Research
Engineering notes from building and testing this site: what broke, how we measured it and what fixed it. The short version is on each card. Longer write-ups are still being written.
Web performance
Setting preload to none does nothing next to autoplay: a video told to play must load. Two background loops below the fold took our own homepage from 0.25MB to 3.45MB on first paint, 93% of it video nobody had scrolled to. Gating playback on an IntersectionObserver put it back to 0.00MB until the loop is within 300px of the viewport.
CSS architecture
On this site an h2 carrying the text-sm utility rendered at 56px. Not slightly wrong: the utility had no effect at all. Unlayered CSS outranks anything in a cascade layer regardless of specificity, and Tailwind v4 puts utilities in one. Around 200 headings were affected, which is most of why unrelated pages looked identical.
Integration
Every one reported healthy and none of them worked: API keys read at import instead of at call time, an SDK pinned against an incompatible HTTP client, a model in the fallback chain that was never registered, a database six migrations behind, and an agent called at the wrong endpoint with the wrong payload shape. A health check that does not exercise the real path is a green light with nothing behind it.
Mobile layout
The vh unit is the largest viewport. It ignores the browser address bar, so a vh-sized hero is always taller than the screen it is on. Measured across thirteen of our own pages, heroes ranged from 14% of the screen to 444%. Sizing in svh with a clamp brought every one of them into the 39-57% band on both a phone and a desktop.
Start here
We look at page weight, load order, mobile layout and how your listing shows up in local search, then send you what we find. Free, and a person replies within one business day.
