Case study 02

A pipeline that builds, reviews and sends demo websites for local businesses.

Role
Solo engineer
Year
2026, live
Stack
FastAPI, Next 16, n8n, PostgreSQL, Jinja2, Cloudflare Pages
Links
quicksite-ph.com ↗(private repo)
535commits on main1
18locale-driven site looks2
10nodes in the daily n8n pipeline3
6architecture decision records4
  1. git rev-list --count origin/main in the quicksite repo
  2. directories under templates/looks, excluding the two shared ones (_photos, _shared)
  3. n8n_workflows/daily_pipeline.json
  4. docs/adr in the quicksite repo

The problem

A small local business with a weak or missing website is a real lead, but only if someone shows them what a better site would look like. Doing that by hand, one business at a time, does not scale past a few a day.

QuickSite automates the whole path: find the business, check it is worth contacting, build a demo site from its real details, review that site, deploy it, and reach out.

What I built

A daily pipeline plus a dashboard to supervise it.

  • Find and qualify: scrape Google Maps, enrich each result from Google Places, then score and filter the leads.
  • Generate: Claude writes the site copy into a strict schema, which a Jinja2 look turns into a static site. There are 18 looks, driven by locale rather than hard-coded to one country.
  • Review: rule-based checks cover content, data and pipeline state. Only when they flag something does an AI reviewer run.
  • Deploy and reach out: the demo goes to Cloudflare Pages, and an SMS goes out with a suppression list so nobody is texted twice or after opting out.

Architecture

n8n daily run10 nodesFastAPIrouter, service, repoLead sourcesMaps, PlacesClaudecopy, reviewSemaphoreSMS outreachCloudflarePagesPostgreSQLDashboardNext 16

AI design

  • Strict schema out: Claude returns site content as JSON validated against a Pydantic schema. A bad shape is retried, and a shared call budget caps the worst case at 7 model calls per site.
  • Fallback: Claude is primary and Gemini is the fallback, behind one small protocol.
  • Cheap reviewer, only when needed: the review model is the smallest Claude, and it runs only when the rule-based checks flag something.
  • Honesty checks: fields that would invent facts are stripped, and long verbatim copying from the source listing is detected.
  • Untrusted input: scraped listings and reviews are treated as untrusted personal data. They never go into the system prompt and are never logged.
  • Spend cap: a monthly AI cap is checked before each call, and a skipped review is logged with which review it was.

How it was built

Specs first, then plans, then parallel implementers per task with a reviewer. A hook blocks destructive database and git operations, so those stay human-only. Technology and model choices are recorded as ADRs.

Tradeoffs

  • No queue worker. Background work runs through FastAPI BackgroundTasks. It is simple and enough at this volume, and it means a crash mid-run is not resumed automatically.
  • Deploy is narrower than the generator. The generator handles any business type in any country, but the deploy step still accepts dental leads only.
  • Manual deploys. GitHub Actions is off for this repo, so merging to main does not deploy. Deploys are manual and need my explicit go-ahead.
WorkHire me