Testing Text-to-CAD on Real Parts: Zoo, Leo AI, and Parametric Reality

We took a standard shop-floor test case and ran it through modern text-to-CAD engines: a 6061-T6 aluminum mounting plate for a NEMA 23 stepper motor mated to a 4:1 planetary gearbox, intended to bolt directly onto 4040 aluminum structural extrusion.

The drawing requirements were ordinary. It needed an 85 mm by 85 mm outer profile, 10 mm plate thickness, a 38.1 mm center pilot bore with a light transition fit, four M5 clearance holes on a 47.14 mm bolt circle, four counterbored M6 holes at the corners on a 60 mm square pattern for extrusion t-nuts, and a 2 mm edge chamfer all around.

In SolidWorks or Inventor, an intermediate mechanical designer models this part in four minutes. A junior machinist can pull up a 2D DXF or raw solid, apply toolpaths in Fusion 360 or Mastercam, and have chips flying on a 3-axis VMC within fifteen minutes.

We fed the exact engineering description into Zoo.dev (KittyCAD / Zoo Text-to-CAD), Leo AI, and Katalyst. We wanted to see what these tools output when given tight mechanical constraints, whether their STEP and B-Rep files carry clean feature trees, and whether importing them into a production pipeline actually saves time or creates a cleanup headache.

The Three Ways AI CAD Engines Attempt Geometry

Text-to-CAD systems on the market right now do not all generate geometry using the same underlying mathematics. They split into three distinct technical approaches, and understanding the difference tells you immediately what kind of geometry will land on your shop floor.

First is direct diffusion and neural mesh generation converted to boundary representation (B-Rep). These systems predict a 3D signed distance field (SDF) or a polygonal mesh, then run a post-processing algorithm to extract planar and cylindrical faces before wrapping them in a STEP container. This approach gives visually convincing thumbnails but frequently outputs non-manifold geometry, warped planar surfaces, and approximated cylinders composed of dozens of micro-faces.

Second is code-based parametric generation, used by Zoo.dev. The neural model does not draw geometry directly. Instead, it writes code in a domain-specific CAD language, like Zoo's KCL (KittyCAD Language), OpenSCAD, or a Python-based CadQuery script. The engine's geometry kernel then executes that code deterministically to build standard B-Rep topology. If the code says cylinder(d=38.1, h=10), the resulting engine output contains a mathematically exact analytical cylinder entity.

Third is conversational feature-tree prompting, seen in systems like Leo AI. These act as engineering copilots, attempting to assemble standard components, parse functional mechanical requirements, inspect dimensional relationships, and integrate with enterprise CAD environments.

Here is how each platform handled the NEMA 23 mounting plate.

Test Run 1: Zoo.dev and the Code-First Pipeline

Zoo approach is the most transparent of the group because it compiles text into KCL before returning geometry. We submitted the following prompt:

"Create an 85mm x 85mm x 10mm aluminum mounting plate. Center bore diameter is 38.1mm. Four M5 clearance holes (5.5mm diameter) arranged on a 47.14mm bolt circle diameter concentric with the center bore. Four M6 clearance holes (6.6mm diameter) with 11mm diameter x 6mm deep counterbores, positioned on a 60mm x 60mm square pattern centered on the plate. 2mm 45-degree chamfer on all exterior edges."

Zoo returned an output in roughly eight seconds. The preview showed a square plate with a center hole and several mounting points. We downloaded the STEP AP242 file and opened it in SolidWorks for inspection.

Geometric Accuracy and Topology

The outer boundary was an exact 85.000 mm by 85.000 mm by 10.000 mm solid. Because KCL compiles directly into an analytical geometric kernel, the planar faces were true planes, perfectly perpendicular, without the surface waviness you see in neural mesh wrappers.

The center pilot bore measured exactly 38.100 mm. SolidWorks recognized the face as a #CYLINDRICAL_SURFACE entity rather than a faceted mesh. This matters immensely for manufacturing. When you drop an analytical cylinder into CAM software, the software instantly identifies the center axis, lets you pick a hole-making cycle, and allows you to apply a standard circular milling or boring routine without manual geometry extraction.

Where Zoo Broke Down

The breakdown occurred in the positional math of the bolt patterns. While the prompt specified a 47.14 mm bolt circle diameter, Zoo's generated KCL script calculated the coordinate offsets for the four holes using an absolute radius from an incorrectly shifted origin. Two of the holes sat at 23.57 mm from center, while the opposite two had their coordinates slightly skewed by an unconstrained sketch offset.

The counterbores for the outer M6 holes were missing entirely. The generated code had created through-holes of 6.6 mm diameter but had omitted the secondary stepped extrusion for the 11 mm counterbore. When we reviewed the raw KCL script, the model had written the code for the outer hole array, attempted a conditional pocketing command that failed engine syntax validation, and silently dropped the feature during final compilation.

Fixing this required jumping into the KCL editor, adjusting the trigonometric variables for the hole positions, and manually typing the secondary pocket extrusions. Total fix time: four minutes. If you already know code-based CAD, that is acceptable. If you were hoping for a drop-in STEP file without code debugging, the tool failed the prompt.

Test Run 2: Leo AI and Katalyst on Assembly Logic

Leo AI and Katalyst position themselves higher up the design chain. Rather than just spitting out a raw single-part solid, they attempt to understand the mechanical context of what you are assembling.

When we gave Leo AI the same NEMA 23 mounting plate prompt along with the constraint that it needed to interface with standard 40-series aluminum structural extrusion, the output handled the mechanical relationships with more context.

Instead of treating the plate as an isolated block of metal, Leo recognized the standard fastener relationships for 4040 extrusion. It adjusted the outer hole centers to exactly 40 mm and 80 mm centerlines, matching the standard t-slot spacing for double-slot profiles, and placed counterbores on the correct side of the plate so socket head cap screws would sit flush.

The Downstream CAD Reality

The friction came when exporting the native geometry. Leo AI focuses heavily on design ideation, requirements capture, and conceptual layout. When you export a STEP file from these types of conversational systems into a traditional CAD package, you receive a "dumb solid" (an unparameterized B-Rep body).

There is no feature tree in your SolidWorks FeatureManager. You do not get an editable sketch containing the 47.14 mm dimension circle. If you need to revise that dimension because your motor supplier changed their gearbox flange to a 50 mm bolt circle, you have two choices. You either use direct editing tools in SolidWorks (Move Face, Delete Face, Re-hole) or you go back to the web portal, re-prompt the model, and download a brand new STEP file.

For any engineer working in an integrated product development cycle, breaking the downstream parametric tree is a non-starter. A change in the mounting plate requires updating mating assembly constraints, regenerating drawings, and recalculating FEA stress models. A dumb STEP file breaks every assembly mate that was attached to its faces.

Inspecting the Geometry: Tolerances, Fits, and CAM Toolpaths

To understand why dumb solids and AI-generated geometry cause headaches on the shop floor, you have to look at how CAM software calculates toolpaths from digital solids.

When a CNC programmer loads a 3D model into Mastercam or Fusion 360, the software inspects the mathematical definitions of the surfaces. A true analytical cylinder is defined by a center axis, a radius, and two bounding planes. When you select that surface for a 2D Bore milling operation or an adaptive clearance roughing pass, the CAM engine uses the exact mathematical radius to compute the tool center point (TCP) offset path with zero interpolation error.

Expected Analytical Geometry (B-Rep):
  Face 1: #CYLINDRICAL_SURFACE (Radius = 19.050 mm, Axis = [0, 0, 1])
  Result: G02/G03 circular arc G-code interpolation (True tolerance +/- 0.005 mm)

Mesh-Derived / Approximated STEP Geometry:
  Face 1..48: #B_SPLINE_SURFACE (NURBS approximation with 48 distinct patches)
  Result: Point-to-point linear G01 moves (Chordal error, faceted bore, chatter)

When we tested geometry generated by purely neural mesh-to-STEP converters, the center 38.1 mm bore was not stored as an analytical cylinder. It was stored as a collection of 48 distinct B-spline surface patches wrapped into a tube.

When we selected the hole in CAM, the software refused to recognize it as a drillable or boreable feature. It treated each patch as an independent freeform surface. Generating a toolpath required switching to a 3D surface contouring strategy, which converted what should have been a four-line G-code program with G02 and G03 circular interpolation into a 4,000-line program of tiny G01 linear segments.

If you run that faceted G-code on a standard vertical machining center, the machine's look-ahead buffer stutters over the micro-segments unless high-speed machining smoothing is cranked up. More critically, the resulting bore will not be round. It will be a polygon with microscopic flats. Your NEMA 23 motor pilot, which requires an ISO H7 or H8 fit (plus 0.025 mm / minus 0.000 mm), will bind on the points or wobble in the clearance.

The Parametric Reality Check

There is a wide gap between generating an object that looks like a bracket and generating a part that integrates into an engineering workflow. That gap comes down to three operational realities.

1. The Cost of Missing Mates

In a real mechanical assembly, a part does not exist in isolation. Our mounting plate must align its corner holes with t-nuts inside the extrusion slots, its center bore with the motor register, and its bolt pattern with the motor faceplate.

When you import an AI-generated STEP file into an assembly containing 40 other components, you must establish mates manually between surfaces. If you later re-generate that STEP file using a revised prompt because the motor frame size changed, SolidWorks or Inventor assigns new internal persistent identification numbers (Face IDs and Edge IDs) to every geometric entity.

Every single mate in your assembly tree referencing that part will explode with broken references. Fixing broken mates in a 200-component robotic joint takes substantially longer than simply sketching the bracket natively in the first place.

2. The Machining Feature Definition Problem

A drawing is a legal document specifying tolerances, surface finishes, and manufacturing methods. A 10 mm plate with counterbored holes has implicit shop-floor rules:

• The bottom of the counterbore must be flat to seat the bolt head evenly.
• The corner fillets must accommodate standard endmill radii (an internal corner radius of 3.175 mm allows a 6 mm or 1/4-inch endmill to clear without slowing down).
• The through-hole must clear standard tap drills if it is meant to be threaded.

When a prompt creates an internal pocket with sharp 90-degree internal vertical corners, a human designer instantly flags it as impossible on a 3-axis mill without EDM or broaching. Modern text-to-CAD engines still routinely place zero-radius internal corners in pockets where a cutting tool must travel, forcing the machinist or mechanical engineer to manually add fillet radii before posting code.

3. Revision Friction

Engineering design is an iterative loop of revisions. You model a part, run FEA, find a stress concentration around a mounting lug, thicken the wall by 1.5 mm, and re-run the solver.

In a parametric CAD environment, you double-click the boss extrusion dimension, change 10 to 11.5, press rebuild, and the entire downstream model, including the drawing sheet, updates automatically.

With current text-to-CAD interfaces, making that 1.5 mm adjustment via a text prompt like "Make the main body plate 1.5mm thicker while keeping all hole depths the same" frequently results in the model rebuilding the entire part from scratch, altering hole coordinates, dropping chamfers, or flipping the orientation of the primary datum plane.

Where Modern Generative Tools Are Actually Delivering Value

Text-to-CAD is not useless. It is simply misunderstood by people who do not cut metal for a living.

Where tools like Zoo.dev and specialized generative systems shine is in creating programmatic boilerplate and handling repetitive mathematical layouts.

Generating a complex gear profile, a custom NEMA motor mounting footprint script, or a multi-port fluid manifold block using code generation tools is remarkably fast. When Zoo generates a clean KCL script, it provides the developer with an editable, version-controlled source file. You can check that script into a Git repository, alter variables parametrically, and compile clean B-Reps automatically within an automated CI/CD pipeline for hardware.

Similarly, rapid 2D outline generators like SimuTecra allow machine shops to quickly turn conversational customer requirements into clean DXF profiles for flat stock cutting on waterjets, lasers, and plasma tables. When the requirement is purely 2D profiling without complex stepped features, the output is immediately usable on the shop floor.

At Mexaio AI, the focus within our mechanical and kinematics pipelines is directly tied to this lesson. We treat generative geometry not as an unconstrained text-to-pixel exercise, but as deterministic, constraint-driven parametric code that respects standard manufacturing tolerances, tool clearances, and native CAD feature trees from the very first token.

The Pre-Flight Checklist for AI-Generated CAD Files

If you are integrating text-to-CAD models or AI-generated STEP files into your machining or additive manufacturing pipeline today, run this five-point check before sending anything to the machine shop or CAM department:

1. Verify Analytical Topology: Open the STEP file in your CAD package and click on all cylindrical bores and planar faces. If your software displays them as multiple B-spline faces or mesh facets instead of a single continuous #CYLINDRICAL_SURFACE or #PLANE, rebuild the feature natively. Do not attempt to run circular milling cycles on faceted surfaces.

2. Check Internal Pocket Radii: Inspect every internal corner. If an internal vertical wall meets another wall at a sharp 90-degree angle, add a radius equal to or slightly larger than your standard endmill radius (e.g., minimum 3.5 mm for a 6 mm tool) to prevent cutter chatter and broken tooling.

3. Measure Bolt Circle PCDs: Never trust an AI model's trigonometric coordinate placement on circular bolt patterns. Project the hole centers onto a 2D sketch and drop a reference dimension for the pitch circle diameter (PCD). Verify that the bolt circle is concentric with the pilot bore within your required tolerance.

4. Inspect Chamfer and Fillet Continuity: AI models often drop chamfers on complex intersections or generate them with variable widths across non-planar edges. Ensure all chamfers have consistent edge offsets to avoid gouging on multi-axis deburring passes.

5. Check Wall Thickness and Hole Breakouts: Measure the minimum distance between outer counterbores and the plate perimeter. Generative models frequently place bolt holes too close to part edges, violating standard 1.5x hole-diameter edge margin rules for sheet and plate stock.

What to Watch Next

The real breakthrough in generative mechanical CAD will not come from web chat interfaces that export dumb STEP files. It will come from native CAD add-ins that write directly to the host application's feature tree via official APIs (SolidWorks API, Autodesk Fusion API, Onshape API).

When an engineering copilot can generate an unsuppressed, fully constrained 2D sketch with explicit geometric relations (coincident, concentric, tangent, symmetric) and extrude it using native host features, text-to-CAD will move from a conceptual novelty to a standard daily engineering tool. Until then, treat every generated solid as raw clay that requires a full parametric inspection before it touches a CNC bed or an active assembly drawing.

Sources

• Zoo Text-to-CAD and KCL Engine: https://zoo.dev/zookeeper
• Leo AI Engineering Copilot: https://www.getleo.ai/blog/open-source-text-to-cad-tools-free
• Katalyst 3D Design Assistant: https://mikekalil.com/blog/ai-driven-text-to-cad/
• SimuTecra 2D CAD Outline Generation: https://simutecra.com/blogs/text-to-cad-ai-product-design