ArtisanCAD uses expert knowledge to turn vague prompts into production CAD models

ArtisanCAD tackles a key limitation in text-to-CAD systems: handling vague, high-level design prompts. The core innovation is CAD-IR, an executable procedural representation encoding parameters, operations, dependencies, and verification rules. This representation serves dual purposes: it first captures expert CAD procedures (from CATIA macro logs, drawing notes) as parameterized skills, then provides a scaffold to translate ambiguous user intent into complete executable operations. The agent retrieves relevant skills, instantiates and revises CAD-IR, executes through a CATIA-MCP backend, and uses multi-view visual feedback for iterative refinement.

On the Text2CAD benchmark, ArtisanCAD reduced mean Chamfer Distance from 14.83 to 9.88 for intermediate-level prompts, demonstrating its ability to bridge ambiguous text and precise CAD construction. For real-world validation, the team used four complex automotive components to show that expert CATIA recordings can be distilled into reusable skills, enabling generation of editable CATIA-native B-Rep models for new variants. This represents a significant step toward industrial-grade AI-assisted design automation, where expert knowledge isn't lost but systematically leveraged.