AI Researcher

Engineering · ResearchRemote or on-siteFull-time

Own how well our model reads a drawing. Take the takeoff from a scope an estimator describes to a full set it can measure on its own.

Apply for this role

About the role

Caliente is a copilot for construction takeoff. An estimator describes a scope in plain English, our model reads the sheet, and the takeoff lands as real, editable markups in the tool they already have open. Bluebeam Revu today, PlanSwift and Revit next, and the tool after that.

Reading the drawing is the part that carries. A sheet is linework, symbols, schedules, and a legend that only makes sense if you know the trade, drawn differently by every architect on every job. Once the model reads that the way a senior estimator reads it, the same understanding drives any tool we point it at, and every sheet it reads better is more scope covered end to end, more trades in reach, and quantities we can stand behind on a bid going out the door.

You will own that work. Not a slice of a roadmap someone else wrote, and not a lab off to the side either. What you build ships to estimators who are mid-bid.

What you would do

Own the reading problem end to end: geometry off the sheet, symbols, schedules, and the trade rules that decide where a boundary actually runs.
Keep that reading independent of whatever writes the markups. What the model understands about a sheet has to hold whether the takeoff lands in Revu, in PlanSwift, in Revit, or in the tool a customer asks for next.
Set the bar and then prove you cleared it. Build the eval sets, the ground truth, and the metrics that say whether a takeoff is right, because on a bid there is a real answer and we are either at it or we are not.
Take an approach from idea to shipped: prototype, measure, harden, and hand it to production. Nothing here is done when the notebook runs.
Push the automatic takeoff further every quarter. More scopes covered, less confirming, no invented quantities.
Work with the people who talk to estimators. The failure cases come from real sheets, not from a benchmark.

What we look for

A track record of training and evaluating models on messy real-world documents or images, in research or in production.
You are rigorous about measurement. You would rather find out your idea did not work than ship it and hope.
You can write the code that runs in front of a customer, not only the code that produces a number in a paper.
You are happy without a spec, and comfortable that the first six things you try may not work.

Nice to have, not required

Document AI, layout understanding, vision-language models, or OCR at production quality.
Publications at top venues (NeurIPS, CVPR, ICML, ICLR, ICCV, ECCV, SIGGRAPH) or serious open-source work.
Geometry, CAD, or PDF internals. Anyone who has fought a vector drawing knows why we ask.
Construction, estimating, or AEC experience of any kind.

Not sure you fit?

Nobody ticks every line. If the problem pulls you in and you can do most of it, apply. We would rather read your note and decide than have you rule yourself out.

Apply

Send us your CV.

Five minutes, one file. Applying for AI Researcher.