Bart Lind

Website-to-demo delivery pipeline

A productised pipeline that turns a business's existing website into a deployed demo — a rebuilt site plus a chatbot that answers only from that business's real content — used as a sales asset. I built it to run for education-sector businesses; the two live instances below are real prospect demos, not paying clients.

1. The seven stages

Every stage has a human gate. Nothing moves to the next stage until I (or, in principle, the prospect) sign off on the previous one.

  1. Intake and qualification — take a business's URL, check it's a reasonable fit, and scaffold the demo project.
  2. Scrape, with a three-way coverage diff — inventory the live site through three independent methods and diff them against each other, so nothing gets missed silently before anything is scraped for real.
  3. Design directions — turn the site's existing brand signals into a small set of concrete design directions, rendered as real homepage mocks with the business's own copy and imagery, for the prospect (or me, standing in) to pick from.
  4. Bespoke build — build the picked direction out into the actual demo pages, not a generic template dressed up.
  5. Grounded chatbot, gated by golden-QA evals — a retrieval-augmented chatbot that answers only from the business's own scraped content, held to a fixed set of golden question/answer test cases before it's allowed to ship.
  6. Ship to Cloudflare Pages, with a live re-eval — deploy, then re-run the same golden-QA suite against the live URL, so "it passed in dev" is never the last word.
  7. Priced proposal and outreach kit — a one-pager with a concrete offer, plus the cold-email and outreach copy to go with it.

2. The discipline that matters

The number that actually enforces the discipline is 33/33: every instance's golden-QA suite has to pass in full, both before deploy and again against the live URL after deploy, or it doesn't ship. That re-check after deploy exists because a chatbot that worked against a local build can still break once it's serving from a different origin with different environment secrets — I've had that gap actually matter, which is why the live re-eval isn't optional.

3. Live instances

Two independent education-sector businesses have live demos running the full pipeline output — same grounded-chat architecture, independent content and brand per instance:

4. Evidence