Data scientist building production ML systems — models, pipelines, and the APIs that ship them.
✦Hover the lettersTap the letters✦Click the portraitTap the portrait✦Throw the ball
Final-year Data Science student at RMIT (graduating 2027). I build machine-learning models and the pipelines & APIs that carry them into production — three projects run as live, documented services with MLflow tracking and CI/CD.
Just as comfortable in SQL and Power BI / Tableau, turning large, messy datasets into decisions people can act on.
Hover or tap a tool
✦Keep scrolling: the cards stack
Beat ARIMA & a coin-flip on next-day ASX direction (51.8% vs 48.8%) with engineered RSI/MACD/Bollinger features and leakage-free walk-forward CV across 10 equities (18,460 samples). Served via a FastAPI /predict endpoint + Streamlit, MLflow registry, pytest & GitHub Actions CI — running live.
Caught 85.9% of fraud at 84.7% precision (PR-AUC 0.88, ROC-AUC 0.99) on 284,807 transactions (0.17% fraud) with a cost-sensitive XGBoost + Isolation Forest ensemble and train-only SMOTE. SHAP reason codes streamed live over a FastAPI WebSocket.
Ranked 260 Greater-Melbourne suburbs on an 8-dimension liveability score — real suburb-boundary geometry (GeoPandas, Vicgrid) with PCA setting the weights (PC1 = 42% of variance). An interactive Folium choropleth re-scores as you move the sliders.
An /api/cashback service that pulls live offers from five vetted Australian networks via Apify, runs a Claude pass to drop untrusted sources, and ranks what's left by the user's own spending — with three fallback tiers so the feed always loads.
✦Filter, shuffle, then hover a cardFilter, shuffle, then press a card
Showing 13 of 13 projects
Funnel analysis · 110M events, 5.3M users
K-Means customer segments
Demand & surge-pricing analysis
LogReg / RF / XGBoost compare
Regression modelling & feature selection
Kaggle ML classification
Telemetry & strategy analysis
Exploratory data analysis
Geospatial mapping in R
Full-stack web app
AI-powered commerce app
Discord music bot
ML delay prediction from weather
Four games and eight tools. Beat a fraud model, break a regression, train a neural net, and try to out-guess the market. Collect every card in the Data Dex.
✦Pick a tab · collect all 12 Dex cardsSwipe the tabs · collect all 12 Dex cards
Swipe the card, or use the buttons. This deck is rigged: 1 in 3 is fraud.
My real project catches 85.9% of fraud at 84.7% precision. Here you play the analyst: decide, then see the model's risk score and the reasons it gave.
Red dots train the yellow curve. Faint dots are hidden test data. Raise the degree and watch training error fall while test error blows up.
A real network, learning in your browser. Pick a puzzle, add neurons, and watch the boundary bend around the dots. Reach 95% to catch this one.
Predict whether the next day closes up or down. My real ASX model managed 51.8% against a 48.8% baseline. Markets are noisy, so see how you do.
Play each tool to catch its card. Gotta train 'em all.
Group points into K clusters — the same technique behind my RFM customer-segmentation project.
Drop points and watch a least-squares line fit in real time.
How strong is the relationship? Slide to guess r, then reveal the truth.
Inverse-distance-weighted (IDW) interpolation turns scattered stations into a smooth heat field.
A spring + repulsion physics sim lets the graph lay itself out.
Shows a 1-D distribution without hiding points — collision-avoidance packing.
Kernel-density estimation draws each group's distribution shape.
A Voronoi tessellation builds convex cells whose area encodes value, animated with Lloyd relaxation.
Open for data science collaboration, analytics projects, or data-related opportunities.