I build machine learning and LLM systems and take them all the way into production, working across Adelaide, Melbourne and Sydney. The model is usually the smallest part: around it sit the retrieval layer, the API, the database, the cloud it runs on and the cost of running it. I build those too.
Projects worked on, solo and in teams
AI and machine learning systems built
Years shipping AI into production
Retrieval, prompt design, model routing and evaluation, including real-time voice and avatar agents running against a custom model.
Computer vision models trained, evaluated and deployed to run on the device rather than in a data centre.
Pipelines that take messy event data and turn it into something you can actually query and trust.
The APIs, cloud infrastructure and interfaces that turn a model into something people can use.
Systems that made it into real use, and the problem each one had to solve.
A talking avatar backed by Claude, with hybrid retrieval feeding it and speech going both ways. Answers stay grounded in a document set, and the system picks a model per question so quality holds without the cost running away.
The same idea delivered as an illustrated character rather than a photoreal one. The performance is assembled from pre-rendered art and cut-out layers driven by the conversation, so a client gets a character in their own house style without a live 3D pipeline behind it.
Ask a question in text, by voice or with a photo and get an answer grounded in an uploaded knowledge base, with real links back to where it came from rather than invented ones.
Vision models that identify disease from a field photo and run on the phone itself, because the people using it have no signal. Built the dataset and annotation pipeline behind it, then the training, evaluation and on-device export.
Extracts structured fields from a photographed document across four scripts and checks them against thirteen legal requirements. Two passes must agree before a field is confirmed, so it reports what it could not read instead of guessing.
A camera and accelerator running detection continuously at the edge rather than round-tripping to a data centre, built as a persistent streaming service after the obvious approach turned out to be racing itself.
A router that picks the cheapest model capable of answering and escalates only when the question needs it, with every decision logged. Alongside it, spend caps that pause the expensive call while keeping search working.
A platform for images, video, audio and 3D models that captions, tags, groups faces and makes everything searchable by what is actually in the file rather than by who remembers the filename.
Millions of behavioural events arriving from browsers, mobile apps and headsets. The pipeline deduplicates them, rebuilds the sessions behind them and makes the result queryable, feeding dashboards and written summaries.
A national estimate of depth to groundwater at any location, built from 35,875 monitoring bores with readings spanning two decades. Every answer ships with a confidence value derived from how many bores are nearby and how far away they are, so the estimate carries its own reliability rather than pretending to be a measurement.
A synchronised film running across a fleet of VR headsets on a moving bus, with no laptop, no server and usually no signal. One tablet is the whole brain, with native networking plugins written twice.
The cloud estate behind these platforms: 38 serverless functions, 43 storage buckets and 19 edge distributions, defined as code across eight stacks in two regions. I ran the audit that mapped it, attributed the cost project by project, and cut the bill where the numbers justified it.
Upload a document, get an AI summary, topics and tag suggestions, plus automatic detection of which existing document this one probably replaces. Scaffold to a passing end-to-end run in three days.
A native macOS app that stamps spherical metadata onto video without re-encoding, built on a Google tool that had been unmaintained since 2018 and needed porting forward first.
Alongside these I work on the real-time and device side when a project needs it: synchronised video across a fleet of headsets with no network, native plugins in Swift and Java, and headless build pipelines.
The work above came first on this page for a reason. The study is what made it possible to reason about the models rather than only call them.
Statistics, machine learning and the maths under the models. Where the training, evaluation and error analysis in the work above comes from.
Networking, distributed systems and cloud architecture. Where the other half comes from: the APIs, the deploy pipelines and the bill at the end of the month.
Rather than describe what I build, I built it. This is my own agent, sitting at the bottom of every page. It answers from my projects, and it tells you plainly when it does not know something.
Send me a quick note and my CV will reach you shortly.