About

I’m an AI engineer. I develop machine learning models and design the AWS architecture that runs them, which in practice means owning a problem from the data through to a service a team can depend on. Both halves matter: a model that scores well and a system that stays up are separate achievements, and shipping requires them together.

Background

More than five years as a data scientist and AI engineer, building machine learning solutions that mostly run on AWS. I came up through the modelling side, then spent enough time on deployments to learn that the model is usually the shorter half of the work.

Based in South Africa, working with teams locally and remotely.

What I do

On the modelling side: framing the problem, preparing the data, selecting and evaluating candidate approaches, and being honest about what the evaluation does and does not prove.

On the systems side: designing the AWS architecture that serves the result — deployment and versioning, inference cost, monitoring, retraining, and defined behaviour when a dependency is slow or unavailable. Reliability is a design decision made early, not something added once a system is already struggling.

How I work

I write the evaluation before the feature. It is slower for the first two weeks and faster for every week after, because it turns arguments about whether a change helped into a number that either moved or did not.

I also plan for the system’s second year, not just its launch. Models drift, data changes shape, and costs grow quietly. Deciding early how those are detected and handled is far cheaper than discovering them from a customer.

About the projects on this site

They are chosen deliberately. Each one is a system built to run in production rather than a proof of concept — the difference being that a POC answers “can this work”, and a production system has to answer “what happens when it doesn’t”. Plenty of machine learning work demonstrates well and then does not survive real users, real load, or a real budget. These are the ones that did.

Each write-up sets out what was measured and how, and ends with what I would do differently. That last section is the one I would read first if I were the one hiring.

What I’m looking for

Work on machine learning systems that have to be dependable: taking a model from prototype to production properly, or improving one that is already there and not behaving. I am most useful where a system exists but is not doing what it should — costs drifting, accuracy quietly degrading, or a team unable to tell whether the last change made things better.