What the work actually is

You author computational problems, not labels. A typical task starts from a real workflow you've run — fitting a two-compartment model with first-order absorption, diagnosing a stiff ODE system that silently fails to converge, teasing apart competing versus uncompetitive inhibition from limited assay conditions — and turns it into a self-contained problem with a verifiable answer. Some tasks are fully specified: the model, parameters, and dosing are given, and the AI must execute a long multi-step simulation and analysis correctly. Others are interactive and harder: parameters are hidden behind an oracle, and the model must decide which simulations or perturbations to run, read partial results, and narrow the parameter space efficiently. You write the SBML or Tellurium/Antimony setup, the oracle function, and a validator that accepts correct answers within a defensible tolerance and rejects lucky guesses.

Every problem then goes through a calibration loop against state-of-the-art models. If frontier models solve it on the first attempt, it's too easy; if no approach can succeed because the setup is ambiguous or numerically ill-posed, it's broken rather than hard. Most of the craft is in that gap — making difficulty come from reasoning about identifiability, timescale separation, or experimental design rather than from tedium or arithmetic volume.

What the screen looks for

Mercor's screening is AI-led and pushes on specifics. Expect to be asked which libraries you've actually written code in, what you built with them, and where those tools misbehave — solver choice for stiff kinetics, events and rate rules in libRoadRunner, unit and annotation pitfalls in SBML round-trips, steady-state solver failures. Claims are checked against publications, repos, or described professional work. You'll also be probed on evaluation judgment: whether you can tell a hard problem from a merely long one, and whether your validator would survive a model that reasons its way to a numerically equivalent but differently-expressed answer.

Logistics

  • Fully remote and asynchronous; work is delivered as problem bundles, not shifts.
  • 15–20 hours per week minimum, sustained over the project rather than in bursts.
  • Work happens in Linux/terminal remote compute sandboxes; containerized, reproducible setups are a plus.
  • Observed band is $70–85/hr, typically varying with depth of tool experience and demonstrated task quality. Not guaranteed.