AI Engineer at Nephew

## About the Project Nephew is the chief of staff small business owners can't otherwise afford. It lives in SMS, handles the inbox, manages bookings, follows up on leads, reconciles invoices, sends reminders, files things in the right place. The customer is a plumber, a dentist, a vet, a hairdresser, someone running a real business between jobs, who texts lowercase and short. We're building the agent that does the work they keep meaning to get to. A small team, a wide surface, and a customer base that's pulling us forward faster than we can build for them. ## What You'll Actually Do - Build the agent runtime: The part of Nephew that turns "follow up with the three quotes from last week" into real messages that landed in real inboxes and checks its own work before anything goes out. The agent has memory, acts on its own initiative when the moment calls for it, and stays inside the audience model whether an owner messaged first or not. You'll own how all of that fits together. - Author skills: Skills are the unit of composable behavior, how we teach Nephew to do a new shape of work. Booking flows, invoice chasing, lead triage, daily reflection, whatever the next vertical needs. You'll write them, refine them, and kill the ones that don't earn their tokens. - Build the eval and observability layer: This is where we catch regressions before owners feel them. Build the datasets, score what matters, watch the production traces, and close the loop from "this looked wrong" back into the skill that produced it. If something drifts, you want to know before the owner does. - Ship integrations: Connect Nephew to the tools small businesses actually use. The work is auth flows, webhook plumbing, mapping someone else's schema into ours, and handling the failures real APIs hand you. The proof point is a live owner's calendar on the first try, not a recorded fixture, not a demo account. - Everything else: A small team covers a lot of ground. SMS is the surface now; web, email, and voice are queued behind it. Plenty of the work that matters most this quarter isn't in this list. ## The Bar - Something you wrote goes live most days. A skill you authored is the reason an owner stopped manually chasing invoices. When Nephew sends something weird at 7am on a Saturday, you can trace it end to end, runtime, skill, memory, integration and fix it before lunch. The founders are writing less code because you've taken over surfaces they used to own. - Agentic coding has to already be how you work. We're not going to teach you the workflow on the side; the people we're hiring built parts of it. ## Day-1 Reality Day 1: laptop open, repo cloned, system prompt in one tab, production traces in another. You read a real owner conversation from yesterday, the actual SMS thread, the actual memory writes, the actual skill that fired. You find something that's off. A reply that was technically correct but landed wrong. A skill that loaded when it shouldn't have. A memory write that buried something