About You
You want to work on a system of agents that work together. Some of them talk to customers, and many of them never say a word to anyone. They interpret, decide, and pass work to each other, and you understand that the quality of the whole system depends on every link in that chain holding up in production. Making that behave reliably, at scale, with real customers on the other end, is the problem you want to own.
You treat prompts as engineering artifacts, not text you tweak until a demo works. If you've ever shipped a prompt change that fixed one case and quietly broke nine others, you know why the evaluation layer matters as much as the prompt itself. And if you get a small thrill from a regression suite going green after a nasty bug, you'll feel at home here.
You'll also recognize how we work. We measure what we ship: changes come with evidence behind them, a dataset, a score, a clear before and after. We ship weekly, so ideas get tested against real production behavior quickly rather than studied at length. Craft matters, and so does rigor — getting an agent to sound right is satisfying, but proving it still sounds right across thousands of conversations is the actual work.
The surface area grows every quarter with new agents, new channels, and new products. You'll own a meaningful slice of it from day one.
You Have
- Hands-on experience building with LLMs in production, on systems that real users depend on day to day.
- Real evaluation experience: building datasets, writing evaluator prompts, defining rubrics, running experiments, interpreting results, and acting on them.
- Fluency with LLM observability and evaluation tooling: tracing, prompt versioning, datasets, experiment runs. The specific tool matters less to us than the habits.
- Experience with multi-step or multi-agent architectures: orchestration, tool and function calling, structured outputs, handoffs between agents.
- Experience designing or debugging voice agents, including latency budgets, turn-taking, and the constraints of real-time speech. Voice is a significant part of what we build.
- Comfort with prompt templating and conditional logic, and with treating prompts as versioned, reviewable artifacts.
- Systems thinking. You can hold a chain of agents in your head and reason about where a failure originated.
- Strong writing. Every word an agent says to a customer is a word you wrote.