About

ML systems engineer; spatial, full-stack, simulation & applied AI.

Right now I'm building geospatial ML systems for climate and carbon work.

STA 01

At ArborMeta I work on geospatial ML and the systems around it: maps, pipelines, and models that have to hold up outside a notebook.

STA 02

I grew up on a cotton farm in Queensland. Initially I set out to become a doctor; three months in I found myself programming more than studying medicine, so I switched to computer science.

STA 03

The playground here is the same discipline off the clock: solvers on your GPU, neural operators trained to emulate them, racing in a tab.

STA 04

Open to contract work in climate tech and applied ML: spatial pipelines, fine-tuning, vector search, deployment that does not live on one API.

The stack

Five clusters that matter. Each one points at something real on this site. The full list is below.

Scaling laws · Research

Compute-optimal scaling, byte level

Byte-level Chinchilla-style sweep: 43 runs plus 18 seed repeats, 0.1M to 113M, 1e15 → 1e18 FLOPs. Local b 0.98 → 0.68; a browser parameter cap inverts the optimum. WLS b = 0.99 ± 0.09.

Scaling-law fittingIsoFLOP sweepsAdamW / cosine schedulesDivergence diagnosisPyTorch

Read the study →N_opt ∝ C^0.99 ± 0.09

Deep learning · LLMs

Pretraining under fixed compute

40M parameters, seven epochs, one seed. Validation loss down 33%. Muon, WSD, RoPE, ReLU², value residuals. Full ledger, failures included.

Muon optimizerWSD schedulingRoPEReLU² activationsAblation-driven tuningPyTorch

Read the report →1.7533 → 1.1754

Simulation · GPU

A neural Earth engine

Shallow-water hydrology, erosion, wildfire in hand-written WGSL. A neural operator trained on the solver, raced live in your browser.

WebGPU / WGSLNumerical simulationNeural operatorsProfiling & benchmarking

Run it live →60 fps

Modelling & Optimization

PageRank as a stochastic process

Random surfer past linear algebra: Gillespie simulation, exact Markov chain, simulated annealing to recover α. Live surfer included.

Markov chainsGillespie SSAMetropolis–HastingsEigenvector / spectral methodsSimulated annealing

Read the paper →α 0.730 / 0.730

The full ledger84 across 9 disciplines

Geospatial

9

PostGISMapLibre GL JSPMTilesGeoJSONEarth observation dataGDAL / Rasterio

Deep Learning & LLMs

14

Transformers (GPT-style)CNNsMuon optimizerNewton–Schulz orthogonalizationRotary embeddings (RoPE)ReLU² activations

ML / Data

13

PyTorchPython ML stackNumPy / PandasSpatial analyticsTime-series forecastingFeature engineering

Modelling & Optimization

8

Stochastic simulation (Gillespie SSA)Markov chainsMonte Carlo methodsSimulated annealingMetropolis–HastingsEigenvector / spectral methods

Systems & Performance

10

RustCC#Multithreading & concurrencyMemory optimizationSIMD / vectorization

Backend

9

FastAPIPythonNode.jsPostgreSQLREST APIsAsync I/O

Frontend

8

ReactNext.jsTypeScriptJavaScriptTailwind CSSWebGL

iOS

5

SwiftSwiftUICore MLCoreNFCOffline-first sync

Infra

8

DockerPortainerAWSLinuxNginxGit

Currently

Sydney, NSW → open to EU / US · AEDT / AEST · remote

Open for contract work

Start a project Email me