Mudit Atrey

When global optimizers are coupled with low-order vortex lattice methods (VLM) without geometric regularization, 35.5% of evaluations waste compute exploring chord-inverted topologies with a 29% penalty in max L/D (12.89 vs 18.15) without solver divergence. How I formulated C1-continuous Hermite manifolds to enforce geometric invariants across 10,087 evaluations, and what post-submission 3D RANS CFD revealed about a 124% parasite drag rise at winglet junctions.

Generative 3D polygon meshes cannot be manufactured on physical CNC mills or lathe tooling; real engineering requires Boundary Representation (B-Rep) solids with exact datum planes, mate constraints, and editable parametric feature trees. Architecture of onshape-mcp: a Model Context Protocol server connecting LLMs to the Onshape cloud REST API for bidirectional feature-tree tracking and topological constraint enforcement.

Lessons from 5 merged pull requests in the scikit-learn ecosystem (skrub and skore). How unintended eager evaluations on Polars lazyframes degrade memory locality during complex tabular feature encoding, and designing figure legend export persistence across matplotlib backends.

At 300 rps (18,000 RPM) muzzle rifling spin, canards cannot execute coherent trajectory corrections without actuator saturation. I decoupled the forward steering head from the projectile body using a 1-DOF roll-isolated collar on dual angular contact bearings. Supported by passive permanent magnet eddy-current damping, the collar attenuates roll from 300 rps to below 5 rps in 1.2 seconds, enabling terminal Augmented Proportional Navigation (APN) and 16-state EKF estimation through 25-second GNSS denial windows.