Recent Updates
Building
Finishing VectorPaths
My first AI vibe coding project
Learning
AI Model Deployment
Local deployment & training workflows
Creating
Photography
Helping others capture moments
Earth Online Game Progress
Main Quests
VectorPaths — Founder & Product Lead
Commonwealth Bank — Customer Banking Specialist
University of Melbourne — B.Com (Finance & Business Analytics)
Side Quests
LLM + Schema Data Extraction: 3,000+ Questions Structured
Adaptive Learning System Design: 2,100+ Test Cases
Learning Vibe Coding & AI Workflow
Case Studies
Deep dive into my projects
VectorPaths
AI-Powered Adaptive Learning Platform
The Challenge
Students don't know which questions to practice. Teachers' judgment relies on experience—it doesn't scale.
1My Hypothesis
What if we could turn "teaching judgment" into a data problem? Then AI could deliver personalized recommendations at scale.
2Conversation Flow Design
3How I Used LLM
• Used LLM + Schema constraints to extract structured data from 3000+ questions
• Why Schema? Pure LLM output format is unstable. Schema ensures every question has complete dimension labels.
• LLM limitations: Direct answer generation has hallucination risk → My solution: LLM extraction + human spot-check verification
4Key Design Decisions
Why 3 questions after wrong answer?
Testing showed: <3 = weak retention, >3 = user fatigue. 3 is the sweet spot.
Why collect error reasons?
Calculation error vs. concept misunderstanding need different interventions.
5Validation
0.87
R² Correlation
2,100+
Test Cases
3,000+
Questions Processed
6Failure Paths I Designed
User answers wrong 3× in a row
→ Lower difficulty, build confidence
User says "too hard"
→ Might mean "I need more hints"
User quits mid-session
→ Save state, resume from checkpoint next time
CBA Customer Service
Real Customer Interactions → Design Insights
The Challenge
How do you design AI assistants that actually help real customers with messy, emotional, and often unclear requests?
1What I Learned from the Front Line
Adaptive Tone
Slow down for elderly customers, be direct with tech-savvy users. One-size-fits-all doesn't work.
De-escalation Patterns
"I want to complain" often means "I need to be heard." Recognition before resolution.
Know When to Handoff
Recognize chatbot limits. Some conversations need human touch.
2Design Principles I Developed
1. Understand the real goal, not the literal request
"I want to transfer money" might mean "I'm not sure how to use the app."
2. Design for the frustrated user first
If it works for someone who's upset, it'll work for everyone.
3. Context is everything
Same words, different history = different response needed.