User Input
Parametric 3D Output
Evaluation

ParaGraph — AI Parametric Design Studio

Inspiration

As computational designers, we’ve spent years wiring parametric graphs manually — tweaking sliders, debugging nodes, rebuilding similar models from scratch.

A phone stand takes 2 minutes to describe, but 15–60 minutes to model.

Existing AI 3D tools generate static meshes:

Not editable
Not explainable
Not constraint-driven

We asked:

What if AI didn’t generate geometry — but generated the parametric system itself?

Not AI for 3D.
AI that does 3D like an engineer:

Constraint-aware
Versioned
Objective-scored
Iterative

ParaGraph builds, evaluates, and refines editable engineering-grade CAD, not disposable meshes.

What it does

1. Text → Parametric 3D Model

Input:

Spur gear, 20 teeth, module 2, pitch diameter 40mm

Output:

Editable parametric dependency graph
Involute tooth profile
Center bore
Chamfers
Engineering-grade BREP solid

Four AI agents collaborate in real time.

2. Image → Parametric Model

Drop a photo of any object.

Pipeline:

Image preprocessing
Structured geometry extraction (Design Intent Representation)
Deterministic prompt conversion
Parametric graph generation

Image → Structured intent → Editable parametric system.

3. Natural Language Editing

Click any node and type:

“Make it 23.65% bigger”
“Double the number of teeth”
“Set diameter to 50”
“Make it bigger” (defaults to +20%)

The system:

Parses intent
Regenerates code
Recompiles
Updates the viewport

4. BREP-Aware Evaluation

Unlike competitors who evaluate meshes visually, ParaGraph evaluates actual CAD topology via OpenCascade.

Metrics:

Volume
Surface area
Face/edge topology
Face type classification
Compactness ratios
Symmetry

Two-stage scoring:

Quality: Is the solid valid and well-formed?
SpecMatch: Does it match the user’s intent?

Broken geometry can never score above 40%.

5. Versioning & Export

Full version history
Restore any previous design
Export STL
Import into SolidWorks / Fusion / FreeCAD
Download full Markdown audit report (agents, timing, cost, scores)

How we built it

4-Agent Pipeline

User Input ↓ Agent 1 — Intent Parser (structured JSON via constrained decoding) ↓ Agent 2 — Parametric Graph Builder ↓ Agent 3 — Build123d Code Generator ↓ Build123d Compiler (OpenCascade BREP) ↓ Agent 4 — Deterministic Geometry Evaluation

Agent Responsibilities

Intent Agent: Guaranteed-valid JSON output
Tree Logic Agent: Builds parametric dependency graph
Script Agent: Generates Build123d Python (loops, trig, advanced geometry)
Compiler: Produces STL + geometry metrics
Evaluation Agent: Pure mathematical scoring (zero LLM cost)

Cost: ~\$0.006 per full design generation.

Challenges we ran into

OpenSCAD → Build123d migration
Needed true parametric graphs with BREP kernel support.
3D viewport race condition
STL data arrived before scene initialization. Fixed with state gating.
JavaScript scoping crash
Closure variables unavailable during compilation. Refactored function parameters.
Template literal conflicts
Python backticks broke JavaScript prompts. Reworked string handling.
Model discovery errors
Incorrect NVIDIA model IDs required cross-referencing documentation.
Fillet radius crashes
Built auto-healing compiler that retries without invalid fillets.
Structured output instability
Solved via constrained decoding for 100% schema-valid JSON.

Accomplishments that we're proud of

The only AI 3D tool that scores actual BREP geometry
Image → Structured Intent → Parametric System (not image-to-mesh)
Natural-language parametric editing
Auto-healing CAD compiler
Observable multi-agent collaboration (live tokens, cost, timing)
Engineering-grade evaluation at sub-cent cost

What we learned

Parametric systems > static meshes
Constrained decoding eliminates parsing failures
BREP kernels expose rich geometric metadata instantly
Multi-layer fallback chains make demos reliable
Separating perception (VLM) from generation (parametric graph) is the right abstraction