BuildBrain
AI agent that reads architecture drawings. Extracts structured data from IFC models and PDF construction documents, cross-validates between sources, and flags discrepancies.
Core loop: Drop IFC + PDF files -> agent scans both sources -> extracts elements, properties, schedules -> cross-validates -> generates QTO tables and discrepancy reports
Mental model: Claude Code is the brain, CLI tools are the hands. The agent queries IFC models via IfcOpenShell and PDFs via pdfplumber, reasons about what it finds, and produces construction deliverables. Packaged as a Claude Code plugin with 6 skills.
Architecture
Claude Code (orchestrator — reasons about construction data)
│
├── /scan First run — inventory model + find PDF schedules
├── /ifc-extract Query IFC elements, properties, quantities
├── /pdf-extract Extract text, tables, schedules from PDFs
├── /cross-validate Compare IFC vs PDF, flag discrepancies
├── /report Generate QTO tables, schedules, compliance checks
└── /cli-tools Reference manual for the CLI tools below
│
├── ifc_extract.py IfcOpenShell wrapper — 7 subcommands
└── pdf_extract.py pdfplumber wrapper — 4 subcommands
What It Does
| Capability | How |
|---|
| IFC model queries | Extract doors, walls, windows, slabs, roofs, columns, beams, stairs — any element type. Properties, quantities, materials, grouped by storey |
| PDF schedule extraction | Auto-detect door/window/finish schedules from column headers. Search, extract tables, text |
| Cross-validation | Compare element counts, fire ratings, materials between IFC and PDF. Distinguish MISMATCH (conflicting data) from ABSENT (missing IFC data) |
| Quantity takeoffs | Pre-computed aggregates — counts, areas, volumes. Hybrid approach: reads IFC quantity sets, falls back to geometry computation |
| Data quality | Flags missing property sets, misclassified elements (IfcBuildingElementProxy), orphaned elements. Reports completeness metrics |
| Structured reports | QTO tables, element schedules, material takeoffs, compliance checklists. Australian conventions, NCC terminology |
Tech Stack
IFC: IfcOpenShell 0.8.x (parse, extract, compute quantities from geometry)
PDF: pdfplumber (text, tables, schedule auto-detection)
Orchestration: Claude Code plugin (skills + CLI tools)
Data: pandas, CSV, JSON, Markdown
Prerequisites
pip install ifcopenshell pdfplumber pandas
Usage
Install as a Claude Code plugin, place IFC and PDF files in data/, then:
/scan Inventory the model and find PDF schedules
/ifc-extract door Extract all doors with properties
/pdf-extract Find and extract door schedule from PDF
/cross-validate door Compare IFC doors vs PDF door schedule
/report qto doors Generate quantity takeoff table
Or just ask naturally — "how many fire-rated doors are there?", "what's the total wall area on Level 2?", "does the model match the spec?"
CLI Tools
The plugin bundles two standalone Python CLI tools at skills/cli-tools/scripts/.
ifc_extract.py — 7 subcommands:
python ifc_extract.py <file.ifc> summary # scan model, all element types + counts
python ifc_extract.py <file.ifc> list door # list elements with properties
python ifc_extract.py <file.ifc> props <guid> # detailed properties for one element
python ifc_extract.py <file.ifc> query door \
--property Pset_DoorCommon.FireRating --not-null # filter by property
python ifc_extract.py <file.ifc> quantities wall \
--group-by storey # pre-computed aggregates
python ifc_extract.py <file.ifc> validate # data quality check
python ifc_extract.py <file.ifc> export door \
--output doors.csv # CSV export
pdf_extract.py — 4 subcommands:
python pdf_extract.py <file.pdf> schedules # auto-detect schedule tables
python pdf_extract.py <file.pdf> search "fire rating" # find pages by keyword
python pdf_extract.py <file.pdf> tables --pages 12-15 # extract tables from pages
python pdf_extract.py <file.pdf> text --pages 1-5 # extract raw text
All commands output structured JSON. Use --save to persist results to output/.
Project Structure
