What the work actually involves

You write engineering design tasks that are hard on purpose — multi-constraint, trade-off heavy, with an explicit optimization target and a reference solution you have validated yourself. Each task ships as a self-contained environment: a simulation harness built on an open-source package, a Python test bench, and an autograder that decides pass/fail without a human in the loop. Then you run coding agents against it repeatedly and read the trajectory logs to see how they fail — misreading simulator output, converging on a design before the constraints are satisfied, producing geometries that cannot physically exist, silently mixing units.

The second half of the job is taxonomy and calibration. Failure modes get named and grouped so researchers can act on them, and problem difficulty gets tuned against empirical pass rates. The hard part of calibration is making a task harder without making it ambiguous or under-specified — a problem that models fail because information is missing is a broken problem, not a difficult one.

What the screen looks for

  • A master's or PhD in electrical, mechanical or aerospace engineering, paired with roughly a decade of hands-on design work — not management of design work.
  • Real fluency in at least one open-source simulator (ngspice/PySpice, OpenFOAM, FEniCSx, CalculiX, python-control, CadQuery/build123d, OpenModelica, Cantera, Gmsh) plus Python scripting strong enough to build a test bench around it.
  • Familiarity with LLM and coding-agent evaluation vocabulary: pass@k, failure-mode analysis, nondeterminism across repeated trials, reading execution logs rather than just final answers.
  • Evidence of rigor — convergence criteria, boundary conditions, unit consistency, mesh or solver settings you can defend under follow-up.

Logistics

Contractor engagement, up to 24 weeks, stated as 40 hours per week with four hours of overlap with PST; the listing also notes part-time engagement can be acceptable and that weekend on-call availability is required. Fully remote, on your own machine with reliable high-speed internet. Pay is not disclosed in the posting — treat any figure you hear secondhand as unverified. The process starts with a Job Interest Form for shortlisted candidates, followed by a delivery review before work is assigned.