We propose an accurate and efficient neural operator architecture for learning PDE solutions on arbitrary domains. We demonstrate its effectiveness across a variety of challenging benchmarks, including large-scale 3D CFD problems.