HN Debrief

Launch HN: Adam (YC W25) – Open-Source AI CAD

  • AI
  • Developer Tools
  • Open Source
  • Hardware

Adam posted CADAM, an open-source text-to-CAD tool that generates parametric 3D models from prompts or reference images, emits OpenSCAD, and exposes extracted dimensions as sliders so users can tweak a part without another model call. It runs fully in the browser with OpenSCAD compiled to WebAssembly and supports export to common 3D formats. The company framed it explicitly as "AI TinkerCAD" for makers, while pointing to its commercial product for deeper integrations with Fusion, Onshape, and other CAD systems.

Treat tools like this as fast concept generators for printable parts, not as substitutes for engineering workflows that need constraints, tolerances, manufacturability, and verification. If you build in this space, the adoption hinge is not prettier demos but whether you can carry design intent into real CAD systems and cut verification time instead of adding more of it.

Discussion mood

Mixed but skeptical. People liked the polish, the open-source release, and the usefulness for makers and quick printable parts, but most experienced engineers rejected the broader framing because current systems do not capture tolerances, manufacturability, assemblies, or design intent well enough for real mechanical work.

Key insights

  1. 01

    Fast win for simple printable parts

    A concrete test with a split TPU cable grommet showed where this kind of tool can actually save time. The generated part was good enough on the first pass, made a sensible assumption about an omitted wire diameter, and surfaced that assumption as an editable parameter instead of hiding it in the code. That is a strong fit for quick one-off parts where "close, editable, now" beats perfect from the start.

    Use AI CAD first on low-risk printable parts with a few clear dimensions and material constraints. Judge it on whether the first output is editable in the right places, not whether it is fully correct unaided.

      Attribution:
    • patja #1
  2. 02

    Canonical source beats hidden render overrides

    Keeping the .scad file as the single source of truth is more important than using OpenSCAD's runtime parameter injection. The regex-based slider edits look crude, but they avoid a nasty state split where the rendered geometry no longer matches what the saved file says. For a code-first CAD tool, that consistency is the real product decision.

    If you build editable generated artifacts, make the saved representation and the visible output stay in lockstep. Do not introduce convenience layers that make the preview diverge from the underlying file.

      Attribution:
    • echoangle #1
    • zachdive #1
  3. 03

    Selection context is the real interface upgrade

    The promising part is not better prompting. It is combining prompts with explicit geometric references like faces and edges. That gives the model grounded context for operations across complex assemblies and tolerance edits, while preserving the direct manipulation CAD users already trust. The founder effectively conceded that natural language alone is not the ideal UX.

    For production workflows, prioritize interfaces that let users point at geometry and then express intent. Pure text is too ambiguous to carry precise edits reliably.

      Attribution:
    • zachdive #1 #2 #3
  4. 04

    Low sketch usage may reflect model limits

    The team's observation that users prefer text over sketches may be misleading. One commenter pointed out that professionals normally think visually, so weak adoption of sketch input could just mean the vision model is poor at turning drawings into usable CAD. That reframes product analytics. Users may be choosing the least broken modality, not the one they truly prefer.

    Do not read current interaction data at face value when one modality is technically weaker. Measure latent demand by improving sketch-to-CAD quality before concluding that text is the winning interface.

      Attribution:
    • 8note #1
    • zachdive #1 #2
    • jurgenaut23 #1
  5. 05

    GPL came from the OpenSCAD stack

    The open-source licensing choice was less ideological than infrastructural. The founder said they would have preferred MIT, but the OpenSCAD WebAssembly dependency is GPL. That matters because it shows how core technical choices can lock in downstream business and licensing constraints long before product strategy is settled.

    When choosing foundational open-source components, model the licensing consequences as product decisions, not legal cleanup. Your distribution and commercialization options may be set by one low-level dependency.

      Attribution:
    • mips_avatar #1 #2
    • zachdive #1

Against the grain

  1. 01

    Mechanical design is not the CAD model

    The harshest pushback was that geometry generation is the easy five percent. Real engineering work lives in requirements, tolerancing, materials, stock, tooling, fixturing, purchased parts, and the drawings or assemblies needed to build something. From that view, AI that drafts shapes but cannot carry those constraints is solving the wrong slice of the problem.

    If your buyer is a professional engineer, prove you reduce work outside raw modeling. Features that stop at geometry will be dismissed as side demos, not workflow tools.

      Attribution:
    • incorene2 #1
    • taneq #1
  2. 02

    Words are the wrong tool for precise edits

    Natural language falls apart on simple spatial references that CAD users handle instantly by clicking. Phrases like "the hole at the end" and "move it up" are ambiguous unless the system already has shared spatial grounding. For many edits, describing the operation in words is actually harder than just performing the operation in CAD.

    Do not assume text is the default human interface for 3D work. Build around pointing, selecting, and constrained operations first, then layer language on top where it genuinely saves time.

      Attribution:
    • alnwlsn #1
    • tapia #1 #2
  3. 03

    The system should refuse under-specified jobs

    An automotive engine mount prompt exposed a deeper product failure than just a bad output. The problem was not that the design was imperfect. It was that the system answered with confidence when the required measurements and context were obviously missing. One commenter said the healthier behavior would be to explain what information is unavailable and what must be measured in the real world before any credible design can begin.

    For high-stakes domains, invest in abstention and requirement elicitation before generation. A graceful refusal can build more trust than a polished wrong answer.

      Attribution:
    • lukasm #1
    • rockostrich #1
    • twosdai #1
    • fkilaiwi #1
    • zachdive #1

In plain english

CAD ↩
Computer-aided design, software used to create precise 2D drawings and 3D models for engineering, manufacturing, and product design.
LLM ↩
Large Language Model, a type of AI system used for text generation and coding assistance.
OpenSCAD ↩
A code-based 3D modeling tool where users define shapes and operations with text rather than drawing directly in a graphical interface.
SCAD ↩
The source file format used by OpenSCAD to describe a 3D model as code.
TPU ↩
Thermoplastic polyurethane, a flexible 3D printing material often used for soft or rubber-like parts.
WebAssembly ↩
A format that lets high-performance code run inside a web browser nearly as fast as native software.

Reference links

Product demos and related tools

  • CADAM demo video
    Official demo showing the product experience discussed in the launch post
  • CADAM live app
    The hosted version readers can try directly
  • Adam commercial product
    Referenced repeatedly as the fuller product with professional CAD integrations
  • Modelrift OpenSCAD LLM benchmark
    Shared as a benchmark and competing approach focused on LLM spatial performance in OpenSCAD
  • Tooltrace
    Example of photo-based tool organization that prompted discussion of image-to-model workflows
  • Quidities
    Another early-stage project mentioned as working on similar ideas

Founder posts and benchmarks

  • Bitter Lesson AI CAD blog post
    Founder essay cited to explain the company's view on scaling models and spatial reasoning
  • MineBench
    Spatial reasoning benchmark suggested as a useful way to evaluate model capability

Examples and social posts