Aurora — 3D in the browser.

The problem

Every client project runs the same way: agents do the work, I set the direction and check the result. But I always do that in fields I already know — automation, integrations, data. Which makes it easy to lose track of how much of the result comes from the method and how much from my own experience.

So I picked a field I knew nothing about: 3D modelling. I had never opened Blender, never exported a model for the web, never written a line of graphics code. The question was whether this way of working would deliver something that genuinely runs, in a place where I had nothing of my own to fall back on.

The result is live in the browser — the Aurora colony ship, two counter-rotating centrifuges, a planet behind it and asteroids passing by. You can orbit it, zoom in, and fly along the hull.

What I built

The model was built the same way I build systems for clients — in steps that can be reproduced and reversed:

• every change to the model is a script, not a mouse click — 24 versioned scripts, around 10,300 lines, with Blender running headless and no interface open. Every state of the model can be rebuilt from scratch and rolled back to; the chain reached 27 versions,
• verification by picture, not by log — after each change, renders from several camera angles, and then actually looking at them. Geometry bugs are obvious in an image and completely invisible in a log,
• preparing it for the browser — materials baked down into textures, geometry and images compressed, meshes merged: from around 4,825 objects to about 1,180. The result is a ~10 MB file that loads in an ordinary browser,
• the app — Next.js and Three.js on Vercel. Both centrifuge rotations are driven by the app itself, because the export format doesn't carry their animation over from Blender.

The outcome

What lasted isn't the ship. It's the process — from an empty file, through scripted changes and rendered verification, to export and a finished page — written down so that the next 3D project starts near the end of that path rather than the beginning. It's the same discipline I apply to client work: versioned, reproducible steps, and checking the result rather than trusting the claim.

Three days from a field I knew nothing about to a working application — proof that this way of working holds up even where I have no expertise of my own to rescue it.