The problem
AI site generators can produce pages that look finished but have broken links, missing accessibility basics or made-up business details. A small business owner usually cannot spot these problems.
What I built
A site generation backend, a quality gate that inspects every generated site before it can be published, and publishing to static hosting.
My role
Founder and sole builder.
Architecture
- Express and TypeScript backend with a Prisma data model.
- Quality gate that parses generated HTML and runs accessibility checks without a full browser.
- Provider registry and planner that prefer free model tiers where they are good enough.
- Publishing of finished sites to Cloudflare Pages.
Technologies
How it works
The owner describes the business, Bizia generates the site, the quality gate checks structure, links and accessibility, and only a site that passes can be published.
Key engineering decisions
- Run quality checks on the generated HTML itself rather than trusting the generator.
- Check accessibility with axe-core inside jsdom so the gate runs cheaply without a browser.
- Prefer free model tiers through a provider registry to keep costs low for small businesses.
Challenges
- Keeping generated content honest: no invented addresses, reviews or claims about the business.
- Balancing a strict gate against a smooth first experience.
What I learned
The value of an AI builder is in what it refuses to publish. The quality gate turned out to be the core of the product.
Current status
Early access. The public frontend is live; the backend is not deployed to production yet.
Links
More case studies: Capital Intelligence OS · GridResolve AI · Rankelo · Vaani · TalkBot
