Labs · AI-assisted build

From a floor plan to a walkable interior

I directed an AI coding agent from a floor plan to an editable 3D model and a browser walkthrough. Plain-language feedback shaped the rooms, the materials, the light and the interface over successive passes.

Build time: built from a floor plan in about 6 to 7 hours of directed sessions
Workflow: Floor plan and direction → Blender → glTF → Three.js
What it is: A conceptual study, not a product or a construction plan

Walk through it

Choose a room, compare four finish palettes, switch between daylight and evening lights, and try the furniture options.

Rendered view of the living area, taken from the same scene as the walkthrough

The 3D scene (about 10 MB) loads only when you launch it.

The brief

Turn a plan into a scene that can be explored and corrected, rather than stopping at a single illustration. Keep the model editable, make every room easy to reach, and expose the trade-offs behind each visual choice.

The starting point was a spatial brief. I reviewed each model and asked for changes to arrangement, storage, materials, light and the way the viewer presented them.

Simplified, unlabeled plan showing connected interior spaces — A redrawn, simplified plan for this page.

Direction, correction and another pass

Build an editable scene
The early concept developed into modeled geometry and authored materials, so a change could be made in the model rather than painted onto an image. Room models and storage studies followed, each pass rendered so it could be inspected.
Challenge practical choices
I asked for clearer access, more useful furniture arrangements and quieter finishes. Comparisons exposed space conflicts instead of forcing a fit, and a redundant wall face that left a dark strip in renders was removed.
Light it for the browser
Daylight, room fixtures, soft shadows and ambient occlusion were tuned for real-time rendering. Native renders and browser lighting follow different paths, so each was checked on its own terms.
Refine the walkthrough
Room arrivals, camera framing and navigation had their own revisions. I asked for unobstructed views and simpler controls, with longer notes gathered into one place instead of repeated across the screen.

The useful loop was specific direction, a model I could inspect, and a correction based on what it showed. The agent handled modeling and browser code; I set priorities and challenged the results.

Rendered views

Rendered view of the kitchen cabinetry and work surfaces — Kitchen. Rendered in the browser from the walkthrough scene.

Rendered view of the dining table and the balcony beyond it — Dining. Rendered in the browser from the walkthrough scene.

Rendered view of a bedroom with a fitted wardrobe — Bedroom. Rendered in the browser from the walkthrough scene.

Rendered view of a wall of framed abstract prints — Gallery wall. Rendered in the browser from the walkthrough scene.

How it was checked

Room views, movement limits, load failures, keyboard and touch controls, and reduced motion are checked as software.

The scene uses compressed geometry and newly generated procedural textures. Browser lighting is an approximation of the modeled light, so appearance can differ between devices.

From a floor plan to a walkable interior

Build time

built from a floor plan in about 6 to 7 hours of directed sessions

Workflow

Floor plan and direction → Blender → glTF → Three.js

What it is

A conceptual study, not a product or a construction plan

The brief

The starting point was a spatial brief. I reviewed each model and asked for changes to arrangement, storage, materials, light and the way the viewer presented them.

Direction, correction and another pass

Build an editable scene

The early concept developed into modeled geometry and authored materials, so a change could be made in the model rather than painted onto an image. Room models and storage studies followed, each pass rendered so it could be inspected.

Challenge practical choices

I asked for clearer access, more useful furniture arrangements and quieter finishes. Comparisons exposed space conflicts instead of forcing a fit, and a redundant wall face that left a dark strip in renders was removed.

Light it for the browser

Daylight, room fixtures, soft shadows and ambient occlusion were tuned for real-time rendering. Native renders and browser lighting follow different paths, so each was checked on its own terms.

Refine the walkthrough

Room arrivals, camera framing and navigation had their own revisions. I asked for unobstructed views and simpler controls, with longer notes gathered into one place instead of repeated across the screen.

The useful loop was specific direction, a model I could inspect, and a correction based on what it showed. The agent handled modeling and browser code; I set priorities and challenged the results.