Skip to main content
Development Studio Showcase

Portfolio
Projects.

Three flagship projects demonstrating full-stack expertise, AI integration, and production-grade engineering.

National AI Challenge 2025 Winner

GradGenie

AI-Powered Exam Grading for Irish Leaving Certificate

Multi-agent AI system that grades student essays through collaborative AI conversation.

Visit Site

The Problem

60,000+ Irish students sit Leaving Certificate exams annually. Grading is:

Slow

Weeks of manual marking by thousands of human examiners

Expensive

Massive cost to coordinate and pay examination teams

Inconsistent

Variation between individual markers affects fairness

Traditional AI grading uses single models that miss nuance and introduce systematic bias.

The Solution

A 5-stage agentic workflow where OpenAI and Anthropic models debate each answer:

01

Initial Grade

OpenAI GPT

Grade against official LC rubric

02

Bias Review

Anthropic Claude

Check for known grading biases

03

Grade Revision

OpenAI GPT

Reconsider based on bias feedback

04

Teacher Compare

Anthropic Claude

Align with real teacher examples

05

Final Calibration

OpenAI GPT

Final decision + confidence score

Low-confidence (< 0.85) routes to human review

High-confidence auto-grades with full audit trail

Technical Innovations

Dual AI Conversation

Two Models Debating

Not one model scoring - two models having an intelligent conversation about student work, catching biases and calibrating against real teacher examples.

Smart Retrieval

Question Type Grouping

Groups by type (personal_reflection, text_analysis, writing_evaluation) for smarter teacher example retrieval instead of exact question matches.

Routing

Confidence-based

Psychometric thresholds determine auto-grade vs human review.

Context

Complete Awareness

AI receives full source text, rubric, indicative material, and teacher examples.

Audit

Conversation Logging

Full audit trail of AI decision-making for transparency and appeals.

Database Design

Question
rubric
source
max_pts
StudentAnswer
student_work
teacher_grade
rationale
generates
GradingSession
AI conversation log
confidence scores
LHR decision
audit trail
Tech Stack
FrontendReact, TypeScript, MUI, Vite
BackendDjango, Django REST Framework
AI ModelsOpenAI GPT-5, Anthropic Claude 3 Haiku
DatabasePostgreSQL
DeploymentNetlify (frontend), separate backend
Results
5

API calls per answer

~13s

Processing time

0.85+

Confidence score

Every grading decision logged and explainable.

Role

Akella inMotion Delivered

Full backend architecture
Multi-agent AI system
Database schema
Frontend development

Part of a team effort.