Insights
Write-ups from the parts of the job that are hard to find written down honestly — including the projects that went badly and what we changed afterwards.
28 Jul 2026 · 9 min read
Every scene we deliver carries a triangle count, a draw-call count, a texture-memory figure and a p75 frame rate measured on a four-year-old Android. Here is why that list is the contract, and what happens on the projects that refuse one.
14 Jun 2026 · 11 min read
Three.js and a competent developer will get you 90% of the pixels. What does not commoditise is encoding what applications engineering knows about what can actually be built — and that knowledge is usually undocumented and partly contradictory.
2 May 2026 · 8 min read
A 2.8 billion point E57 registration is an archive, not a product. Everything useful happens in classification, decimation and publishing — and that is the work most scanning quotes leave out.
11 Apr 2026 · 7 min read
Splats give photographic fidelity at real-time frame rates and are the best answer available for captured environments. They are also unmeasurable, uneditable and uncollidable — so we composite meshes into them rather than replacing meshes with them.
19 Mar 2026 · 6 min read
The most common reason 3D gets pulled from a storefront is Core Web Vitals. The fix is architectural and boring: the canvas does not exist until someone asks for it.
6 Feb 2026 · 8 min read
Setting one tessellation tolerance for a whole assembly produces either a faceted bezel or a 40 MB download. The fix is per-part budgets driven by visibility, and it should be automated.
15 Jan 2026 · 10 min read
The rendering is a fortnight. Naming 340 assets consistently, mapping them to tags that were named by six people over eleven years, and keeping the mapping alive through plant changes — that is the project.
8 Dec 2025 · 9 min read
Not text-to-3D. The wins are unglamorous: silhouette-aware decimation, material capture from a photograph, UV atlasing, and finding the part that already exists before someone draws a fourth one.
20 Nov 2025 · 7 min read
Compute shaders and lower draw-call overhead are real wins. Universal availability is not here yet, so every scene we ship still has a WebGL2 path — and will for a while.
9 Oct 2025 · 5 min read
The single most common AR defect we are called in to fix is a model authored in centimetres and presented as metres. It is a one-line fix and it destroys the entire value proposition until someone finds it.
If there is a question in this space you cannot find answered honestly anywhere, tell us. Several of these posts exist because a client asked and we could not point them at anything.
We reply to every enquiry within one business day, in your timezone — with a frame budget, not a brochure.