Selected work · 2024–2026

THEWORKSHOP

Case studies · the timeline

Systems people actually use

Production work across climate, energy, e-commerce, and supply chain. Real systems with real users, not proof-of-concept demos. The whole route, in order, including the chapters before the case studies.

2025 – present

ArborMeta

Machine Learning Engineer

Read the deep-dive →
Problem

Carbon and ecology data was scattered across field surveys and raw LiDAR. None of it sat in one place a policymaker could actually look at, and none of it answered the question that matters: how much carbon, and is it growing?

Approach
  • Canopy-height and above-ground-carbon models from paired LiDAR captures, with growth quantified between flights parcel by parcel
  • Live fire-spread and ember mapping driven by wind fields
  • A multi-LOD PMTiles pipeline that keeps continental-scale datasets fast at every zoom level
  • A PostGIS spatial backend, with parsers that pull Fulcrum field surveys straight into the platform
  • A map viewer that renders live survey data over satellite imagery
Outcome
  • Used to advise federal and state government on carbon credit rules
  • Canopy growth and carbon change measured between repeat LiDAR captures, the numbers under the policy
  • Field data from four countries, readable in one live interface
FastAPIPostGISMapLibre GL JSPMTilesALS / LiDARDockerReact
Inside the platform
Canopy growth and above-ground carbon between two LiDAR captures
Live fire-spread and ember mapping
Processing pipeline with live data provenance
Continental-scale survey coverage
Australia-wide fire and wind density

2025 – present

Step One Clothing

Software Engineer (Contract)

Problem

As the store grew, shoppers struggled to find the right product quickly.

Approach
  • Natural-language product search
  • A new mega-menu the merchandising team can edit themselves
  • Accessible navigation with analytics built in
Outcome
  • Easier product discovery at scale
  • Content changes ship without engineering
  • Kept on as the lead front-end contractor
Shopify LiquidJavaScriptNLP SearchAnalytics
Inside the platform
Natural-language search — describe a need, get a matched product
Schema-driven mega-menu the merchandising team edits themselves
Accessible primary navigation across the storefront

2025

FibreTrace

iOS Engineer

Problem

Brands needed shoppers to scan a product, see its origin, and trust the result.

Approach
  • Native iOS app in SwiftUI
  • A QR / NFC scan opens a signed, tamper-proof passport
  • Works offline
  • On-device ML flags fake tags and broken chains of custody
Outcome
  • Scan to verified passport in under three seconds
  • Works with no connectivity — verification holds up in a warehouse or a paddock
  • Launched a pilot with a partner brand
SwiftSwiftUICore MLNFCOffline-first
Inside the platform
Trace a material at every stage — fibre to consumer, with provenance
Embed, verify, share — how fibre integrity is proven end to end
A searchable network of FibreTrace-verified suppliers

2023, 2024

Australian Energy Market Operator (AEMO)

Software Engineering Intern

Melbourne

  • Python analytics over terabyte-scale NEM/MMS datasets with real-time dashboards used by operations.
  • High-throughput C parsers for market/telemetry files; C# micro-services integrating MMS schemas and SQL stores.

2023 – 2025

Monash University

B.S. Advanced Computer Science (AI/ML)

Melbourne

  • Concentrations in Artificial Intelligence / Machine Learning and Advanced Algorithms & Data Structures. High Distinction GPA.

2022

James Cook University

Medicine (MBBS) — deferred after 6 months

Townsville

  • Left a medical degree to build geospatial tools full-time — the decision that set this whole path in motion.
Research & writing

Papers

Galah: compute-optimal scaling at byte level

Independent · Scaling-law study · Ongoing · 2026

A Chinchilla-style sweep on one GPU: 39 byte-level runs, 0.1M to 113M params, six budgets. Frontier b = 1.04, with four runs that diverged on the longest schedules. Ongoing.

Adapting frontier LLM-pretraining techniques under fixed compute

Maincode · Mainrun assessment · 2026

40M parameters, seven epochs, one seed. Validation loss 1.75 → 1.18. Muon, RoPE, WSD, ReLU², value residuals.

Web-page navigation as a stochastic process

Monash FIT3139 · Computational Modelling & Simulation

PageRank past linear algebra: Gillespie simulation, exact Markov chain, simulated annealing to recover α.

Personal projects

Built for myself

catchment

physics vs neural

2026

A browser-native Earth engine: shallow-water hydrology, erosion, and wind-driven fire in raw WebGPU. A convolutional neural operator trains against the solver, exports to hand-written WGSL compute passes, and races the real physics with a live error field. No servers, no runtime dependency.

WebGPUNeural OperatorsPyTorchWGSLShallow-water

genesis

artificial life lab

2026

Particle Life on the GPU: flocking, chase, and predation waves, summoned from a text prompt. In-browser CLIP scores candidate worlds while a separable CMA-ES search breeds the simulation toward the description. Entirely client-side.

WebGPUCLIPCMA-ESEmbedding RetrievalArtificial Life

agri-iq

Research

ML research on smart ear tags for livestock: pregnancy prediction and health indexing from biosensor time series. Pre-commercial R&D on cattle, sheep, and goats.

Time SeriesIoTPyTorchAgricultureLivestock

modaics

2024

Digital wardrobe and sustainable-fashion marketplace with AI styling. Built through the RMIT Activator accelerator program.

Computer VisionNLPReact NativeSwift