From Wix to an HTML concept, and from there to a living, bilingual, editable platform with its own 3D engine. A year of learning inside Conqr, turned into product.
At Conqr, the agency where I work, I took its website from a sealed HTML concept to WordPress + Elementor Pro + Polylang, with a custom 3D engine, in 9 days. I launched it on conqr.mx on October 30, 2026: bilingual, editable by the team, and from 3.99 to 766 requests per second on the Home.
The site lived on Wix. By the time the domain moved, it no longer answered, and its URL map had to be rebuilt from the Wayback Machine.
The Wayback Machine kept the text and layout, not Wix's images or videos: that is why they show as empty boxes.
A beautiful, sealed prototype: one logic file, a 57 KB stylesheet and its media packed as base64. It looked like the site Conqr wanted to be, but nobody on the team could change a comma.
Its real weight could not be measured: its base64 data files never arrived. The media that did arrive add up to ~1 MB of base64, before counting the video.
WordPress + Elementor Pro + Polylang, with a custom 3D engine: bilingual, click-to-edit and live on conqr.mx.
151 URLs in two languages, all returning 200 and indexable.
Intro → hero → manifesto. On the left, the HTML concept (in the reconstruction used for QA, without the chrome video that never arrived); on the right, conqr.mx.
The project's first rule fits in one line: when in doubt, the original rules. This was not about "looking similar": every section was measured against the original, served locally, at 1440×900 and 390×844, in English and Spanish.
Measuring this closely also exposed bugs in the original itself: a footer that overflowed by 7 px on Spanish mobile, and a pin that grew on mobile because the page spilled sideways. Those bugs were not copied. The original rules, but it gets measured.
Original on the left, WordPress on the right. Tap one to see it in full.
The 40 frames of the original chrome video next to the live 3D figure, with each frame's silhouette IoU (mean 0.96).
Home intro in deterministic frames at 0.5, 2, 3.5, 4.3 and 5 s: 0 differing pixels.
Menu opening at 0.2, 0.5 and 1.1 s and the open menu. Geometry ∆ 0.00 px; the chrome preview is the live figure.The liquid drop of "we are a digital, creative & btl agency." at three points of the pin, at 1440 and 390 px.
Full About page at 1440: hero, con·quer, manifesto, three reasons why, AI, map, clients and CTA.
Nosotros (About, Spanish) on a 390 px phone, section by section.
Expertise: hero, service list with the light row on hover, CTA and footer.
Formula 1 case: hero, "the challenge.", facts, services and "Next project".The breathing, shape-shifting chrome figure used to be a video. Now it is a living object: it bulges toward the cursor, bounces when touched, spins when dragged, jumps to its next shape on double click and follows your finger on an iPhone. It keeps the original video's cycle, every 10 seconds: sphere, torus, triangular torus, cross, clover, wavy sphere and back to the sphere.
A single WebGLRenderer for the whole page. Each figure is drawn into its own viewport and copied into its own 2D canvas: one WebGL context no matter how many figures there are, and each keeps its place in the DOM.
The shape lives in the vertex shader, with two-octave simplex noise that makes the surface "liquid".
Physical metal lit by a blue-panel environment generated in code, with no HDR download.
Bulging toward the cursor, bouncing on click and spinning on drag, with damped physics.
The cycle was measured on 40 real frames of the video: silhouette IoU 40/40 with the QA goal at ≥ 0.90, mean 0.960 and minimum 0.932 (before tuning: 29/40, mean 0.919).
JavaScript per frame: 1.75 ms on desktop, 8.62 ms with a 4× slower CPU and 3.77 ms on an emulated phone with a 4× CPU. All within the 16.7 ms of a 60 fps frame. The test server's fps do not count: it has no GPU and draws WebGL in software.
A single requestAnimationFrame that stops off-screen or in a hidden tab, automatic low quality on modest devices and, without WebGL, a still image rendered by the engine itself.
A beautiful site nobody can touch ages in weeks. The second rule was: whatever can live in Elementor lives in Elementor. Every piece of the original was sorted into three types:
Plain Elementor.
Regular containers with a cq-* class that adds motion. For example, cq-hscroll turns a row of containers into the pinned horizontal carousel.
Only what Elementor cannot do, always with its controls in the panel.
136 sections edit with a direct click and 26 from the WordPress editor or automatically. The team got a 6-page guide with real screenshots.
English is the base, at /, and Spanish lives at /es/, with its own slugs: /about/ ↔ /es/nosotros/ and /work/ ↔ /es/portafolio/. Free Polylang cannot share a slug across languages; the custom plugin handles it with filters and rewrite rules, and sends a 301 when a URL is opened under the wrong language. The EN/ES button navigates to the translation and returns to the same scroll position, without replaying the intro.
PHP-FPM, a 256 MB OPcache, a private Redis and a page cache where Apache serves the HTML without touching PHP. The Home went from 3.99 to 766 requests per second with 10 users.
Mail goes out signed with DKIM, with SPF, DKIM and DMARC passing. It is only sent from a domain that passes all three.
A unique meta description per URL, Open Graph, structured data, hreflang and a Yoast sitemap that kept every previous title.
Forms with Jotform's official embed on Contact, careers and the internal portal.
4 internal noindex pages, outside the sitemap and the cache.
341 redirects from the old Wix URLs, with the inventory rebuilt from the Wayback Machine.
9 calendar days and 6 working days. There are no log entries on the 24th and 25th; the 27th ran into the early hours of the 28th.
Oct 22
Inventory of every section of the original and the NATIVE / NATIVE + BEHAVIOR / WIDGET classification.
Oct 22
Child theme, Polylang, custom plugin, content types and content seeded in EN and ES.
Oct 22 → 27
The live figure replaces the video; on the 26th it was measured against 40 frames (IoU 40/40).
Oct 22 → 23
∆ 0.00 px on header and menu; 0 differing pixels in the intro.
Oct 26 → 27
Hero, manifesto, con·quer, Expertise, footer, page curtain and chrome cursor.
Oct 27
About, Expertise, Work, Careers, Blog, Contact and 404, from 01:00 to 01:50.
Oct 27 → 28
162 sections reviewed, none left uneditable, and the team guide.
Oct 27 → 30
Final URLs per language with 301s and the new SEO & GEO service (8 in total).
Oct 29
Three photos turned into layered scenes.
Oct 27 → 29
20 linked EN↔ES pairs and 152 old URLs with 301s.
Oct 29
4 internal noindex pages, outside the cache.
Oct 26 → 30
PHP-FPM, OPcache, page cache and a private Redis: from 3.99 to 766 req/s.
Oct 27 → 30
hreflang, unique meta descriptions, OG, JSON-LD and Yoast with 151 URLs.
Oct 30, 00:24–00:40
16 minutes, a 503 gate from 00:30 to 00:39 and 341 301 rules from Wix.
Oct 30
Smooth scroll, page curtain, safe SVG, adaptive favicon, social links and Yoast.
The concept showed its chrome figure as a 10 s video embedded as base64, in the hero, inside the con·quer letters, in the menu, in the cursor and in inner heroes. And the rule was never to load the original's base64 files.
The concept was built as a self-contained prototype: the video was the fastest way to show the figure, but it could not be edited, weighed down the HTML and did not react to the user.
The video was replaced by a live 3D figure (Three.js r160, local). The cycle was measured on 40 real frames, one every 0.25 s: from 29/40 frames with IoU ≥ 0.90 (mean 0.919) to 40/40 (mean 0.960, minimum 0.932).
The original rules, but it gets measured: every likeness is validated with numbers against the reference, and nothing loads as base64.
Nuance. The QA goal was IoU ≥ 0.90. That all 40 frames also clear 0.93 follows from the minimum, but 0.93 was not the stated goal.
The figure shows up at once in the cursor, the menu, the hero, floating elements, thumbnails and About scenes, and each appearance needs its own place in the DOM: above the menu, behind the letters with multiply, with masks and floating.
One WebGL canvas per view multiplies contexts, and browsers limit how many active contexts a page can have. A fixed full-screen canvas forces everything above or below the content.
One WebGLRenderer off the DOM: each view is drawn into its own viewport and copied with drawImage into its 2D canvas. A single requestAnimationFrame, paused off-screen or in a hidden tab, and a still image when there is no WebGL.
Widgets consume the engine instead of duplicating it: one API and one context.
Nuance. The log justifies the decision with the single context and layer order; it does not cite a specific context limit.
The team had to change text, images, colors and order without touching code, and the site had to keep scroll pins, GSAP timelines, a magnetic cursor, liquid effects and the 3D figure.
Elementor does not natively do canvas, WebGL, pinned scroll or liquid effects, and turning every section into a closed widget kills click-to-edit.
Sort every piece into NATIVE, NATIVE + BEHAVIOR or WIDGET: 23 widgets and 14 behaviors. 162 sections reviewed, 136 with a direct click, 26 from the WordPress editor or automatic, and 0 left uneditable.
Whatever can live in Elementor lives in Elementor. What cannot is a widget with controls, never loose code.
On October 26, when page caching was switched on, the staging site returned a 500 error for about 5 minutes.
With PHP-FPM, PHP runs as the account user. WP-CLI ran as root and the plugin rewrote the config with root as owner; PHP-FPM could no longer read it.
The correct owner and permissions were restored. It happened on staging, not in production.
WP-CLI always as the account user, never as root. It is the first rule in the production site's document.
The move to conqr.mx had to happen without anyone seeing a half-built site, with the domain's DNS already pointing at the new server and the Wix site no longer answering.
There was no old site to keep running in parallel: the domain was already on the new server.
Full backup, copy and URL rewrite, caching rebuilt, and a 503 "coming soon" gate with Retry-After and noindex from 00:30 to 00:39. It opened with the client's OK, and by 00:54 the 341 301 rules from Wix were live.
Publish behind a gate and verify from outside before opening. Every old URL has a destination.
Nuance. "Zero downtime" holds as "no window of a broken site": for 9 minutes visitors saw a 503 "coming soon", which is the right thing for search engines.
At launch, conqr.mx's DNS ended up with no MX or SPF records, and the Google Workspace inboxes stopped receiving.
Moving the DNS to the new server dropped the mail records, and the server had the old DNS cached, which hid the problem.
It was flagged on the spot and the client restored MX, SPF, DKIM and DMARC around 02:00. A delivery test at 01:59 completed.
Check MX against a public DNS, never the local resolver, and only send from a domain with SPF, DKIM and DMARC passing.
Nuance. No email landed in spam because of DMARC: the risk and the rule that prevents it were documented. The actual "spam" was the relay rejecting 3 reports for their content; the text was rewritten and it went through.
In the Expertise carousel, the last panel stayed off-position, the background light went dark on the way back, and the exit to the footer felt like a bump.
The pin ended exactly where the track ended, the footer script stored the opacity "from page load", and the wheel moved the scroll in jumps.
A 30% resting stretch at the end of the pin, the light keeps the opacity it has when reaching the footer, and Lenis smooth scroll synced with ScrollTrigger. No Lenis with reduced motion or in the editor.
No scroll-linked animation stores "page load" state.
The SVG logo could not be uploaded to the Media Library and, later, the black logo nearly vanished on dark tabs.
WordPress does not allow SVG by default, and a single-color icon cannot look right on both light and dark tabs.
SVG with sanitizing on upload, for administrators only. The final favicon is adaptive: black on light tabs, white on dark ones.
SVG only with sanitizing and only for administrators; the site icon is printed from a single place in the code.
Nuance. The test WebKit does not switch the SVG to white in dark mode (Chromium does). It still needs checking in real Safari.
Phases, rules and what the client decides, before touching the server.
At the end of each one: stop, report, wait.
On the server, with WP-CLI as the account user, scripts that can be re-run and nothing in WordPress or Elementor core.
135 backup points across 6 working days.
Chrome, Safari and mobile, in English and Spanish. Measurements, not impressions.
122 work reports in 9 days: what was done, what changed, the backup and what is pending.
Rules, decisions and exact commands, so each session starts where the last one ended.
AI speeds things up; I direct, decide and verify. The AI wrote a lot of code. The judgment, the rules and the decisions about what not to do were human.
Requests per second on the Home with 10 users
Home p95 response time
CLS (layout shift) versus the concept
Home SEO score in Lighthouse
Lighthouse in production, median of 3 runs: SEO 100 and CLS 0 on all five pages measured. Mobile performance lands between 54 and 66 because the test server has no GPU and draws the 3D figure on the CPU; the concept scores higher there because it draws no figure at all. The concept's numbers come from the reconstruction used for QA.
Visual design: Conqr design team.
Development, 3D engine, WordPress architecture, infrastructure, migration and launch: Samuel Torres, with Claude and Claude Code as AI copilots.
Want something like this on your site? Book 20 minutes and we'll review it together.
Book 20 min
How to integrate a 3D graphics engine into your website (without killing performance)What I learned putting a live 3D figure on a real website: architecture, performance and accessibility.Read the article →
Why you should renew your website now (and how to know it's time)Eight signs your site has aged, and a 10-question self-check.Read the article →