A metadata-driven, deterministic compiler that reads BOM data and produces verified 3D building coordinates — the same way an ERP system explodes a manufacturing BOM into work orders.
Every output element traces to a library input. Nothing is invented. No AI inside. Pure arithmetic.
| 35 buildings compiled (126K elements largest) | 77 verbs, 2,475 products |
| 1M elements loaded in a single federated session | 6 mathematical gates prove every output |
| ERP-native data model (iDempiere) | Blender/Bonsai live GUI |
# Prerequisites: Java 17+, Maven 3.8+, SQLite3
git clone https://github.com/red1oon/BIMCompiler.git && cd bim-compiler
mvn compile -q # Compile all modules
./scripts/run_RosettaStones.sh classify_sh.yaml # Compile Sample House + verify gates
./scripts/run_tests.sh # Full test gate (392 tests)https://red1oon.github.io/BIMCompiler/
50 specs with full-text search, dark mode, and sidebar navigation.
See a walkthrough of how Claude does pair programming with the Creator.
IFC file → extract → classify.yaml → IFCtoBOM → BOM.db → compile → output.db → gates
(once) (human intent) (once) (recipe) (repeat) (elements) (proof)
Every BIM viewer loads the file, then queries it. We query the database. The geometry streams directly from SQLite BLOBs into Blender — no intermediate files, no baking, no server. Two database files, one Python timer.
Step 1: Extract python3 extractIFCtoDB.py --ifc building.ifc --library component_library.db
Step 2: Preview Ctrl+Shift+P → 1M GPU wireframes in ~13s (instant orbit)
Step 3: Stream Ctrl+Shift+A → solid geometry appears, largest pieces first
Three speed secrets:
-
Geometry hashing — 1M elements compress to 50K unique meshes. A hospital has 10,000 doors but only 15 unique door shapes. Mesh creation runs 15 times, not 10,000.
-
Spatial ordering — walls and slabs stream first (
ORDER BY bbox volume DESC). The building looks complete after 5% of elements are placed. -
No format conversion — IFC is tessellated once at extraction. After that, streaming is binary unpack →
from_pydata(). SQLite reads are effectively memcpy from disk cache.
| Metric | Proven |
|---|---|
| Time to first geometry | 2-3s (shell walls appear) |
| RTree wireframe load (1M) | ~13s |
| Orbit / pan | Instant (60 FPS) |
| Geometry deduplication | 95% (1M → 50K meshes) |
| Total data footprint | ~1.1GB (two SQLite files) |
| Server required | None |
See RTree.md for architecture.
Template-driven engine generates schedules (4D), cost estimates (5D), carbon (6D), lifecycle (7D), and safety plans (8D) from the same compiled database. 37 buildings + 1M sandbox costed at MYR 1.59B. See 4D5DAnalysis.md.
bim-compiler/
├── DAGCompiler/ # 12-stage compilation pipeline (G1-G6 gates)
├── BIM_COBOL/ # 64 domain verbs, witness engine
├── BIMEyes/ # Geometric comprehension: 28 shape proofs
├── IFCtoBOM/ # IFC extraction → BOM database pipeline
├── BonsaiBIMDesigner/ # GUI server + validation (TCP :9876)
├── BIMBackOffice/ # ERP reporting + portfolio (HTTP :9877)
├── orm-core/ # Base ORM, BIMLogger, shared utilities
├── library/ # SQLite databases (product catalog, BOMs)
├── migration/ # SQL migration scripts (append-only)
├── scripts/ # Build, test, docs, and audit scripts
└── docs/ # 50 specifications (mkdocs site source)
Redhuan D. Oon (red1) — Kuala Lumpur, Malaysia. Led ADempiere (2006), paved the way for iDempiere (2010), authored Open Source ERP (Pearson Malaysia, 2010). Two decades of ERP manufacturing BOM expertise applied to construction.
| Java | 1,140 files, 261K lines across 12 Maven modules |
| Tests | 129 test classes, 392+ assertions, all GREEN |
| SQL | 106 migration scripts (append-only) |
| Databases | 55 SQLite DBs (4-DB architecture per building) |
| Specifications | 50 docs, governed by SystemContract.md |
| Buildings | 35 compiled (34 extracted + 1 generative) |
| Library | 123,573 tessellated meshes in component_library.db |
| Scale | 1,063,563 elements federated, 35 buildings, 6 disciplines |
Alpha v1.0 — April 2026
Code: GPL v2 · Documentation: CC BY-SA 4.0
Copyright (c) 2026 Redhuan D. Oon. All rights reserved.