Meramia already works inside construction and property teams. This seat owns the product decisions for that work: what gets automated, what stays a human call, and how a model is scored before it touches a safety or a commercial decision.
You will write the roadmap for tools that take field capture, plan sets, submittals, estimates and progress and turn them into something an operator can act on. The harness has to ground language and vision models in BIM, CAD, the schedule baseline and the feeds that actually exist on the project — and it has to fail loudly when the data is not there.
What you will do
- Set the sequence: field-capture processing, plan parsing, submittal review, cost, progress — and say which of those is a harness problem this quarter.
- Specify developer-facing and internal harnesses that bind LLMs and vision-language models to the project record, so a hallucination is a test failure rather than a site surprise.
- Stand up agents that can flag a hazard, follow a delivery, check a build against the drawing, or draft a change order — always with an eval attached.
- Go to site. Talk to project managers, superintendents and subcontractors. Write the bottleneck down before you write the ticket.
- Own the evals for accuracy, safety and reliability wherever the output can move money or injure someone.
- Work with the people who sell and support the work so hours saved, risk avoided and rework cut are measured, not asserted.
What you bring
- Three or more years running AI or ML products: agent architectures, retrieval, computer vision, and an eval harness you have actually used.
- The construction lifecycle — pre-construction, bid, schedule, execute, handover — and the tools the trades already live in: Procore, Autodesk Construction Cloud, Primavera P6, Revit, OpenSpace.
- A plan for unstructured site data: photographs, 360 scans, drone footage, PDF specs, BIM metadata.
- Four or more years of product work in B2B software, PropTech, ConTech, or developer infrastructure.
- Interviews, prototypes, a backlog you will defend, and the ability to talk to researchers, engineers, designers and a superintendent in the same week.
Useful, not required
- Civil engineering, construction management, architecture, or computer science as a first training.
- Prompting and model-eval tools (LangSmith, MLflow, a custom CLI) used as instruments, not as theatre.
- A zero-to-one AI product shipped into a trade that still runs on paper and a WhatsApp group.
What to send
- A CV and the two profile links the form asks for.
- One product you owned where the model had to be scored before it reached an operator — what you measured, and what you refused to ship.
The first quarter
The first month is discovery on live accounts and an audit of the current models and their evals. The second is a twelve-month map: which agentic workflows, which harness gaps. The third is one shipped feature — a plan-set delta check or a live risk alert — with retention and accuracy written down before it leaves the bench.