A queue-driven pipeline that finds local service businesses without a website, researches their brand, builds them a site with an agentic CLI, and deploys it over SFTP, end to end, unattended.
Google Places enrichment, agentic build, SFTP deploy, status writeback
Between a queued row and a live URL in the happy path
Output is verified before anything reaches the host
Lead generation for web work is manual on both ends: finding businesses that genuinely have no website, then producing something concrete enough to be worth their attention. Doing that by hand caps the whole motion at however many sites one person can build in a week.
Built the risky half first, a Supabase jobs table plus a worker that consumes a queued row, builds a real site through an agentic CLI against a fixed rule set, verifies the output, deploys it over SFTP, and writes the live URL back. Discovery and dashboard were only built once one hand-inserted row reliably produced one live site.
The scraper and the dashboard are well-understood mechanical work. The genuinely unknown question was whether a database row could reliably trigger a real, deployed website built by an agentic CLI. That got built and proven manually before anything upstream or downstream existed.
The Google Places discovery stage writes queued rows into the same jobs table the worker already consumes, and the dashboard reads and triggers against it. No new integration surface was introduced for phases two and three.
The build step runs against a fixed rule set rather than free-form prompting, so output quality is a property of the rules, versioned and improvable, instead of a property of any individual run.
A long-running local worker polls for queued jobs rather than receiving webhooks. Builds run on a machine with the CLI and credentials already present, and a crashed worker resumes by simply polling again, no missed-delivery replay logic to write.
The worker's only obligation to the rest of the system is writing status and live URL back to the job row. That single contract is what let the dashboard be built last without changing anything already working.
If something here looks like what you're facing, tell me about it. The first conversation is an architecture discussion, not a sales call.
