Discovery, architecture, a prototype that can be scored, then a rollout that stays up. The work is full-stack when the interface is the delivery, and it is infrastructure when the model will not stay honest without it.
You write the code. You also write down the trade: what shipped, what was refused, and which eval the live environment changed.
What you will do
- Own a deployment from the first architecture note to a production system with a health check and an owner on the client side.
- Write and review production code across the services the delivery actually needs — Python, TypeScript, the stack the client already runs.
- Embed with their engineers. Map the workflow. Remove the blocker before it becomes a status slide.
- Turn a pattern that worked into a library, a playbook, or an internal SDK so the next estate is not a rewrite.
- Bring eval numbers and operator notes back to the people who train and productise the models, so the next build is not a guess.
What you bring
- Five or more years shipping full-stack applications and cloud services — Python, TypeScript, Node, React, or the equivalent you will defend.
- LLM or foundation-model systems in production, and a clear view of how non-deterministic output changes the interface and the tests.
- Microservices, APIs, containers, and an observability path you have actually paged on.
- The ability to turn a workflow into a software requirement, and to say the same thing to a sponsor and to the engineer who will merge the PR.
- A record of simplifying a system, naming the deployment risk early, and cutting scope when the calendar and the quality bar cannot both move.
Useful, not required
- AWS, Azure or GCP; Kubernetes; Terraform — used on a live estate, not a tutorial.
- Eval harnesses, prompt pipelines, or retrieval systems on a workflow where a wrong answer costs money or trust.
What to send
- A CV and the two profile links the form asks for.
- One production system you took from a sketch to a release the client still runs — what you measured, and what you would not ship.
The first quarter
The first month is the model APIs, the deploy path, and the security bar on the accounts we already hold — you sit in on a live cycle before you lead one. The second is a customer integration you own: a prototype in production with evals attached. The third is that release made stable, and the lesson written into a tool or a note the next FDE can use.