ML / AI Infrastructure Engineer

The problem.

A model that cannot be reproduced is a model nobody can debug. A model that degrades quietly is worse than one that stops. Infrastructure must make reproduction cheap, deployment boring and failure loud. Today the training path and the inference path are further apart than they should be.

What you would do.

• Own the training pipeline: fixed seeds, recorded data slices, an artefact that carries enough provenance to be interrogated a year later.
• Own the inference path, compiled and inside a strict latency budget, with the numerical behaviour of training preserved.
• Build the refusal behaviour: when a window cannot be fitted, the trainer raises rather than emitting a constant model the rest of the system will trust completely.
• Monitor for drift in the inputs, not only in the output, and make the alert actionable.
• Manage whatever accelerator capacity we run, on cost as well as on throughput.

What we would want evidence of.

• You have reproduced a model fit from six months earlier, and know why that was hard.
• You have deployed inference into a latency-sensitive path and can describe the budget.
• You have built a system that refuses rather than degrades, and defended the choice.
• You are careful about training and serving skew from feature code existing twice.
• You know what your accelerators cost per useful experiment.

What would make this a bad fit.

This role is a poor fit if you expect a large cluster. Hardware here is modest. The constraint is usually the data rather than the compute.

This role is a poor fit if model serving means a container and an endpoint. Requirements here are stricter and will feel fussy at first.