RoomMuse is an AI-powered interior design configurator I’ve been building for the Singapore market, in partnership with Ewins Group. The premise: turn a real room into buildable, priced design options — without requiring the user to know anything about interior design.

The project secured US$20K in Phase 1 development funding from MOZU, a Singapore furniture company, by delivering a working prototype to a real market partner. That commercial validation shaped every technical decision: this had to work for actual customers, not a demo video.

Three Ways to Capture a Space

Not every user will walk around their room with a phone, so RoomMuse supports three space-input modes:

1. AR scan — iOS RoomPlan/ARKit with real-time wall detection
2. Manual entry — type in your dimensions
3. Floor-plan upload — OCR extracts dimensions directly from a floor-plan image

All three converge into the same structured room geometry, which flows through a cross-device spatial pipeline (WebSocket session sync) into design briefs.

The Hard Part: Layouts That Are Actually Buildable

Generating a pretty layout is easy. Generating a layout a contractor can build is not. The core of RoomMuse is a rule-based spatial validation engine covering 5 check dimensions:

• Clearance width
• Door-swing radius
• Traffic flow
• Singapore building code
• Occupancy-based thresholds

Clearance uses 3-tier dynamic thresholds — 900mm / 1200mm / 1500mm depending on user count — so a layout for a family of four is held to different standards than one for a single occupant.

On top of validation, the AI layout engine generates 2–3 layout variations per space, each with a configuration overview and a price estimate, drawing from 50+ curated material finishes and 5 appliance categories.

3D Assets Without a 3D Team

Furniture catalogs are 2D photos; AR previews need 3D models. We built a 2D-to-3D product pipeline using Hyper3D Rodin with trimesh scale normalization, exporting GLB / USDZ — at roughly $0.24 per model, which makes converting a real catalog economically viable.

What I Learned

1. Constraints are the product. The validation engine — not the AI generation — is what makes output trustworthy enough to price and build.
2. Meet users where they are. Supporting three input modes tripled the input-handling work but removed the biggest adoption barrier.
3. Commercial validation changes your engineering. A paying partner forces choices a side project never would — like caring about the per-unit cost of 3D generation.

RoomMuse sits at the intersection I care most about: turning AI capabilities into products with real-world constraints — spatial, regulatory, and economic.