About this role Interview language: English AI agents are getting good at reading a screen and clicking through web forms. Put one in front of a parametric CAD package and it falls apart.
About the role
About this role
Interview language: English
AI agents are getting good at reading a screen and clicking through web forms.
Put one in front of a parametric CAD package and it falls apart.
FreeCAD is an unforgiving test case: multiple workbenches with overlapping capabilities, a constraint solver that fails in ways the interface doesn't explain, feature histories where the correct next action depends entirely on what was built twenty steps earlier, and the topological naming problem lying in wait for anyone who references geometry carelessly.
Engineering intent lives in the model tree, not on the screen — and that is exactly what current agents cannot see.
In this remote, hourly contract role you will help close that gap. You will design realistic, challenging CAD tasks, then record yourself completing them on your own computer using recording tools we provide.
Each recording captures the full picture of how expert work actually gets done, the sequence of actions, the interface state at every step, and the files before and after. It becomes both a demonstration the model learns from and a benchmark it is measured against.
Because a model can only learn from work that can be scored, an equally important part of the job is defining what "done correctly" means.
Alongside each task you will write the objective and the steps you took. You will specify how the result should be verified, what the finished part, sketch, assembly, or drawing has to satisfy for the outcome to count as correct, and which intermediate checkpoints a partially-successful attempt would hit.
In CAD this is a sharper question than it looks. A part can be dimensionally correct and still be wrong because the sketch is under-constrained, the feature order won't survive a parameter change, or the model references a face that will migrate the moment anything upstream is edited.
Encoding that distinction, dimensionally right versus actually right, is the contribution we need from you. …
Official source: www.jtlabs.ai. Details can change; the platform's page is the final word. How we check listings.
