作品。

I rebuilt this in 2026 to turn a static information page into calculators that each cite the provision they apply. The calculators cover meeting quorums (Schedule 1 clause 17 and Schedule 2 clause 12), levy shares by unit entitlement (section 83(2)), and the thresholds for ordinary, special and unanimous resolutions (section 5 and Schedule 1 clause 14), checked against the Strata Schemes Management Act 2015 on 11 October 2026. This is an informational tool only, not legal advice.

A read-only terminal tool that explains what your Homebrew setup actually does on macOS. It shows which packages overlap or shadow each other, how often you and your AI agents use each one, and what a move to a new Mac needs to carry across. It is built for AI coding agents first, with JSON on stdout, stable schemas, meaningful exit codes and no prompts, while people at a terminal still get a readable table. The web demo demonstrates brewscope's capabilities in the browser with synthetic output.

I revived a 2024 timezone calendar prototype. The original embedded a live calendar, so this version uses synthetic events instead, which makes it safe to share. Events are generated in the browser from a fixed seed (2026), so the demo is reproducible and shows no real schedule. The grid runs Monday to Sunday by the hour, colour-coded into work, meetings, focus time and personal events.

Personal study notes and code following the DataTalksClub Data Engineering Zoomcamp 2024, a free hands-on course that builds an end-to-end data pipeline on NYC taxi data. I completed Week 1 (Docker, Postgres, ingestion) in 2024. Weeks 2 to 6 (workflow orchestration, data warehouse, analytics engineering with dbt, batch processing with Spark, and stream processing with Redpanda) were completed in October 2026 with Claude as a pair programmer, using free local stand-ins for the Google Cloud services. Each week answers the official 2024 homework with numbers computed by code in the repo, with tests in CI. The Week 7 project is not started yet.

This is the generic version of a structure I use every day, with every personal detail taken out. One behaviour contract (AGENTS.md), Areas as the backbone (ongoing responsibilities), Projects (time-bound, each tagged with its area), HQ (goals, reviews, intake), and memory files that stay small (MEMORY.md with a 300-line ceiling, append-only LOG.md). The web demo showcases the template with an interactive vault navigator, landing page, and guided tour.

A local-first tool that analyses WeChat chat exports and distils your texting voice into a persona prompt for Claude or other LLMs. The Python pipeline extracts 12+ linguistic features — message length, emoji frequency, code-switching between English and Chinese, catchphrases, punctuation habits — and maps them into LLM instructions. The web demo runs the methodology on six synthetic personas, with feature validation, bootstrap confidence intervals and a privacy threat model.

Ninety-five South Australian and Australian government datasets curated for research and civic tools, each documented with source, licence, retrieval date and field definitions. The collection spans ten domains — health, justice, education, infrastructure, environment, social services, planning, economy, demographics and governance — with provenance tracking for every joined or transformed dataset.

A Claude Code skill that triages Gmail by auto-labelling senders into provider categories — Banking, Shopping, Subscriptions, Government and more. The Python skill classifies by sender domain and content, generates importable Gmail filter XML, and works in dry-run mode so nothing moves until you review it. The web demo showcases the provider taxonomy, classification rules and dry-run simulation on synthetic inbox data.

Tutoring and mentoring university students across computer science subjects through ClassBro, from programming fundamentals to advanced topics in machine learning, AI, databases, systems and cloud computing. Working one-on-one with learners to bridge concepts and code.

Brings four ranking systems (QS, Times Higher Education, and two U.S. News rankings) together for 3,790 universities across 40 years. A no-hard-coding pipeline pulls each ranking live and recovers QS history from archived edition IDs, published as an open, FAIR-licensed dataset.

Built as sole developer for my mum's meal-prep studio: one system for members, prepaid meal cards, daily orders, kitchen and delivery, and finance. A single Expo codebase builds for iOS, Android and the web, on a Hono API on Vercel with Turso (libSQL) and Drizzle. It was built to replace group-chat sign-ups that were tallied by hand and re-keyed into an Excel workbook.

Research officer role in the Intelligence & Probity Unit of SAPOL's Ethical and Professional Standards Branch. Conducts strategic research on ethical frameworks and professional standards, performing analysis to support branch decision-making and strategic planning.

As ASO7 Senior Data Analyst in the Intelligence & Probity Unit of SAPOL's Ethical and Professional Standards Branch, produces the quarterly Use of Force and Vehicle Pursuit statistical reports, led an end-to-end review of the complaint administration workflow, and analysed a financial year of expiation notices. Also built a Python client and web console for the complaint-management system APIs, covering more than 1,100 endpoints.

A maths-complete explainer of the expectation-maximisation algorithm that I wrote in September 2025, framed around sparse movie ratings, with a notebook that fits a two-component Gaussian mixture from scratch. In 2026 I rebuilt it as a web lab where you can step through the worked example by hand, run EM on the notebook's 200 ratings, and trigger label switching, stopping too early, local maxima and variance collapse. The TypeScript port reproduces the notebook's printed results, and the slips in my own hand-worked numbers are corrected beside the originals.

South Australia publishes its gaming-machine statistics as separate PDF releases. In September 2025 I transcribed about 110 of them, from FY 2009/10 to FY 2024/25, into one workbook with a Power BI report. In 2026 I rebuilt it as a website on statewide revenue and tax, council areas, licences and manufacturers, with every figure checked against its PDF and the gaps in the series shown. It is a personal project on public releases from Consumer and Business Services, not affiliated with CBS, the Government of South Australia or my employer, and it takes no position for or against gambling. If gambling is affecting you or someone close to you, call the Gambling Help Line on 1800 858 858.

Co-founded a Melbourne startup building a map-first app for discovering people and events nearby. As Dev Lead, owns the technical architecture (an Expo React Native client on a Django, Rust and PostgreSQL backend, shipped through GitHub Actions) and leads code review for a part-time engineering team.

A personal command-line tool I wrote in August 2025 that makes mock South Australian addresses for software testing, each with a suburb, postcode, council and remoteness level, and looks up real addresses through the Mapbox API. Porting it to the web in 2026 showed that its remoteness and socio-economic weights were never applied and its socio-economic column was empty. The revival keeps that uniform mode as built, adds the promised weighting from ABS remoteness and SEIFA data, and checks each sample against its target mix. Every output is labelled as synthetic test data.

Two notebooks I wrote in August 2025 that collect public pages from The American Presidency Project at UC Santa Barbara, with a thread pool, retries and resumable runs: 7,556 campaign documents from 2016 to 2024 and 179 debate transcripts from 1960 to 2024. In 2026 I revived the collection as Campaign Text Lab, a descriptive reading room that reads the original CSV files without scraping again, with distinctive-word comparisons, debate talk shares and term timelines. Every speaker goes through the same code, and nothing is scored or predicted.

Built the first intelligence analytics capability within the CBS Prevention Team — integrating ABS, SA Health, ACCC, and DataSA data into unified dashboards and GIS maps used by the Minister's Office.

A retrieval-augmented generation prototype I built in October 2024 to turn course notes into revision material, with DSPy, sentence-transformer embeddings and SQLite. In 2026 I rebuilt it to run in the browser. It finds the passages that answer a question and drafts revision agendas, worked examples and practice questions that cite those passages, and nothing is exported until a person has reviewed each item. Generation runs on the visitor's own API key, and the demo uses an openly licensed OpenStax statistics textbook instead of my private notes.

A clinician-facing and patient-facing mental health app for the University of Melbourne's Department of Psychiatry, built as sole developer. Rebuilt it from Uniapp to Expo React Native with a clinician dashboard on a Flask backend, kept hosting under $500 a month by running each service in its own Docker container on AWS LightSail, and kept it GDPR-aligned before handing it to a professional team for production.

Worked on Genomics Metadata Multiplexing (GMM), an R Shiny tool that builds CEL-Seq2 sample sheets for plate-based single-cell sequencing. It merges flow-cytometry sort metadata from FACS index-sort files with the plate layout and primer indexes, so nobody has to stitch them together by hand. Wrote the Shiny side of the app, added test inputs with expected outputs, and contributed a column clean-up step to WEHIGenomicsRnD/celseq-sample-sheet-generator.

A June 2024 prototype of Echo, a companion chatbot for Psyckitchen built to remember a person's profile and past conversations. A Flask backend on Azure App Service kept both in MongoDB and sent them to GPT-4o with every new message. A Next.js test page sat alongside it, with a chat box and a mock wearable display. The prototype has since ended.

A two-stage fact-checker for climate-science claims, built for COMP90042 with Wei Zhao and Xuan Wang. TF-IDF retrieval pulls evidence from about 1.2 million Wikipedia passages, then a Transformer trained from scratch labels each claim as supported, refuted, disputed or not enough information. I worked on the system design, the preprocessing and the TF-IDF retrieval, and Wei built the classifiers. I revived it in 2026, with the original pipeline re-run as it was and the weak results shown.

Built and refined interactive digital educational content aimed at bridging the gap between education and professional success. Conducted market research on leveraging advanced technologies to improve course engagement and learner outcomes for students across Australia and beyond.

A survey of self-driving databases for COMP90050, written by Group 40 with Xiaoyi Liu, Runqiu Fei and Qingxuan Yang. Runqiu and I covered index selection, from heuristics to bandits, and Xiaoyi and Qingxuan covered workload-driven optimisation. I revived it in 2026 as Self-Driving DB Lab, where the index advisors the survey described race on a live SQLite database in the browser.

Harvested and analysed Twitter and Mastodon data alongside ABS SUDO spatial data to produce a Social Sense Dashboard illuminating Australian sentiment, social trends, and regional behavioural patterns.

Processed a large-scale Twitter dataset on the University of Melbourne's SPARTAN HPC cluster using MPI and Python, identifying tweet distribution across Australian cities and top contributors.

This was our MAST90106 and MAST90107 industry capstone, hosted by CSIRO in 2023. I worked in Group 36 with Ritwik Giri, Jiaqi Hu, Xiangyi (Emma) He and Jihang (Jonathan) Yu. In 2026 I rebuilt it from scratch as a clean-room revival on public World Bank, NOAA and FRED data, with no CSIRO or team data, code or documents. It asks whether lagged climate indices such as El Niño help forecast next month's prices for wheat, maize, rice, soybeans, palm oil and sugar, comparing forecasts with and without the climate input on months the models never saw. None of the eight commodities forecast better with it.

Forty projects I worked through for practice in late 2022, one notebook each, following Aman Kharwal's 180 Data Science and Machine Learning Projects with Python series on thecleverprogrammer.com. The project ideas and methods come from his tutorials. In 2026 I turned them into a study log that records which notebooks ran, which broke and what each printed, plus eight audits where a little more rigour changes the conclusion, such as stock forecasts that lose to a flat forecast once the split respects time.

A pair project with Wei Zhao as team _4399 for COMP30024. Cachex is a two-player connection game like Hex, with captures and a steal rule. We wrote an A* solver for the search task and a minimax agent with alpha-beta pruning to play the full game. I revived it in 2026 as Cachex Arena, where you can play the agent or run a seeded tournament.

A small community you can only join with an invite code from a member, where each member holds at most two unused codes. Jiahong Zheng and I built it in early 2022 to learn a full stack, with an Express and MongoDB API behind a Next.js front end. The database is long gone, so in 2026 I revived it as one Next.js app on SQLite (Turso), with the 2022 invite rules ported and checked by parity tests, an invite tree, demo logins and a guided tour.

This was our MAST30034 industry capstone with CSL. I worked in Group 07 (Team 4399) with Yuchen (Cynthia) Cai, Rongshun Li, Zhiliang (Tommy) Tang and Quzihan (Martin) Wu, and was the team's Scrum Master. We parsed free-text HPLC lab-result spreadsheets into tidy records in the spirit of the FAIR principles, rebuilt control charts for four assays to find out-of-control runs, and compared PCA, t-SNE and UMAP with K-Means and DBSCAN as an aid for spotting unusual runs. I revived it in 2026 as HPLC QC Lab, which runs the team's methods on synthetic data, with no CSL data, code or documents.

A mobile-first personal CRM built by Team 4399 for the COMP30022 IT Project, with Bin Liang, Hongji (Harrison) Huang, Wei Zhao and Yixiao Tian. It has a React front end, an Express REST API and a MongoDB database. I was the team's Scrum Master and worked on the contacts, meeting records and map screens. I rebuilt it in 2026 as 4399 CRM, a single Next.js app with a guest sandbox.

My individual project for MAST30034 Applied Data Science. I cleaned every 2019 New York yellow-cab trip in PySpark, more than 84 million records, and fitted an elastic-net regression to estimate trip times. I revived it in 2026 as NYC Taxi 2019, with zone maps, weather, events and collisions, and the 2021 regression running in the browser.

Five major iterations of a personal website — currently rebuilt in v5 with a Nothing OS-inspired minimal design language. A living record of technical and professional development since 2020.

I completed twelve self-paced job simulations on Forage, each designed by the company that set the brief. The companies were KPMG, BCG, British Airways, Quantium, Tata, Cognizant, GE Aviation, Accenture, Red Bull, PwC (two) and Standard Bank. The work covered customer analytics, churn and booking models, dashboards and data joins.