AI engineer
The one who puts a model they did not train to work
An AI engineer builds a product around a model they did not train and have no intention of training, like a cabinetmaker who buys timber rather than planting the forest.
- Floor 2The Harness
- P2Writes the application
- Established titleIt has been in adverts long enough to mean something.
What they answer for
The deliverableA feature that answers real users, with its cost per call and its quality measurement. If they leave, the product stops moving.
The profile
Five activities, scored 0 to 3- ResearchProducing knowledge that does not yet exist.
- BuildShipping a system that runs, deploys and breaks.
- OperateKeeping it in production: cost, incidents, drift, on-call.
- VerifyMeasuring, testing, attacking. Producing a verdict that holds.
- LeadDeciding, persuading, driving adoption, answering to others.
What the work is
The reversal
The role was born of a reversal. For ten years, doing AI meant training a model on your own data. Now that general-purpose models are rented by the call, the question has changed. It is no longer "how do we train this" but "how do we wire it in, how do we measure it, and how do we avoid burning the company’s money in tokens".
The border
The border with the ML engineer is training, and nothing else. An AI engineer picks a model, writes what goes into its window, constrains what comes out, and wires up tools. They touch neither the parameters nor the architecture. It is integration work in the full sense, which does not make it easier: nothing resembles a test suite less than a product that answers in prose.
The trap
The trap is the title itself. It pays better than "ML engineer" for comparable work, so it lands on roles that have nothing to do with it. One advert describes data analysis, another describes pre-sales. The deliverable is the only reliable test. Ask what breaks if the person leaves.
A week in the role
- Production code: calls to the model, guard rails, retries when the answer does not hold the expected shape.
- Evaluation sets to keep current, because no unit test tells you whether an answer is good.
- Watching the cost per call, the one metric a finance function asks for unprompted.
- Back and forth with product about what you promise the user when the model gets it wrong.
Ways in
- From application development, with no detour through machine learning. This is the shortest and commonest way in.
- From data science, accepting that the deliverable changes: you no longer hand over a conclusion, you hand over a service.
- What the role does not require, whatever the advert says: a doctorate, or a single publication.
Reading an advert
3 signs- The advert asks you to train and deploy models "from scratch".
- That is an ML engineer role under a better-paid title, or an advert written without technical review. Both deserve a question at interview.
- No mention of evaluation, test sets or quality measurement.
- The role has not met production yet. You will build the measurement yourself, which is good work but not the work described.
- The advert lists fifteen libraries and not one problem.
- Tooling turns over every six months, the problem does not. An advert that will not say what has to work does not yet know what it is hiring for.