AI Engineer
You would build the harness, not the model. Context that reaches the agent before anyone types, routing keyed to the job rather than the application, and evaluation honest enough to tell you a cheaper model already cleared the bar.
Most of the week is spent making a model useful inside one specific company: pulling their real material into a harness, writing down what good looks like for a job they do every day, and then proving the output moved. Some of it is retrieval and prompt architecture; a lot of it is measurement, because without a bar, routing is guesswork with a confident tone.
- Build organisation, role and person harnesses — and the inheritance that stops a personal tweak overriding a company guardrail.
- Define a quality bar per job, then build the evaluation that checks whether a run met it.
- Route work across tiers so routine steps stop running on frontier models, with attribution per application and per person.
- Ship the utilities people actually open, in chat and embedded in a client’s own product.
- Write down what you learned in a form the next person inherits — the memory file is part of the job, not paperwork after it.
- Strong Python, and enough TypeScript to be useful where the harness meets an interface.
- Real experience putting an LLM system into production — retrieval, tools, structured output, failure handling. Papers are welcome; a thing that ran is better.
- An instinct for evaluation. If you have ever argued that a demo proved nothing, you will fit here.
- Plain writing. A finding an owner cannot act on is not finished.
- Comfort saying “I could not verify this” in writing, to a client.
Not required: a degree in the field, agent-framework brand loyalty, or benchmark trivia.
Send. Use the form below to send us one thing you have built and are proud of. We reply either way, within a week, with what we thought of it.
Review. Two paid days on a real repository through one lens, written up the way we write them.
Decide. A conversation with the two people you would sit beside, then a straight yes or no with the reason attached.