+ exam boards and replaces hours of manual marking per batch.
## Problem
Teachers spent hours per batch applying mark schemes by hand, and each exam board formats its schemes differently.
## Approach
- Parsed question papers and mark schemes directly from PDFs.
- Graded each answer sheet against per-criterion rubrics.
- Applied follow-through logic and mark dependencies, so one early error doesn't cascade through the rest of a script.
Architecture diagram: Question papers and mark schemes are parsed straight from PDFs into a generic rubric model that covers 3+ exam boards. Each answer sheet is graded against the per-criterion rubric by LLM-based evaluation, with follow-through logic and mark dependencies, and the result is marks per criterion.
## Key decisions
- **A generic rubric model.** The engine handles 3+ exam boards with no board-specific rules hardcoded into it.
- **Follow-through as a first-class rule.** A student who carries a wrong earlier answer forward correctly still earns the later marks.
## Result
Marking that took teachers hours per batch now runs automatically, across 3+ exam boards.
## Stack
Python, FastAPI, PDF parsing, LLM-based evaluation
Write-up: [I Asked 3 LLMs for JSON. I Got 3 Kinds of “JSON”.](https://medium.com/@gadhiyajay7676/i-asked-3-llms-for-json-i-got-3-kinds-of-json-794fd969289f)
Page: https://jaygadhiya.com/work/ai-exam-grading-engine/
Contact: hello@jaygadhiya.com
Read the write-up
(a) Gradient of the line through (1, 3) and (3, 7)
3
0/2
(b) Equation of that line
y = 3x
3/3carried through from (a)
Total
3/5
1Problem
Teachers spent hours per batch applying mark schemes by hand, and each exam board formats its schemes differently.
2Approach
- Parsed question papers and mark schemes directly from PDFs.
- Graded each answer sheet against per-criterion rubrics.
- Applied follow-through logic and mark dependencies, so one early error doesn't cascade through the rest of a script.
3Key decisions
A generic rubric model.
The engine handles 3+ exam boards with no board-specific rules hardcoded into it.
Follow-through as a first-class rule.
A student who carries a wrong earlier answer forward correctly still earns the later marks.
4Result
Marking that took teachers hours per batch now runs automatically, across 3+ exam boards.
Want to go deeper on this project?
I can walk you through the decisions and trade-offs behind it.