Processing Pipeline
How requirements become a structured building object graph, costs, BOQ, and review findings — and how this connects to a real CAD worker in production.
Active: working in this demo. Demo: working at conceptual fidelity. Production path: implemented as a server-side worker outside the browser. Future: planned.
1. Input
What enters the system.
File intake
DWG, DXF, PDF, images, manual requirements, company rules, supplier rates.
Demo accepts manual project inputs and stores uploads. Real DWG/DXF/PDF parsing requires the production CAD worker.
2. Parse & Understand
Convert files into normalized geometry and semantic meaning.
DXF / DWG conversion path
ODA File Converter or Autodesk APS to normalize DWG → DXF for parsing.
Runs as a Python worker outside the browser. Out of scope for this in-browser demo.
Geometry parser
ezdxf / IfcOpenShell extract walls, doors, windows, rooms, and slabs.
Produces a normalized geometry stream the semantic classifier can read.
Semantic classifier
Tag entities as walls / doors / windows / circulation / MEP / parking using rules + ML.
AI vision and rule-based heuristics map raw geometry to architectural meaning.
Room / zone detection
Wall-loop detection groups segments into rooms and circulation areas.
Demo derives rectangular zone loops from the generated layout JSON.
3. Reason & Optimize
Build the structured graph, apply constraints, and use AI for explanation.
Building object graph
Structured nodes (zones, walls, doors, windows, materials, costs) and typed edges.
Lives in frontend state. Each generated project carries a versioned graph used by 2D, 3D, costs, BOQ, and review.
Constraints & scoring
Adjacency, circulation, egress, MEP access, envelope rules.
Rule-based today. Layouts that violate constraints surface as issues in CAD Review.
Layout optimizer
Deterministic placement of zones based on project type and selected zone set.
Future: solver-based or AI-assisted iterative optimization scored against constraints.
LLM reasoning & explanations
Claude / Gemini explain the layout, audit findings, BOQ notes, and document drafts.
AI is advisory only. Code calculates geometry, quantities, and costs — AI never invents numbers.
4. Cost & Output
Materials, quantities, costs, BOQ, and CAD export.
Cost & BOQ engine
Materials → cost items → BOQ sections, all linked back to graph nodes.
Click a 3D object to see its material, unit rate, and BOQ line. Click a BOQ line to highlight source objects.
CAD geometry generation
FreeCAD / IfcOpenShell worker assembles real CAD geometry from the object graph.
Replaces the conceptual browser massing with engineering-grade CAD entities.
Export
JSON, CSV today; DXF, IFC, PDF, OBJ via the production CAD worker.
Demo exports layout JSON and cost CSV. Engineering exports require the worker.
Active in this demo
- · Layout JSON generation from requirements
- · Browser-based 2D plan rendering
- · Browser-based 3D conceptual massing
- · Building object graph (zones, walls, doors, windows, materials, costs)
- · Click-through metadata on 3D objects
- · Rule-based CAD review findings
- · Cost sheet and BOQ derived from graph
- · AI explanations via OpenRouter (Claude / Gemini)
Production path
- · DWG → DXF conversion (ODA / Autodesk APS)
- · DXF / IFC parser worker (ezdxf, IfcOpenShell)
- · Semantic classifier for parsed geometry
- · FreeCAD worker for CAD geometry generation
- · Real DXF / IFC / PDF / OBJ export
- · Vision parsing for uploaded PDF / image plans
- · Embedding-based semantic search across rules and projects
AI outputs are advisory and must be reviewed by qualified engineers, QS teams, and project stakeholders. This system does not provide engineering or structural approval.