Selected work

Research

My research focuses on making 3D geometry more accessible, useful, and interpretable, from CAD representations to language-guided material understanding.

Point cloud samples of diverse CAD models from Figure 1 of the Better STEP paper
01 / CAD representations

Better STEP

Better STEP introduces an open, HDF5-based representation of boundary representation (B-Rep) geometry and an associated dataset and Python library. It makes CAD geometry and topology easier to query, sample, and use in learning pipelines.

Figure: Better STEP paper, Izadyar et al. (2025).
Two material assignments for a 3D cooking pot, as shown in the LLM-Guided Material Inference paper
02 / Multimodal 3D understanding

LLM-Guided Material Inference

A two-stage, zero-shot approach to assigning plausible materials to parts of 3D point clouds. The model first reasons about object semantics, then uses that context to infer material compositions for segmented parts.

Figure: LLM-Guided Material Inference paper, Izadyar & Schneider (2025).
Conceptual illustration of a CAD modelTOPOLOGY · GEOMETRY · MESHING
03 / Geometry processing

Topology-First B-Rep Meshing

A method that preserves the topology encoded in boundary-representation CAD models throughout the meshing process, while controlling geometric approximation with a user-defined tolerance.

Illustration is conceptual; see the paper for original figures.