Blog
Occasional thoughts on AI, technology, and building things.
How to Analyse an Agent: Sixty Years of Answers to a Two-Year-Old Question
Agent evals are a measurement problem, and measurement is an old discipline — from Cohen’s kappa and Krippendorff’s alpha to McNemar’s test and pass^k. Co-authored with Fable 5.
I Tested Fable 5 on This Website. It Found a Real Vulnerability.
Claude Code, running Anthropic’s Fable 5 model, reviewed this website like a seasoned engineer — then fixed what it found.
25 Years of NSW Property Sales
4.35 million transactions, $3.6 trillion in value — every suburb in NSW from the raw Valuer General data, with a searchable explorer.
Five Things I Learned From 7 Years of APRA Banking Data
Macquarie's quiet takeover, $154B in COVID deposits, and why investment lending never recovered.
Australian Banking — Bar Chart Race
An animated visualization of 23 years of Australian banking, built with APRA data.
Vibecoding My Personal Website
What happens when you let AI drive the keyboard and you drive the vision.
AI in Product Development
Lessons learned from building AI-powered features in financial products.
Building Interactive ML Demos
How I approach creating educational tools for complex concepts.
Understanding Self-Attention
A deep dive into the mechanism that makes transformers so powerful.