Skip to main content
AI for Education

Exam Paper Evaluation

AI-assisted evaluation of handwritten answer sheets against a versioned rubric, with evaluator override and annotated PDFs for students.

Exam controllersEvaluatorsRubric authors
The problem

The bottleneck in Indian university examinations isn't writing the paper, it's evaluating tens of thousands of handwritten answer sheets against detailed rubrics, in time for results to publish. Evaluator burnout, batch reassignments, and grievance backlogs are the norm.

The QverLabs approach

Exam Evaluation is built around four actors, the faculty rubric author, the evaluator, the Controller of Examinations, and the student, each with a purpose-built view that shares the same versioned source of truth. Scoring is calibrated against your own human-scored corpus to QWK (quadratic weighted kappa) alignment with your graders, and the dashboard is role-aware: the Vice Chancellor, the CoE office, and each evaluator see the same session through their own lens. The AI doesn't replace evaluators; it handles the volume so they can focus on judgment.

How it works

1

Author the rubric

Upload the question paper PDF. The system extracts questions into a structured rubric, sections, marks, Bloom's level, that faculty review and freeze.

2

Process answer sheets

Scanned handwritten sheets feed into the pipeline. The AI reads handwriting, aligns answers to the rubric, and scores against criteria.

3

Evaluator review

Per-question scores with criteria-level reasoning, evaluator feedback, and override controls. Annotated PDF emitted on freeze.

4

Student grievance

Students review their annotated sheet, acknowledge marks, or raise a question-level grievance within the acknowledgement window.

Four views, one workflow

Faculty author rubrics, evaluators score, the COE tracks the pipeline, and students review their results, each view tuned to the role.

Faculty

Rubric Designer

AI-assisted rubric authoring from the question paper PDF. Sections, questions, marks, and Bloom's level extracted automatically.

UPLOAD
EXTRACT
REVIEW
FREEZE
Q01Explain properties of object-oriented languages and give example.UNDERSTAND4 m
Q02Illustrate use of classes and objects in Java with example.UNDERSTAND4 m
Q08.aAnalyse and explain Java exceptions. How are they handled?APPLY5 m
Evaluator

Evaluations

Per-question score table with criteria-level reasoning, original sheet thumbnail, and one-click override.

RESULT · E2510247
23.5 of 75 marks · 31%
Q012 / 4INSUFFICIENT
Q022 / 4SATISFACTORY
Q052 / 4SATISFACTORY
Q12.b8 / 8EXCELLENT
Controller of Examinations

Office of the CoE

Role-aware executive dashboard, switch views as Vice Chancellor, Controller of Exams, or evaluator. Session KPIs, marks distribution, programs at risk, and per-program completion, streaming live.

Student

My Results

Per-subject card with marks, grade, and acknowledgement state. Open the annotated PDF, raise a grievance per question, or accept marks.

RM501
Research Methodology
55.5 / 75
REVIEW PENDING
CSE301
Operating Systems
62 / 75
ACKNOWLEDGED
CSE304
Computer Networks
51 / 75
GRIEVANCE RAISED

What the student receives

A page-by-page annotated PDF, plus a one-page evaluation summary. Students review every annotation before they acknowledge marks or raise a grievance on a specific question.

Q4 · 0.5 / 4 · INSUFFICIENT4

Command-Line Arguments: Command-Line Arguments in JAVA can be defined as the element which helps in running the program.

→ It is denoted by system.out.println( string.args[]).

→ It helps in performing operations and programme which are coordinated manually.

Code: public class A {
  public void main();
  {
    system.out.println( string.args[]);
    } system.out.println("Argument o");
  }
}

→ It helps in processing datatypes efficiently, moving them from primitive datatypes to wrapped datatypes.

EVALUATOR · INSUFFICIENT
Your explanation of command-line arguments is unclear, and the provided code example contains several syntax errors, including an incorrect main method signature, preventing it from compiling or demonstrating the concept.
EVALUATION SUMMARY
23.5 / 7531%
run 019dff3d · paper 019dff31
Excellent × 0Good × 0Satisfactory × 4Insufficient × 9
QAwardedLabelFeedback
12 / 4SATISFACTORY
22 / 4SATISFACTORY
32.5 / 4SATISFACTORY
40.5 / 4INSUFFICIENT
7.b1 / 5INSUFFICIENT
9.b0 / 8INSUFFICIENT
12.b3.5 / 8INSUFFICIENT
+ 6 more questions

Reads the way students actually write

Cursive, block print, code, diagrams, comparison tables, math notation. The pipeline handles what an Indian university exam paper actually looks like, not just clean dataset PDFs.

CURSIVE
Inheritance: It can be defined as the child class or progeny class inherits levels from the parent class.
Joined script, personal letter forms, slope variations
BLOCK PRINT
SECTION-A

Object-oriented Languages: OOL can be defined as the language where it follows properties like Inheritance, Polymorphism, Abstraction.
Capital letters, deliberate spacing, ruled-line discipline
CODE
class Animal {
  void eat();
  {
    system.out.println("Animal eats food");
  }
}
class Dog extends Animal
  void barks();
Brackets, semicolons, indentation, mixed-case method names
DIAGRAMS
BYTECODE
CLASS LOADER
RUN-TIME
EXECUTION ENGINE
Hand-drawn boxes, arrows, flow charts, hierarchies
COMPARISON TABLES
thissuper
Refers to current classRefers to parent class
Slower in functionFaster in function
Multi-column tables drawn freehand with ruler lines
MIXED CONTENT
Example: a + b where a = 10, b = 20
Result = 30
→ Operators: ×, ÷, ≤, ≥, ==
Math, arrows, symbols, mixed prose and formula

What makes it work in production

AI-assisted rubric authoring

Question paper extraction with automatic Bloom's level inference. Faculty edit in place, run AI edge-case review, then freeze a versioned rubric.

Handwriting-aware extraction

Reads handwritten answer sheets, aligns to rubric questions, and emits structured text alongside the original scan for evaluator review.

Criteria-level audit

Every score breaks down into the rubric criteria that earned or lost marks. Students see why they got Insufficient on Q4 and Excellent on Q12.

Annotated student PDF

Each scored sheet emits a 30+ page annotated PDF with sticky-note evaluator feedback per question. Replaces the back-and-forth of grievance review.

Role-aware dashboard

One pipeline, three lenses. The Vice Chancellor sees programs at risk, the CoE office tracks completion and flagged sheets, and each evaluator sees their own queue, switched without leaving the dashboard.

Versioned rubrics

Frozen rubrics stay immutable for sheets that referenced them. Editing creates a new version, older evaluations remain reproducible.

Frequently Asked Questions

No. Every score has a human in the loop. The AI reads handwriting, aligns answers to the rubric, and proposes scores with criteria-level reasoning. The evaluator reviews, can override any score, and only frozen sheets emit the annotated student record. AI carries the volume; evaluators keep the judgment.

Faculty upload a question paper PDF. The system extracts questions into a structured rubric with sections, marks per question, and inferred Bloom's taxonomy levels (Remember, Understand, Apply, Analyze, Create, Evaluate). Faculty edit in place, run an AI edge-case review, then freeze a versioned rubric used for all subsequent evaluations.

Each student gets an annotated PDF of their scored answer sheet, page-by-page evaluator feedback per question, plus a summary banded by Excellent / Good / Satisfactory / Insufficient. Students can acknowledge to accept marks, or raise a grievance on a specific question within the acknowledgement window. Earlier sessions are sealed and view-only.

Yes, the pipeline is built for handwritten sheets. It reads handwriting, identifies question boundaries, extracts answer text into structured form, and aligns each answer to its rubric question. The original scanned PDF stays available alongside the extracted text so evaluators can verify any anomaly inline.

Through a dedicated calibration step. You upload a human-scored corpus — a ZIP of scanned answer-sheet PDFs plus a scores.csv listing roll number, file name, question number, max marks, and the human-awarded score. The system builds calibration datasets from it and aligns AI scoring to your graders, measured by QWK (quadratic weighted kappa), before live sheets are scored.

A live, role-aware dashboard shows session-level metrics: sheets evaluated, average and median marks, flagged sheets awaiting human review, programs at risk, and historical trail across sessions. A view-as switcher renders the same session for the Vice Chancellor, the Controller of Exams, or an individual evaluator, and per-program drill-downs surface subjects, evaluator assignments, and live activity, so the COE can spot stuck pipelines before freeze.

When students answer more sub-questions than required (e.g., attempting Q6.b after Q6.a counts), the system flags excess attempts and only counts first-required-N toward the total. Grievances are raised at the question level, route to the evaluator with full context, and resolutions are versioned in the audit log.

From handwritten sheet to acknowledged result.

See the rubric designer, evaluation flow, QWK calibration, role-aware CoE dashboard, and student review walkthrough with our team.

Schedule a call