A PROJECT BY RAY ZHANG

PANTOGRAPH

I.A.D
INTELLIGENCE AIDED DESIGN

WORK IN PROGRESS

(A):

PHYSICALLY ACCURATE GENERATIVE 3D MODELING WITH WORLD FOUNDATION MODELS

(B):

PARAMETER-CONTROLLABLE AI-DRIVEN DESIGN WORKFLOWS FOR CAD USERS

(C):

SCALABLE GENERATIVE POWER WITHIN AND BEYOND EXISTING WORKFLOWS

STATUS
WORKSPACE ↗GITHUB ↗
INFO@PANTOGRAPH.AI
01PREMISE

A PANTOGRAPH COPIES A DRAWING AND SCALES IT. THIS PROJECT ASKS: CAN AN AGENT COPY DESIGN INTENT — AND SCALE IT INTO GEOMETRY?

Pantograph is a study of agentic modeling inside CAD. A designer describes what they want in plain language; an AI agent plans the steps, writes real rhinoscriptsyntax, executes it in the live Rhino document, and checks its own result by looking at the viewport. Not a render, not a mesh import — native, editable CAD geometry, in the designer's file, on the designer's machine.

FILES STAY LOCAL. THE AGENT WORKS OVER A LOOPBACK BRIDGE — NOTHING IS UPLOADED. THE WORKSPACE WILL BE OPEN-SOURCED.

ANCHORYOUR INTENTAI OUTPUT ×2
02METHOD

WORKSPACE

CONVERSATION + CANVAS

AI AGENT

PLANS THE DEFINITION

DEFINITION GRAPH

THE FIRST-CLASS OBJECT

COMPILER

GRAPH → RHINOSCRIPT

RHINO 8

LIVE DOCUMENT

WIRE LOG — MODEL CONTEXT PROTOCOL over stdio · JSON over loopback TCP

→ graph_add_node { "id": "twist", "op": "MathMap", "clause": "3.2° more per level" }

← ok (definition v6)

graph_add_node(id, op, params, provenance)

Adds one typed node to the definition. Provenance is required: the prompt clause this node answers, and why it exists.

graph_connect(from, to, semantics)

Wires an output port into an input port, with a note on what the dependency means. Validation errors come straight back for repair.

graph_execute()

Compiles the definition to rhinoscriptsyntax, performs it in the live document, and returns a viewport capture the agent inspects.

01

DESCRIBE

Design intent stated in plain language — “a lofted skin over 50 floors that twist and taper.” No scripting, no node graph to hand-build.

02

AUTHOR

The agent authors a definition graph through narrow, validated mutations — never baked geometry. The graph is the deliverable.

03

VERIFY + HAND BACK

It performs the definition, inspects the capture, repairs — then the designer keeps editing the same graph: drag, rewire, retune.

03DEMO

RHINO VIEWPORT
AWAITING OUTPUT…
04PARAMETRIC STUDIES
BASE STUDIES — ONE RELATION AT A TIME
(A): SPHERE GRID
A 10×10 grid of spheres with random radii.
AGENT100 spheres placed — radii randomized 1.5–4.5.
(B): TWISTING TOWER
Stack 40 floors, rotating each 2.5° more.
AGENTFloor stack built and twisted per level.
(C): SINE FIELD
15×15 cylinders, heights following a sine wave.
AGENT225 cylinders modulated across the grid.
(D): LOFTED VASE
Loft a vase through 6 scaled circles, then cap it.
AGENTSections lofted and capped into a solid.
(E): ATTRACTOR GRID
12×12 spheres, radius decaying with distance to a point.
AGENTField generated — radii mapped to attractor distance.
(F): RADIAL ARRAY
24 boxes in a ring, scaled up with angle.
AGENTRadial array placed with angular scaling.
COMPOSITE STUDIES — THE SAME VOCABULARY, COMPOUNDED
(G): TAPERED TWIST SKIN
Loft a continuous skin over 50 floor profiles that twist 3.2° per level and taper toward the top.
AGENTOne surface carries both moves — twist and taper compound through the same stack.
(H): PHYLLOTAXIS FIELD
Scatter 350 spheres in a golden-angle phyllotaxis spiral, growing with their index.
AGENTSunflower spiral placed — 137.5° divergence, radii mapped to index.
(I): HELIX STAIR
A spiral staircase: 36 treads winding up around a central core.
AGENTTreads swept up the helix; core column holds the center.
(J): ATTRACTOR FACADE
160 facade panels that rotate open as they approach an attractor point.
AGENTPanel field generated — rotation driven by attractor distance.
(K): WAVE LOFT TOWER
A tower lofted through 26 rings whose radii follow a sine wave.
AGENTProfile modulated and lofted into one undulating surface.
(L): DOUBLE HELIX
Two helices of spheres, phase-offset 180°, like a double helix.
AGENTBoth strands wound — same parameters, opposite phase.
05RESEARCH

THE EDITABLE RETURN

PANTOGRAPH AND THE CASE FOR AN ORCHESTRATIVE DESIGN MACHINE

Why an AI design system should return a definition rather than an object — the argument read through cybernetics, notation theory, media theory, and the documented brittleness of parametric models, with this prototype as the evidence that the position can be built.