N Admin
Product designer
Portfolio 2026Systems thinking
Workflow architecture
Product auditsHealthcare · Fintech
Education · Retail
Civic dataDelhi, India
28.6139 N
77.2090 E01

EducAItors: the problem wasn’t grading, it was everything before it

Case 02
Education · AI
Product Designer
Instructor Module
6 months
Discovery, design, validation
Case 02 / Education · AI

The problem wasn’t grading. It was everything before it.

(Asked for)An AI tool to help instructors grade.
(Found)Grading fails before it begins. Intent is captured first, so every score traces to approved criteria.
(Changed)Assignment setup became the foundation that grading depends on.
Role
Product Designer, Instructor Module (Pre-Evaluation phase)
Methods
System design, workflow architecture, intent modelling, AI integration design
Scope
The Pre-Evaluation phase: assignment creation, rubrics, calibration and preview, before any grading happens
The evidenceEstimated setup effortPre-Evaluation phase02

The evidence

to set up an assignment, down from 3–4 hours
45⁠–⁠60min
phases consolidated into three stages: pre-evaluation, evaluation, post-evaluation
7 → 3
principles every score answers to: trust, accountability, auditability
3
approaches tested before the breakthrough: templates, AI-first, manual with hints
3
ContextAI-assisted evaluationInstructors, students, institutions03

Grading was already making decisions it couldn’t explain.

EducAItors is an AI-assisted evaluation platform, meant to make grading consistent, explainable and auditable across instructors, courses and institutions. The institutions are mainly Tier 2 and Tier 3 colleges in India.

An instructor might give one answer 7/10 and the next 8/10, from standards that were never written down. Existing AI grading tools returned a score like 73/100 with little explanation of how it was reached. Students received generic feedback, and institutions had little evidence to defend how a grade was reached.

I traced the problem upstream to assignment creation. Expectations were often unclear, rubrics inconsistent, and standards lived largely in the instructor’s head.

  1. (01)AssignmentExpectations often unclear.
  2. (02)RubricInconsistent from one to the next.
  3. (03)InstructorWhere the standards actually lived.
  4. (04)Score7/10, then 8/10, with no way to explain the difference.
ApproachPre-Evaluation phaseInstructors and the AI04

Designing for instructor intent.

Pre-Evaluation is the first of three product stages: Pre-Evaluation, Evaluation and Post-Evaluation. I designed it in two layers. Quick decisions come first: the course, and how to create the assignment (reuse one, start from scratch or use AI assistance). Structured setup follows: details, rubric, calibration, then preview and publish.

  1. 01Course
  2. 02Source
  3. 03Details
  4. 04Rubric
  5. 05Calibration
  6. 06Preview
The turnAutomate structure, not judgement05

Don’t automate decisions. Automate structure.

The system maps questions to learning outcomes and organises the evaluation logic. AI only suggests, in three moments: question phrasing, rubric gaps and calibration. Every suggestion stays visible and editable, and nothing happens until the instructor confirms it.

What was actually brokenThree core problemsUpstream of grading06

What was actually broken

Diagnosis

The foundation was missing.

(001) Unclear expectationsAssignment expectations were often unclear before grading began.
(002) Inconsistent rubricsInstitutions used different rubric formats, scoring systems and performance levels.
(003) Standards in the headEvaluation standards lived largely in the instructor’s head, not in the system.
ImpactThree design decisionsStructure, not judgement07

Pre-Evaluation became the foundation for grading.

  1. 01

    Intent captured first

    Rubrics, learning-outcome mappings, calibration and a final preview set the standard before any grading.

  2. 02

    One rubric language

    A normalization layer turned different rubric formats into one internal structure, while instructors kept the formats they knew.

  3. 03

    Calibration as intent

    Instructors defined what strong and weak answers contain. Reused assignments could inherit that intent; bigger changes triggered recalibration.

What I learnedEvaluation starts with clarity08

What I learned

The important problems come before the interface.

Cleaner forms would have been superficial. The system first needed a way to understand what instructors wanted to evaluate.

AI should support judgement, not replace it.

In high-stakes domains, automation without control erodes trust faster than it creates value.

Structure enables fairness at scale.

Normalised rubrics, explicit intent and evidence mapped to criteria make fair evaluation possible across many instructors and students.

When evaluation begins without clarity, the system has already failed.

Next caseCase 03 / Fintech4 months09

Paisalello

Read the next case