The problem
KarLev builds accessory dwelling units and custom homes in Los Angeles on an engineered space-frame system: components fabricated off-site by robots to half-millimetre tolerance, assembled on-site in as little as four weeks. Its founder, architect Alexis Rochas, came from a space-frame practice and SCI-Arc, and the product is genuinely different from a stick-built backyard unit. The brand promise is "Full-Size Living."
The difficulty is that a product like that sells on feel and on trust, and those pull in opposite technical directions. The founder wanted an Apple-style cinematic walkthrough, a building assembling itself as you scroll. The business needed to rank for ADU searches across Los Angeles, and a canvas that is the page is a beautiful site that ranks for nothing. On top of that, the sales conversation, models, options, pricing and financing, was happening in person with printed sales sheets.
What KarLev needed
- Cinema and SEO in one page: The scroll experience could not cost the site its indexability or its Core Web Vitals.
- Own everything: Explicit anti-requirements: no Shopify, no Lovable, no hosted CMS, no backend-as-a-service. The code and the content had to belong to KarLev.
- A digital sales tool: Let a prospect configure a unit, see a price, and arrive at a consultation with the conversation already half done.
- Not brick a phone: Two hundred and forty decoded frames is roughly 84 MB of RGBA. Mobile Safari will kill that tab.
How AI is used
An agent-driven build with a written sequence
The site was built as a twelve-step prompt sequence, from scaffold and design system through SEO, shell, scroll-cinema, landing, models, configurator, contact and journal, Core Web Vitals, deploy and Stripe, with each step ending in a committable state. A persistent context file carried the design tokens, the guardrails and a "never do" list into every step. That is how a site of this scope went from first commit to launch in five days.
AI-drafted content and imagery, human-owned
Ten technology pages
Generated hero imagery
Nothing AI-driven runs at request time. The configurator's pricing is deterministic code. Lead handling is a database insert. The site does not contain a chatbot.
How we built it
The canvas is never the page
The homepage renders its content as ordinary semantic sections, server-side. They scroll over a sticky canvas marked hidden to assistive technology, onto which GSAP scrubs a 240-frame sequence in step with Lenis smooth scrolling. Because the canvas is never the largest contentful paint, the cinematic layer costs nothing in the metrics that decide ranking.
On phones the canvas never mounts at all. Both layouts exist in the server HTML, a desktop canvas and six stacked full-height panels, and CSS picks one, so the hidden tree wastes no lazy loads. The choice is gated on a coarse pointer and starts false to stay hydration-safe.
Design Your ADU
The configurator is a server page that decodes a configuration from the URL and hands it to a client component: model, exterior panel, flooring, and add-ons such as solar and battery, deck, premium kitchen, half-bath, smart home and HVAC, each priced live against the model's base. The call to action routes to the contact page with the configuration prefilled. A Stripe reservation checkout and webhook are built and gated off until KarLev is ready to take deposits online.
What shipped
Timeline
- 24 May 2026: Initial launch commit: marketing site and configurator.
- 24 to 28 May 2026: Twenty-six commits: design system, models, quiz modal, features page, deploy fixes. Site live.
- July 2026: Technology content library drafted and shipped; codebase knowledge graph built to navigate the project.
- Ongoing: Admin CRM, Supabase lead pipeline and reservation checkout in the working tree ahead of the next release.
Problems worth talking about
84 MB that must never decode on a phone
A CDN cache that would not forget
Scroll lock without the jump
Leads that anon cannot write
Where it stands
What we can and cannot claim
The site is live at karlev.co. The performance and SEO targets it was built against, Lighthouse SEO of 100, mobile performance of 90 or better, LCP under 2.5 seconds, CLS under 0.1 and INP under 200 ms, are documented as targets in the architecture, and we do not have the client's analytics to publish traffic or lead figures. The counts on this page are from the codebase.
If the canvas is the page, you get a beautiful site that ranks for nothing.
— KarLev architecture note
What This Project Does Well
Technologies Used
Next.js 16 App Router with Turbopack, React 19, TypeScript strict, Tailwind CSS with custom tokens, GSAP and Lenis, MDX via next-mdx-remote, Supabase (Postgres, SSR auth, leads CRM), Stripe Checkout and webhooks, ImageKit CDN, Vercel
Ready to Build Something Similar?
Let's discuss how we can help transform your business with a custom digital solution.
