A field of sunflowers beneath a towering cumulus cloud, posterized into flat bands of teal, cream and yellow.

I build and evaluate ML systems.

Rust and Python. Training pipelines, generative models, evaluation.

↑ posterized to gruvbox

I write Rust and Python, mostly for machine learning work.

Most of what I do is training and evaluation pipelines. I usually end up writing the evaluation as well as the model. On my thesis that turned out to be the part that mattered: the standard utility metric ranked the worst model first.

Right now I am working on running inference in the browser. Rust compiled to WebAssembly, WebGPU where a GPU is needed.

Work

13× Generated recommendations for niche users were 13 times more popular than those users' own histories. Real behaviour drifts 0.45. All three architectures produced the same number.

Synthetic session generation

MSc thesis. A multi-head autoregressive transformer that generates synthetic e-commerce sessions for recommender simulation. Trained on 225 million interaction events. A constraint layer keeps generated sessions structurally legal.

The evaluation is the more useful half: five seeds, pre-registered pass conditions, and a result showing that the standard utility metric ranks a mode-collapsed model first.

Python · PyTorch · Evaluation design

Elsewhere

github.com/mofumail