Detailed Analysis
Adam, a Y Combinator Winter 2025 company founded by Zach, has launched CADAM, an open-source text-to-CAD platform that uses AI agents to generate parametric 3D models from natural language input. The system operates on a "text-to-code-to-CAD" pipeline, converting user prompts into OpenSCAD code that is then compiled and rendered directly in the browser. The platform supports multiple input modalities, including natural language descriptions and image references, and exports to a broad range of formats including STL, SCAD, OBJ, GLB/GLTF, FBX, and DXF. Built as a React application using TanStack Start with a Supabase backend for authentication, database, and file storage, CADAM positions itself as an AI-powered analogue to TinkerCAD, targeting accessible, browser-native 3D design.
A technically notable architectural decision is CADAM's approach to parameter editing. Rather than invoking a large language model for every user interaction, simple parameter adjustments — such as moving a dimension slider — are handled through deterministic regex operations on the underlying SCAD source code. This design significantly reduces latency and cost for iterative design work, reserving LLM calls for substantive generative tasks. The system runs OpenSCAD compiled to WebAssembly inside a Web Worker, keeping the browser UI responsive while rendering is handled by Three.js via React Three Fiber. The platform is model-agnostic, routing through the Vercel AI SDK to support Anthropic's Claude, Google's Gemini, and OpenAI-compatible models via OpenRouter. Notably, internal evaluations found Gemini 2.5 Pro to be the top-performing model, a finding that cuts against assumptions of Claude or GPT-4-class models dominating code-generation tasks in specialized domains.
The code-as-CAD paradigm Adam has adopted reflects a broader shift in how generative AI is being applied to structured, constraint-driven creative domains. Traditional CAD software has long relied on direct manipulation interfaces, but the emergence of capable code-generating LLMs has made programmatic geometry description — through tools like OpenSCAD, CadQuery, and build123d — a viable alternative for AI-mediated design. By generating code rather than geometry directly, the system gains reproducibility, editability, and parametric flexibility that mesh-generation approaches lack. The roadmap toward supporting CadQuery and build123d specifically signals an intent to move beyond constructive solid geometry (CSG) primitives toward more sophisticated constraint-based modeling, which would bring CADAM closer to the capabilities of professional-grade CAD environments.
The open-source release and YC backing together position Adam at a strategically important intersection: AI infrastructure for physical product design. While AI code generation for software has matured rapidly with tools like GitHub Copilot, the equivalent tooling for mechanical engineering remains nascent. The mechanical CAD market is dominated by expensive, proprietary incumbents such as SolidWorks, Fusion 360, and CATIA, and the barrier to entry for AI-native challengers is high given the domain complexity of tolerances, manufacturing constraints, and spatial reasoning. Adam's approach — starting with an accessible, browser-based open-source tool — mirrors successful developer-tool go-to-market strategies, building community adoption before addressing the harder problems of professional-grade precision. The planned additions of face and edge selection UI and viewport image integration indicate awareness that spatial grounding remains a core limitation of current LLMs in geometric reasoning tasks, and that hybrid human-AI interaction models will be necessary to close that gap.
Read original article →