The work

You are a task designer for a benchmark measuring whether advanced AI systems can do research-level computational engineering with real scientific software. Day to day, that means picking a workflow you know cold — a Timoshenko beam under a nonstandard load path, a mesh convergence study, a reacting-flow calculation in Cantera, a reliability or layout optimization in Pyomo or gurobipy — and turning it into a problem with a verifiable answer and a defensible solution path.

Two problem archetypes recur. The first is fully specified: the setup is complete and the difficulty lies in executing a long multi-step numerical workflow without drifting. The second is interactive and harder to build: information is deliberately hidden, and the model must plan a sequence of simulations or measurements to infer it — deciding what to probe, reading partial results, and narrowing the space efficiently. You'll write the environment, the oracle function that produces ground truth, and the validator that accepts correct answers arriving by different routes.

  • Draft problems grounded in workflows you've actually run, not textbook exercises
  • Run them against state-of-the-art models, inspect failure traces, and diagnose whether a miss reflects genuine difficulty or ambiguous specification
  • Iterate until the item lands in the target difficulty band — hard for reasoning, not merely long
  • Package everything reproducibly: pinned dependencies, deterministic seeds, containerized where required

What the screen looks for

Mercor's process is AI-led and follow-up heavy. Expect it to push past tool name-dropping toward specifics: which solver you configured, which formulation you chose and why, where the library broke, how you established that your numerical result was actually converged. Claims that don't survive two or three follow-ups tend not to advance. The other thing under evaluation is design judgment — can you distinguish a problem that's hard because it demands careful reasoning from one that's hard because it's tedious, and can you write a validator that doesn't reward the wrong answer arrived at by the right-looking method.

Logistics

Fully remote and asynchronous, with review cycles rather than fixed shifts. Listings in this family cite 15–20 hours per week minimum; observed rates for this band run $70–85/hr, set per engagement and not guaranteed. You'll work in a Linux/terminal environment with remote compute sandboxes, so comfort with CLI workflows and containerized environments is assumed rather than trained. MS or PhD in a relevant field, or equivalent research track record, is the baseline.