This is a senior individual contributor role for an engineer who arrived at production work through research.
The foundation is a first-principles understanding of how these models actually work. When behavior breaks, that understanding is what turns guesswork into an explainable engineering decision, reasoning through tokenization, attention, sampling, context window, and training distribution. It is also what model strategy rests on: whether to stay on hosted models or invest in fine-tuning and self-training, a decision this role owns and defends with data, cost, and technical trade-offs.
What this role is hired for is that depth pointed at a product. Model selection and prompt tuning are the starting point, not the substance. The substance is the design of AI behavior for specific business contexts: defining what the agent is permitted to do in a given scenario, where its boundaries sit, when it must refuse or escalate, how it handles ambiguity and adversarial input, and how each of those decisions is measured. Scenario-level configuration of this kind determines whether the product is trusted in production.
The expected pattern of work is to read the literature, form a position, run the experiment, and land the result in production. Direction is set by this role rather than handed to it.