Founding AI Engineer

Job description

The problem The construction industry loses $1.6 trillion a year to inefficiency, and most of it is avoidable. The knowledge that would prevent it already exists. It's just trapped in PDFs, meeting protocols, and email threads nobody can search. Every new project relearns what the last one already knew. We're building the opposite of that: software for the people who run large, complex construction projects. It reads the documents, tracks the decisions, and catches the mistakes early. The model isn't the moat. The project memory is, and it compounds: every project makes the system better, and a competitor starting today is already years of projects behind. Our software runs today on a live autobahn construction program and an S-Bahn transit program: multi-year timelines, hundreds of thousands of pages of specs, protocols, and site communications, and real consequences when we get an answer wrong. We're pre-seed, led by Realyze Ventures, whose LPs include Zech and other large European construction groups. Co-investors: D11Z, the family office behind Aleph Alpha, and the CDTM Venture Fund, backed by 300+ CDTM alumni including the founders of Personio and Alasco and DeepMind's Technical Director. 25+ live customers. Tasks What you'd work on Two problems, both of which need state-of-the-art answers. Agent harness engineering for construction documents. One summary doesn't fit all. "Structural risk" means something different in an RFI, a cost review, and a schedule reconciliation. We build multi-agent harnesses (specialized extraction, reasoning, and evaluation stages) that route a 400-page tender document or a protocol archive into the right pipeline with the right context. Part of that is context compression : what the system should remember, forget, and surface, and at which decision point, when a project runs five years and touches 50 stakeholders. There's no clean top-k answer, so this gets solved in production, not in a library. If you've read what Anthropic and Vercel have written about agent harnesses and thought "yes, that's the hard part of shipping production agents," this is the job. Project memory as a compounding moat. We started with meeting transcripts. Now we're building a decision graph that grows with every project: not just what was decided, but why, by whom, against which alternatives, and how it played out. That graph feeds the next project…