Vibe-CADing is a research-driven system that transforms natural language into functional 3D CAD part designs. It combines state-of-the-art generative models, including VAEs, diffusion models, geometry-based retrieval, and LLMs, to generate, refine, and retrieve CAD components aligned with human intent.
Inspired by the philosophy of vib-coding (coding with intuitive creativity), Vibe-CADing extends that concept to the CAD world, empowering users to describe rather than draw. Whether it's “a slotted bracket with curved ends” or “a balanced support beam,” the system converts that intent into CAD-like voxel representations, suggests related parts from open-source libraries, and refines designs based on feedback (e.g., symmetry, balance, or user edits).
The long-term vision is to support:
Text-to-3D generation
Interactive feedback loops
Multi-part assembly understanding
LLM-guided refinement and GUI-based conversations
Manufacturability gap in SOTA generative models for 3D objects
Era Key Innovations Description
1960s–70s Text Commands Tools like Sketchpad (1963) enabled light-pen input but lacked GUIs—CAD was command-line only.
1980s 2D GUI + AutoCAD AutoCAD (1982) introduced a graphical interface, marking the shift toward visual design tools.
1990s Parametric CAD SolidWorks, CATIA enabled parametric modeling with part constraints and feature history.
2000s Design Automation Scripting (e.g., OpenSCAD) and simulation tools streamlined repetitive or functional design.
2010s AI/ML + Web CAD Tools like Onshape and Fusion 360 cloudified CAD and introduced ML for part recognition.
2020s LLM + Diffusion The rise of Text-to-CAD systems (e.g., ShapeCrafter, DeepCAD) brought language into design via GPT, CLIP, and diffusion.
The rise of generative AI has unlocked the ability to bridge abstract human intent and concrete geometric structure. CAD, once restricted to skilled designers, can now become language-accessible, intuitive, and collaborative.
Vibe-CADing positions itself at this frontier—augmenting traditional CAD workflows with intent-driven generation, interactive feedback, and scalable retrieval to make design faster, smarter, and more human-centered.
Not at all! Like many other AI-driven systems, several existing tools attempt to convert text into 3D shapes or CAD models. However, when it comes to product-level design, manufacturability, and most critically, geometric precision, current models often fall short of what's needed in industrial applications.
Moreover, just as early LLM-based chatbots lacked access to search tools before RAG became standard, today’s CAD generation tools typically overlook the potential of open-source CAD libraries. A truly capable CAD agent should be able to search, retrieve, and reason over existing parts to accelerate design. Yet, as far as I know, no system has seriously attempted this—likely because retrieval based on geometry is far more complex than retrieval based on text semantics.
In a more product-focused implementation of Vibe-CADing, I've developed a modular pipeline using LangChain and LangGraph that breaks the problem into composable agents. The workflow consists of:
Part specification extraction from natural language prompts
Retrieval of related CAD models from open-source libraries
Geometry generation using diffusion or VAE models
Constraint checking and refinement
To support multi-step collaboration between components, I used LangGraph to implement a multi-agent refinement loop, where four agents interact:
Planner: Interprets the user’s intent and decomposes it into actionable subgoals
Retriever: Searches for similar CAD parts across external libraries
Generator: Produces a new geometry from scratch or by adapting retrieved parts
Critic: Evaluates the output against constraints (e.g., symmetry, balance)
Each agent operates with a defined role, state memory, and a prompt template. Together, they collaborate to translate user language into viable, meaningful 3D part designs—iteratively improving the output with feedback, context, and retrieval augmentation.