Case study · AI pipeline

WebPulse: AI Website Redesigns

A pipeline that finds local businesses with outdated websites and rebuilds them as live, personalized preview pages, at roughly $8 per thousand leads.

Cinematic 3D render of a cracked stone slab transforming into luminous glass, lit in purple, fuchsia, and orange
RolePipeline architect + dev
ClientWeb design agency
StackNode.js + AI vision
StatusIn production use

Cold outreach that shows, not tells

A web design agency's best pitch is a finished redesign: send a prospect a link to their own site rebuilt, and the email sells itself. But doing that by hand means finding businesses with bad websites, judging each one, and mocking up a redesign. Hours per prospect, when cold email needs thousands of prospects.

The agency needed the whole chain automated: find the leads, judge the sites, build the previews, put them online, and deliver everything in a spreadsheet their team could work from.

Spend AI only where it earns its keep

The kit runs as twelve composable skills with one autopilot command that chains a full city run end to end. The client talks to it in plain English: "scrape pressure washers in Louisville" is a valid instruction.

The core design decision is cost discipline. AI handles only the judgment calls: vision models qualify sites and score photos. Everything mechanical is deterministic code that costs nothing: brand extraction, template filling, palette repainting, deployment. That's the difference between $8 and $40 per thousand leads.

Six stages, one command

  1. 01

    Scrape

    Apify pulls around 140 businesses per city from Google Maps, with rule-based pre-flags catching franchises, big brands, and businesses with no website at all before any money is spent judging them.

    Apify · Google Maps · ~$0.50 per city
  2. 02

    Qualify

    Playwright screenshots every site, then DOM signals plus a vision model sort leads into HOT, GOOD, BAD, and reference buckets. The same call extracts each site's brand colors and mood for later reuse, at about a tenth of a cent per lead.

    Playwright · GPT-4.1-mini vision
  3. 03

    Extract assets

    A free deterministic crawl collects each business's logo (with a styled text wordmark generated when none survives), hero candidates, and up to 200 real photos per lead from their website, Facebook, and Google Maps listing.

    Cheerio crawl · $0
  4. 04

    Build the preview

    A polished niche template is filled per lead. An AI scorer picks the best real photos into the hero, about, and gallery slots (finished work first, crew shots last resort), and the whole template is repainted with the lead's actual brand palette. Their business, their photos, their colors.

    Template fill · AI photo scoring
  5. 05

    Deploy

    One Vercel project per niche, with every lead at its own route. New leads are file adds plus a single redeploy, so the system scales to thousands of previews without spawning a project per lead.

    Vercel · 1 project per niche
  6. 06

    Deliver

    Everything lands in the client's Google Sheet: qualified leads in tabbed buckets, live preview URLs, and optionally an AI-drafted outreach email per lead, ready to paste into their CRM and send.

    Google Sheets · outreach drafts

The dashboard the agency runs it from

WebPulse Studio puts the pipeline behind a login. A business owner requests a free preview through a short form, and the agency reviews each AI-built draft, fixes its copy, photos and logo, and redeploys from one screen.

Use the arrows or swipe to move between screens, and tap or click any screen to see it full screen.

WebPulse Labs intake form asking for business name, website or listing link, cell phone and email
Intake form. A business owner asks for a free AI website with four fields: business name, a website or listing link, cell phone and email.
Requests list with each business, its link, submission time and status, phone numbers blurred
Requests. Every request with its link, time and status. With no Google Maps listing, the draft is built from the website and the row says so. Phone numbers blurred.
Website drafts dashboard listing businesses with city, status, Google rating, photo count and deployed link
Website drafts. Every AI-built draft in one list, with its Google rating, the photos scraped for it and its live link once deployed.
Draft editor with a live website preview on the left and the business details form on the right, reviewer name blurred
Draft editor. The live preview beside the business details, with a check that the testimonial is a real review. Reviewer name blurred.
Photos tab of the draft editor showing the hero pick, image slots marked stock or AI pick, and photo filters
Photos. Each image slot shows the AI's pick and marks stock photos, and any slot can be swapped from the photos scraped for that business.
Logo tab of the draft editor showing the scraped logo on a dark header and on white
Logo. The scraped logo previewed on the dark header and on white, with an upload for a cleaned-up version.
Template demos for cleaning, pressure washing, roofing and tree service pages with placeholder text
Template demos. One live demo page per service, with placeholders where a lead's name, city and phone drop in.

Agency-scale prospecting on autopilot

~$8 Per 1,000 leads

Full pipeline baseline, against a client quality budget of $30 per thousand.

12 Composable skills

Each stage runs standalone or chained by the one-command autopilot.

100% Real business content

Every preview uses the lead's own name, photos, reviews, and brand colors. Nothing invented.

Client KitNode.jsPlaywrightApifyGPT-4.1-miniVercelGoogle Sheets

Want to see the output quality? The Web Projects page has a drag-to-compare slider showing a real business's old site against its WebPulse rebuild.

Available for remote work

Have a process an AI pipeline could run?

If your team does something repetitive at volume, there's probably a pipeline in it. Tell me about it and I'll sketch the stages.