Rapid Site Builder — Hire Your AI Ops Team

Inspiration

Small businesses don't need another site-builder wizard — they need an ops team they can't afford. We asked: what if hiring your web team was as easy as describing your business in one sentence?

What it does

Type one line about your business and watch a real AI agent team — Theo (orchestrator), Aria (research), Leo (layout), Noa (copy), Sam (SEO), Max (publish gate) and Phoenix (observability) — design, write and ship a hosted website live in front of you. First draft appears in ~3 seconds, the crew refines it on screen, and one click publishes it to a public URL. The same team then keeps operating the site from a Monday-style Agents Dashboard: uptime checks, update proposals, security catches — every action behind a human-approval gate. You're not buying a site builder; you're hiring a supervised, auditable AI ops team that replaces 1–3 ops hires and is live in minutes.

How we built it

• Google Cloud Agent Builder / ADK: a multi-agent crew deployed on Vertex AI Agent Engine, driven via create_session + phased stream_query, streamed to the browser as a character conversation (SSE).
• Gemini everywhere: the crew runs on gemini-2.5-flash; hero imagery is generated by Gemini 3 Nano Banana Pro on the global endpoint with a GCS cache.
• Arize partner track: every single run fires a real Arize Phoenix MCP tool call (dataset website_build_run_quality_signals) before the approval gate — observability is in the loop, not bolted on.
• Infra: Express proxy on Cloud Run; publishing renders script-free static HTML to GCS served at /sites/{id}; Firebase Hosting fronts the app.

Challenges we ran into

• Agent Engine ends a stream_query turn at the first transfer_to_agent — we built a phased multi-turn session driver to walk the whole crew.
• Firebase Hosting kills SSE streams after ~60s — the frontend streams straight from the Cloud Run origin with CORS.
• Gemini 3 text models are allowlist-gated for crew use on Agent Engine — we kept Gemini 3 in the running stack via the Nano Banana Pro image path and a one-var rollback plan.

Accomplishments we're proud of

A build can never end in an error: schema-bearing retries plus a deterministic, category-aware fallback spec guarantee every visitor walks away with a publishable site. Draft-first UX in under 3 seconds. A real partner-MCP turn on camera in every run.

What we learned

Multi-agent UX is a show — streaming the crew's conversation turned a 2-minute wait into the product's best feature. And observability-as-an-agent (Phoenix) makes eligibility and quality the same thing.

What's next

Login + cross-device My Sites, BYOK vs managed pricing tiers, user photo/video uploads feeding builds, and the full interactive agents dashboard — the control plane for hiring your AI ops team.

Built With

• adk
• arize-phoenix
• cloud-run
• express.js
• firebase
• gemini
• gemini-3
• google-agent-builder
• google-cloud
• javascript
• mcp
• node.js
• vertex-ai