Senior AI Engineer

You will own significant parts of the Wren runtime. That is the system that watches a building, reads its record, drafts the filings and narratives a licensed professional will sign, and prepares the package for human review. It is not a chatbot. It is a careful research assistant with a long memory, and the standard it has to meet is the standard an architect or engineer would apply to their own work.

Day to day you will be writing production TypeScript and Python, designing tool interfaces for Claude-based agents, wiring retrieval against a growing corpus of statutes and filings, and thinking hard about where the system should stop and ask a human. Much of the work is unglamorous: improving citation accuracy by a few percent, making a draft read less like a model output and more like something a practitioner would write, closing loopholes where the agent could confidently produce something wrong.

We ship slowly on purpose. A typical week involves more reading and writing than coding. You will be expected to write the design document before you write the patch, and to defend the design in prose before anyone reviews the code.

Who you’ll work with

You will work closely with our research lead, who owns the building-side knowledge, and with our founding engineer. You will also spend time with the licensed professionals who use Wren on their own buildings, because their feedback is the only signal that matters. Expect to sit in on review sessions where an architect goes through a draft filing line by line.

What you’ll do

In your first six months you will take ownership of the draft-and-review loop: the path a filing takes from the first retrieval, through tool use, through the intermediate critique step, into the format a licensed reviewer expects to see. You will harden citations so that every claim ties back to a statute section, a prior filing, or an on-building observation. You will design evaluations that look more like professional exams than machine-learning benchmarks.

Beyond that loop, you will help shape how the runtime handles uncertainty. That means building the parts of the system that decide when not to draft, how to flag a condition the agent cannot assess, and how to surface an open question to the licensed human in language a practitioner would use. You will also contribute to our internal tooling for prompt and tool-use regression, because the only way we keep this system honest is by catching quiet regressions before they ship.

What we’re looking for

Seven or more years of software experience, with at least a year of serious work on LLM-based systems in production. You have built something people depend on. You have debugged a tool-use loop at two in the morning. You have opinions about evaluation that go beyond accuracy on a held-out set. You write well enough that your design documents do not need a rewrite before they can be reviewed.

In character, we are looking for someone who would rather be slow and right than fast and impressive. Someone who reads the statute before touching the prompt. Someone who, given a choice between a clever architecture and a legible one, picks the legible one. If you are uncomfortable shipping work that another professional will have to countersign, this is the wrong role.

Nice to have

Prior work in a regulated domain, legal tech, medical records, or building systems. Familiarity with Anthropic’s Claude models and the Managed Agents pattern is useful but not required — we will teach it.

Compensation and benefits

Competitive cash for a senior IC at a small, well-capitalized company, early-stage equity on a standard four-year schedule, full health and dental, 401(k) with match, five weeks of paid time off, and an office in lower Manhattan you are welcome to use as much or as little as you like. Hybrid by default, fully remote considered for the right candidate.