We build agentic CAE/CFD pipelines and surrogate models, then run studies on them end to end.
Each study is documented method-first: what we ran, how it was verified, how the two methods compared, and where each reaches its limits.
A genetic algorithm drives a fast BEMT evaluator to a Pareto front. A blade-resolved RANS run then verifies the winner, with review agents running the whole loop.
A post-processing sign error slipped into an automated study. A verification agent flagged it as physically implausible before it reached the report.
A surrogate trained on high-fidelity thermal runs predicts component temperatures in milliseconds. The hard part is knowing exactly where to trust it.