Thanks for checking out my portfolio!

I’m still polishing up the tablet experience. In the meantime, please check out my portfolio on desktop or mobile.

Thanks for checking out my portfolio!

I’m still polishing up the tablet experience. In the meantime, please check out my portfolio on desktop or mobile.

Maze Companion

Helping students apply computational thinking skills in maze design and solving

Deliverables

Defined 3 student behavioural profiles
Designed a 4-level adaptive AI role in learning
Built a knowledge and reasoning framework

Year

2025

Software

ChatGPT

Figma

Team

Prompt Engineer

Pedagogist

Expertise

Human-AI Interaction

Conversation Design

Prompt Engineering

Knowledge Preparation

Instructional Design

Context

Context

Reimagining how 3,000 students learn with AI

In a 3,000-student undergraduate course, the goal was to shift from using generic public chatbots to a custom one that understood the 3D Maze platform. The new chatbot would provide context-aware, learning-focused guidance that would help students meet the learning outcomes.

Reimagining how 3,000 students learn with AI

In a 3,000-student undergraduate course, the goal was to shift from using generic public chatbots to a custom one that understood the 3D Maze platform. The new chatbot would provide context-aware, learning-focused guidance that would help students meet the learning outcomes.

Reimagining how 3,000 students learn with AI

In a 3,000-student undergraduate course, the goal was to shift from using generic public chatbots to a custom one that understood the 3D Maze platform. The new chatbot would provide context-aware, learning-focused guidance that would help students meet the learning outcomes.

Reimagining how 3,000 students learn with AI

In a 3,000-student undergraduate course, the goal was to shift from using generic public chatbots to a custom one that understood the 3D Maze platform. The new chatbot would provide context-aware, learning-focused guidance that would help students meet the learning outcomes.

Reimagining how 3,000 students learn with AI

In a 3,000-student undergraduate course, the goal was to shift from using generic public chatbots to a custom one that understood the 3D Maze platform. The new chatbot would provide context-aware, learning-focused guidance that would help students meet the learning outcomes.

Problem

Problem

Inaccurate AI responses

Public LLMs generated unsolvable mazes or ignored platform rules and mechanics, wasting students’ time correcting errors. This unreliability led many to lose patience and ask directly for the final answer instead.

Inaccurate AI responses

Public LLMs generated unsolvable mazes or ignored platform rules and mechanics, wasting students’ time correcting errors. This unreliability led many to lose patience and ask directly for the final answer instead.

Inaccurate AI responses

Public LLMs generated unsolvable mazes or ignored platform rules and mechanics, wasting students’ time correcting errors. This unreliability led many to lose patience and ask directly for the final answer instead.

Inaccurate AI responses

Public LLMs generated unsolvable mazes or ignored platform rules and mechanics, wasting students’ time correcting errors. This unreliability led many to lose patience and ask directly for the final answer instead.

Inaccurate AI responses

Public LLMs generated unsolvable mazes or ignored platform rules and mechanics, wasting students’ time correcting errors. This unreliability led many to lose patience and ask directly for the final answer instead.

High effort and low motivation in prompting

Students had to write lengthy, detailed prompts to make AI understand their tasks and context, which was demotivating for beginners during the two-hour course time.

High effort and low motivation in prompting

Students had to write lengthy, detailed prompts to make AI understand their tasks and context, which was demotivating for beginners during the two-hour course time.

High effort and low motivation in prompting

Students had to write lengthy, detailed prompts to make AI understand their tasks and context, which was demotivating for beginners during the two-hour course time.

Student profiles

Student profiles

Understanding diverse learning motivations

Because the course was compulsory and unrelated to many students’ majors, motivation ranged widely, from those rushing to finish to others exploring out of curiosity or genuine interest.

Understanding diverse learning motivations

Because the course was compulsory and unrelated to many students’ majors, motivation ranged widely, from those rushing to finish to others exploring out of curiosity or genuine interest.

Understanding diverse learning motivations

Because the course was compulsory and unrelated to many students’ majors, motivation ranged widely, from those rushing to finish to others exploring out of curiosity or genuine interest.

Chatbot requirements

Chatbot requirements

Translating learning behaviours into chatbot design

The chatbot adapts its interaction types to each learner profile, adjusting depth and pacing to match engagement and learning needs.

Translating learning behaviours into chatbot design

The chatbot adapts its interaction types to each learner profile, adjusting depth and pacing to match engagement and learning needs.

Translating learning behaviours into chatbot design

The chatbot adapts its interaction types to each learner profile, adjusting depth and pacing to match engagement and learning needs.

Translating learning behaviours into chatbot design

The chatbot adapts its interaction types to each learner profile, adjusting depth and pacing to match engagement and learning needs.

Translating learning behaviours into chatbot design

The chatbot adapts its interaction types to each learner profile, adjusting depth and pacing to match engagement and learning needs.

AI role

AI role

Defining an adaptive interaction manner

The chatbot adapts its interaction types to each learner profile, adjusting depth and pacing to match engagement and learning needs.

Defining an adaptive interaction manner

The chatbot adapts its interaction types to each learner profile, adjusting depth and pacing to match engagement and learning needs.

Defining an adaptive interaction manner

The chatbot adapts its interaction types to each learner profile, adjusting depth and pacing to match engagement and learning needs.

Defining an adaptive interaction manner

The chatbot adapts its interaction types to each learner profile, adjusting depth and pacing to match engagement and learning needs.

Defining an adaptive interaction manner

The chatbot adapts its interaction types to each learner profile, adjusting depth and pacing to match engagement and learning needs.

AI knowledge

AI knowledge

Building the platform knowledge

Since no online data described the 3D Maze platform, a custom knowledge base was built to teach the AI how assets, mechanics, and coding logic worked accurately.

Building the platform knowledge

Since no online data described the 3D Maze platform, a custom knowledge base was built to teach the AI how assets, mechanics, and coding logic worked accurately.

Building the platform knowledge

Since no online data described the 3D Maze platform, a custom knowledge base was built to teach the AI how assets, mechanics, and coding logic worked accurately.

Building the platform knowledge

Since no online data described the 3D Maze platform, a custom knowledge base was built to teach the AI how assets, mechanics, and coding logic worked accurately.

Building the platform knowledge

Since no online data described the 3D Maze platform, a custom knowledge base was built to teach the AI how assets, mechanics, and coding logic worked accurately.

Framework for maze difficulty

A difficulty framework defined how maze complexity scaled, from simple linear paths to multi-branch logic structures. This ensured the AI can confidently guide the students in applying the concept appropriately.

Framework for maze difficulty

A difficulty framework defined how maze complexity scaled, from simple linear paths to multi-branch logic structures. This ensured the AI can confidently guide the students in applying the concept appropriately.

Framework for maze difficulty

A difficulty framework defined how maze complexity scaled, from simple linear paths to multi-branch logic structures. This ensured the AI can confidently guide the students in applying the concept appropriately.

Framework for maze difficulty

A difficulty framework defined how maze complexity scaled, from simple linear paths to multi-branch logic structures. This ensured the AI can confidently guide the students in applying the concept appropriately.

Framework for maze difficulty

A difficulty framework defined how maze complexity scaled, from simple linear paths to multi-branch logic structures. This ensured the AI can confidently guide the students in applying the concept appropriately.

Interaction flow

Interaction flow

Example of adaptive student–AI exchange

This example illustrates how the chatbot tailored its guidance to each student profile. The student’s first reply is used to identify which motivation profile they fit into. From there, the chatbot adjusts its tone, depth, and guidance style.

Example of adaptive student–AI exchange

This example illustrates how the chatbot tailored its guidance to each student profile. The student’s first reply is used to identify which motivation profile they fit into. From there, the chatbot adjusts its tone, depth, and guidance style.

Example of adaptive student–AI exchange

This example illustrates how the chatbot tailored its guidance to each student profile. The student’s first reply is used to identify which motivation profile they fit into. From there, the chatbot adjusts its tone, depth, and guidance style.

Example of adaptive student–AI exchange

This example illustrates how the chatbot tailored its guidance to each student profile. The student’s first reply is used to identify which motivation profile they fit into. From there, the chatbot adjusts its tone, depth, and guidance style.

Example of adaptive student–AI exchange

This example illustrates how the chatbot tailored its guidance to each student profile. The student’s first reply is used to identify which motivation profile they fit into. From there, the chatbot adjusts its tone, depth, and guidance style.

Reflection

Reflection

Realities of building a custom chatbot for education

Designing the chatbot revealed that crafting AI for education requires more than a system prompt. It required building a knowledge base, reasoning framework, and interaction logic that worked together to support real educational use.

Realities of building a custom chatbot for education

Designing the chatbot revealed that crafting AI for education requires more than a system prompt. It required building a knowledge base, reasoning framework, and interaction logic that worked together to support real educational use.

Realities of building a custom chatbot for education

Designing the chatbot revealed that crafting AI for education requires more than a system prompt. It required building a knowledge base, reasoning framework, and interaction logic that worked together to support real educational use.

Realities of building a custom chatbot for education

Designing the chatbot revealed that crafting AI for education requires more than a system prompt. It required building a knowledge base, reasoning framework, and interaction logic that worked together to support real educational use.

Realities of building a custom chatbot for education

Designing the chatbot revealed that crafting AI for education requires more than a system prompt. It required building a knowledge base, reasoning framework, and interaction logic that worked together to support real educational use.