How We Built a Procedural 3D Massing Engine for CRE in 9 KB of JavaScript
We replaced a 470 KB Three.js bundle with a 9 KB lazy-loaded wrapper and a procedural engine that generates a 22-floor tower at runtime — complete with floor-plate efficiency, glazing ratios, and camera views. Here's how.
How We Built a Procedural 3D Massing Engine for CRE in 9 KB of JavaScript
When we set out to build the Keystone Asset Management template — a white-label institutional CRE landing page — the 3D massing study was the differentiator. Not a decorative hero. A functional underwriting tool.
The brief: investors need to see the asset. Floor plates, glazing ratios, clear heights, vertical zoning. And they need it before capital is committed.
The constraint: no external model files. No GLB, no CDN, no 50 MB download. The template had to be a single deployable artifact — npm run build → dist/ → Vercel.
The Architecture
dist/
├── index.html # 23 KB (with JSON-LD, CSP, modal)
├── assets/
│ ├── index-*.js # 9.4 KB ← main bundle (NO Three.js)
│ ├── massing-study-*.js # 7.5 KB ← procedural engine wrapper
│ ├── three-*.js # 470 KB ← Three.js (LAZY, on viewport)
│ └── index-*.css # 18.5 KB
├── models/
│ ├── asset.meta.json # 6 KB ← 22-floor underwriting data
│ └── floor-metadata.schema.json
└── img/ # 240 KB ← portfolio renders Key insight: Three.js (470 KB gzipped → 117 KB) only loads when the hero enters the viewport via IntersectionObserver. First paint is 9.4 KB.
The Procedural Engine
src/three/massing-study.js generates the tower at runtime using seeded RNG. No two builds are different — same seed = same building.
// Floor plate resolution with zone-aware geometry
const widthAt = (floorIndex) => {
if (floorIndex < PODIUM_FLOORS) return LOT_WIDTH;
if (floorIndex >= FLOORS - CROWN_FLOORS) return LOT_WIDTH * 0.62;
return LOT_WIDTH - TOWER_INSET * 2; // typical tower
}; Features:
- 22 floors — Podium (4), Tower (15), Crown (3)
- Office + Residential variants — punched windows + balconies for multifamily
- 4 camera views — Overview, Street, Crown, Plan (smooth spherical interpolation)
- Floor-by-floor underwriting — plate, efficiency, glazing, clear height bound to interactive strip
- Keyboard + touch + mouse — arrow keys, pinch zoom, Home/End, Enter to select
- Reduced motion respected —
prefers-reduced-motionpauses auto-spin
Lazy-Loading Three.js
The 470 KB chunk loads only when needed:
// main.js — IntersectionObserver triggers load
const heroSection = document.querySelector('.hero');
const io = new IntersectionObserver((entries) => {
entries.forEach((entry) => {
if (entry.isIntersecting) {
io.unobserve(entry.target);
initWebGL(); // dynamic import('./three/massing-study.js')
}
});
}, { rootMargin: '200px 0px', threshold: 0.01 });
io.observe(heroSection); Result:
- LCP unaffected by Three.js
- Mobile users who don’t scroll to hero never pay the cost
prefers-reduced-motionusers never load it
Floor Selector + Underwriting Data
The floor strip isn’t decorative — it’s a data browser. Each floor maps to real underwriting metrics:
| Floor | Zone | Plate | Efficiency | Glazing | Clear Height | Use |
|---|---|---|---|---|---|---|
| 1–4 | Podium | 1,180 sq ft | 78% | 64% | 13’6” | Retail frontage |
| 5–19 | Tower | 1,180 sq ft | 82% | 68% | 13’6” | Grade A office |
| 20–22 | Crown | 940 sq ft | 74% | 71% | 12’6” | Penthouse suite |
Clicking a floor updates the specs panel in real time. The data lives in public/models/asset.meta.json — versionable, reviewable, replaceable.
GLTF Upgrade Path (For Buyers)
The procedural engine is the demo. Real clients have Revit/Rhino models.
We built a drop-in replacement with identical API:
// src/three/gltf-loader.js — same { setFocus, setView, destroy }
const { loadGLTFModel } = await import('./three/gltf-loader.js');
study = await loadGLTFModel(canvas, {
gltfUrl: '/models/building.glb',
metadataUrl: '/models/building.meta.json',
onReady: ({ floors }) => { /* same */ },
onFocus,
}); What buyers need to provide:
building.glb— Draco-compressed (< 10 MB)building.meta.json— floor metadata matching GLTF mesh names (Floor_01,Floor_02…)
Camera views auto-compute from model bounds. No manual config.
Verification Suite (25 Automated Checks)
Every deploy passes 25 Playwright checks:
# Run locally
BASE=http://localhost:4173 node scripts/verify.mjs | Category | Checks |
|---|---|
| Runtime | No console errors, WebGL canvas present/sized, hero CTA in viewport |
| 3D | Floor strip (22), selection updates specs, camera chips respond |
| Forms | Rejects invalid, accepts valid, honeypot-ready |
| A11y | WCAG AA contrast (4.75:1 light / 6.95:1 dark), keyboard nav, reduced motion |
| SEO | JSON-LD (Org, RealEstateAgent, InvestmentFund), sitemap, robots, OG/Twitter |
| Performance | No mobile overflow, hero CTA visible, image payload < 600 KB |
| Security | CSP meta, service worker, no emoji/gradient headlines |
The License Inquiry Modal
Not a separate page — a modal triggered from Hero + Footer CTAs:
// Tracks in dataLayer for GTM
dlPush('license_modal_open', { trigger: 'cta' });
dlPush('license_inquiry_submit', { aum, use_case, firm }); Form captures: name, email, firm, AUM range, use case, message. Validates client-side, submits to your endpoint.
Deploy & Sell
npm run build # → dist/ (production-ready)
npx vercel --prod # → live in 30s Package for Gumroad:
dist/zip (production)src/+models/+GLTF_INTEGRATION.mdzip (source)- Price: $397 single / $1,497 agency
Why This Matters
Most CRE landing pages are static marketing. This one carries underwriting data.
When an LP asks “what’s the efficiency on floor 17?”, the answer is in the model — not a PDF buried in a data room.
That’s the product. Not the 3D. The data.
View Live Demo → | License Inquiry → | Technical Docs →
Built by RyderTech — we ship these systems for paying clients.
Tags
Free Tools That Actually Help
Every RyderTech Labs tool is free, instant, and built to solve a specific problem we see Nigerian businesses hit every day. No signup, no credit card.
Explore All Labs Tools