**41 MB → 12.8 MB on a public heavy scene, and the step that silently broke it**
I ran a measured glTF delivery pass and put both versions on a page that loads them live, so it can be checked rather than believed:
Geometry untouched — `renderVertexCount` is 4,497,216 before and after, same bbox, same 63 meshes. In-browser load 6.4 s → 2.3 s. The scene is ABeautifulGame from the Khronos sample assets, picked because it is heavy, public, and ships with an official optimized build to check against.
The part worth having if you do this yourself: my first pass came out at **7.0 MB**, smaller than the official build, and it was wrong. Meshopt stores normals as 8-bit octahedral, and that destroys smooth `KHR_materials_transmission` — the glass spheres in that scene turned into dark faceted blobs while every size metric reported success. Draco keeps normals at f32 and they survive. No file-size report catches this; only rendering both and looking does.
The trade-off I did not win, on the same page for completeness: `join` cut draw calls but pushed upload vertex count from 417k to 568k, against 266k in the reference build. On a mobile target that can go the wrong way, so it is a per-project decision rather than a preset.
Two other things of mine that run in a browser rather than in a video, if the approach is useful to anyone: a [parametric duct configurator](Ventprom — parametric duct-fitting configurator · Artur Smirnov) where the costing reads the same parameters as the geometry, so the preview and the price cannot describe different objects; and a [shelving layout tool](Standes — 3D shelving configurator · Artur Smirnov) with AR on a phone, where the hard part was spatial constraints on mid-range Android rather than arithmetic.
Available for this kind of work — parametric configurators, WebGL/WebGPU delivery and optimization, custom GLSL. Belgrade, UTC+2, remote.
Artur Smirnov — Telegram @smirnovarturr, GitHub smirnov-artur