A universe built from data. Every planet is a system I shipped.
An AI-ethics tool that finds and flags bias in AI text generation.
Real-time monitoring and alerting catches harmful content as it's generated, while a harm-classification system sorts issues by type so they can be handled for compliance and safety. A feedback loop channels every flagged case back into the model, improving robustness with use — a guardrail sitting between an LLM and its users.
An end-to-end system classifying colon-cancer histology images.
Before modelling: duplicate detection, class-imbalance analysis and dimensionality reduction to understand the data and shut down leakage risks. Then multiple CNN architectures, heavy augmentation, and patient-level stratified splits — so accuracy holds on genuinely unseen patients, not just unseen images. Benchmarked independently against published results, with a hard look at limitations and what deployment would actually demand.
An online music-subscription application built across five AWS services.
User authentication, registration and validation run against DynamoDB with backend logic guarding data integrity. Search and subscription features persist to storage built for reliability, with Lambda and API Gateway handling the request layer and S3 holding assets. Deployed and operated on an Apache web server on an Ubuntu EC2 instance — architecture, integration and system operations end to end.
Melbourne property prices, predicted with expert priors and honest uncertainty.
A Bayesian linear regression in R and JAGS, fusing observed market data with expert-informed priors. MCMC sampling was designed and tuned across multiple chains, with convergence diagnostics and posterior estimates confirming the model could be trusted. Every prediction arrives as a distribution with 95% credible intervals — a defensible valuation that states its own uncertainty, rather than one black-box number.
A filing cabinet walks in. One ordered PDF walks out.
Companies still pay people to digitise decades of paper by hand. This reads an unsorted stack, finds the page numbers wherever they were scribbled, understands dates however they were written — and rebuilds the document in order. How it does that is the interesting part.
The places the search engines never suggest.
Not another ranked list of the same landmarks. It thinks in moods, not keywords — a night drive, a coast road, a mountain pass, a table worth dressing up for — and learns which ones are yours. The recommender behind it is where the real work lives.
Swipe right on the outfit. Then wear it before you own it.
Catalogues become a deck of cards — left, right, learning your taste with every flick. Like something and it lands on you, fitted, ready to buy, keep, or show off. Making that fit look real is the hard problem, and it's the one being solved.
A rock full of terminals — three roles, three systems. Step inside to read them.
The bright core everything here orbits — where all of it began.
Wherever your data is raw, unruly, or hard to trust — that's where I start.