Don't take our word for it.
See for yourself.

The marketing site and the application are deliberately different things on different origins. clubpolls.io is a static Astro build; the product lives at dash.clubpolls.io as a Next.js App Router app on Cloudflare Workers via OpenNext. The point of the split is that someone reading the pitch never downloads a byte of the application — the demo route alone pulls roughly 176 KB of compressed JavaScript across nine chunks that the marketing homepage never touches.

The officer-console tab switcher and the monthly/annual pricing toggle look like scripted components and aren't. Both are hidden radio inputs driving sibling selectors in CSS. The radios are clipped rather than hidden, so they stay focusable and the tab group inherits native arrow-key navigation — no roving-tabindex code to write, and nothing to go wrong when a script fails to load.

The hero is a hand-written dot field, and most of the code in it exists to stop it costing anything. The animation frame is cancelled outright when the element scrolls out of view or the tab is hidden. Device pixel ratio is capped at 2. Easing is frame-rate independent and the delta clamped, so returning to a backgrounded tab doesn't produce a jump. The glow sprite is pre-rendered once to an offscreen canvas and only drawn for dots above an alpha threshold.

Zero img elements, zero picture elements, no video — the only vector on the page is the wordmark. What looks like a product screenshot is real markup, which is exactly why CSS can tab-switch between three views of it, why it costs no image bytes, and why it can never go stale against the product. Typography is self-hosted and split by unicode-range, so an English reader downloads two font files rather than four.

default-src 'self', with script-src enumerating a sha256 hash per inline script and no 'unsafe-inline' anywhere. The authored page references no external host but its own: no tag manager, no font CDN, no third-party analytics.

A demo route is readable without an account, and it's the one that exposes the maths: a ranked ballot, then the officer's side of it — a Borda count with the weighting stated in plain English, a consensus score, and a per-option polarization readout that tells you when a 'popular' option is actually splitting the room. The tally bars carry screen-reader labels rather than being decorative.

The only build asset on the entire site is one stylesheet. There is no generated script bundle anywhere in the markup. Every interactive behaviour is a hand-written inline module — the homepage carries five, totalling around 2.3 KB, covering the mobile menu, reduced-motion handling, the Instagram fetch, the FAQ fallback and scroll reveal.

The projects page watches four videos at four intersection thresholds, keeps a running map of how visible each one is, then plays only the most-visible video above a quarter on screen and pauses every other. Four videos never decode at once. If a play call is rejected by the browser's autoplay policy, that element switches to native controls rather than sitting there silently broken.

The FAQ uses the details element's name attribute — the modern HTML feature that gives one-open-at-a-time behaviour with no script at all. A 200-byte listener covers browsers that don't support it yet. The first item ships open, so the answer is visible and indexable on load rather than hidden behind a click.

Not a name-email-message box. A division selector, a bid due date as a real date field, project location, the general contractor's name, and a hidden field recording which page produced the lead. Spam control is a honeypot correctly hidden from assistive technology, and the result is announced through a live region rather than only appearing visually.

The projects gallery is server-rendered with nine hand-picked photographs, each with alt text written for the actual image rather than derived from a filename. A script then attempts to replace them from the Instagram endpoint and bails out on any of five conditions — missing template, bad response, wrong shape, empty payload, thrown fetch. The visitor never sees a broken grid or a loading state that doesn't resolve; they see the curated gallery.

Every page's FAQ schema carries exactly as many questions as there are FAQ elements rendered on it, with no drift on any page — the signature of schema generated from the same source as the markup rather than maintained alongside it by hand. The organisation node encodes the Florida contractor licence as a typed identifier rather than loose text.

Four content files hold 278 answered questions across general, personal injury, veterans disability and workers' compensation. Between them the answers carry 340 outbound links to primary legal sources, rendered as labelled chips beneath each answer. A dedicated function normalises how each citation displays — stripping the boilerplate from a Florida Statutes reference while leaving C.F.R. and U.S.C. numbers intact — so a citation reads as a lawyer would write it rather than as a URL.

Content is authored in a bespoke plain-text format rather than a CMS. Each file declares its own item count in a header, and the parser slices the parsed result to that number — so a truncated file or a half-finished paste renders nothing rather than silently publishing a partial library. The declared count is the contract. Multi-paragraph answers survive intact, and every field is escaped before it reaches the page.

A whole-phrase hit inside a question scores fourteen, inside an answer seven; individual terms score four and one. Results sort by score with original order as the tie-break. The 'all areas' view interleaves the four categories round-robin rather than concatenating them, so the first screen shows one question from each rather than sixty-two general ones before anything else appears.

The messaging-consent checkbox ships disabled and is only enabled once the phone field holds a real ten or eleven digit number; the submit button stays disabled until consent is actually given. The number reformats as it's typed. Three anti-spam measures run alongside — a screen-reader-only honeypot, a start timestamp for timing analysis, and a 32-character cryptographic nonce persisted for the session so it survives a page change.

Two routes are held behind a pre-launch gate driven by a data array. Every link pointing at them is marked aria-disabled, given a tooltip, and has both click and keyboard activation intercepted — so a keyboard user hits exactly the same wall as a mouse user rather than being navigated somewhere half-built. The placeholder card grid elsewhere is marked inert so it can't be tabbed into behind its overlay.

Nineteen clients are maintained in two parallel files — short excerpts for the homepage, full-length versions for the testimonials page — so the same quote appears at the right length in each context instead of being truncated mid-sentence. On the page, cards are distributed by greedy least-weight packing, each quote weighted by its length, which gives balanced columns rather than the ragged bottom a naive three-way split produces. The homepage rotator advances each of three slots by the slot count, so two slots can never land on the same quote.