AI generation · 3D · real estate
We ran real estate listing photos through photo-to-3D generation — three mansions, four methods each, from one flat shot to a hand-finished model — to see whether it could go on a production line.
It could not, yet: the input photo decides the result more than the generator does. The model reads light as shape, so shadows and glare turn into false geometry and a tree in front of a facade fuses onto the wall. Every mansion and method is embedded live as a 3D viewer you can spin.
Read the full experiment →
AI agents · web design
Our framework for building creative landing pages: generate many versions in parallel with agents, then pick one — instead of polishing a single design.
It came out of a client project where about twenty-two versions were cut round by round to four finalists and one shipped. Two diagrams lay out the principle and the actual selection cascade.
Read the full experiment →
AI agents · web design
We cloned about sixty strong landing pages in a single day — each a different site with a different signature effect — to build a library of WOW effects we can reuse.
Agents reproduce a genuinely complex effect — canvas animations, particle fields, kinetic type, 3D scenes on scroll — with real fidelity, as long as they have the live site to work from. Five of them are embedded live, each looping its signature motion.
Read the full experiment →
AI models · coding agents
Three models were given the same zero shot prompt to clone one hard site — canvas animations, editorial typography, a long scroll. No shared work between them.
The write up covers what each produced, how each approached the task, and why reverse engineering the animations out of the minified bundle is what decided the outcome. All three clones are embedded live, so you can open them and disagree with the verdict.
Read the full experiment →
See the case studies →
Start a conversation