Google for Startups AI Agents Challenge · Track 1 · Build

The salon owner sleeps.
The agents work.

An autonomous multi-agent system that drafts WhatsApp messages to recover cancelled bookings, responds to Google reviews in Turkish, and proposes off-peak promotions — all while the owner is busy with a customer in the chair.

Try the live dashboard → View source on GitHub

Three agents, one orchestrator

An ADK orchestrator on Gemini 2.5 Flash takes booking, review, and weekly-calendar events from production and routes each one to the right specialist sub-agent.

📵

No-Show Recovery

When a booking is cancelled, ranks the waitlist on a 50-30-20 service / time-window / loyalty weighting, picks the best match, and drafts a personalized Turkish WhatsApp opener.

28 s live

⭐

Review Responder

Classifies a new Google review by rating, drafts a tone-matched Turkish reply (empathetic for 1-2★, thankful for 5★), and pushes it to the owner approval queue. Never auto-publishes.

17 s live

📅

Calendar Optimizer

Runs every Monday morning. Finds the single biggest off-peak gap in the next 7 days and drafts one targeted Turkish promotion with an audience tag. Refuses to invent on empty data.

4 s live

Architecture

Every box on this diagram is a real Cloud Run service or MCP tool. Each MCP server runs as both a stdio subprocess for local dev and a Streamable HTTP service in production — same binary, two transports.

Google ADK 2.1 · Python Vertex AI Gemini 2.5 Flash MCP TypeScript SDK · 3 servers Cloud Run europe-west4 Supabase · copilot schema isolated FastAPI · cookie session + CSRF Secret Manager · 5 secrets

Built for judges to verify, not just trust

Each surface is a different angle on the same agents. The dashboard is bilingual: English decision summary above, Turkish customer draft below. Click any session, see the full reasoning trace.

Owner dashboard with bilingual decision pills and approval queue — Owner dashboard — green pill = English decision summary; raw payload = Turkish draft for the customer.

White-box reasoning trace showing 13 ADK events — White-box reasoning trace — every tool call, response, and final action captured in chronological order.

Marketplace economics calculator with live sliders — Marketplace economics — drag the sliders, the ARR projection re-computes live.

Five live surfaces

Same auth (cookie session or Authorization header), different lens on the same data.

/dashboard

Operator console

Pending owner approvals + recent agent actions, with a one-click button that fires all 3 agents in parallel.

/dashboard/trace/{session_id}

White-box reasoning

Every ADK event for a session — routing, tool calls, responses, final draft. The opposite of a black-box agent.

/dashboard/stats

Observability

Rollups over agent_actions and the approval queue. Shadow-mode percentage is rendered explicitly.

/dashboard/economics

Live ARR calculator

Five sliders, three KPI groups, two pricing tiers. Drag the assumptions, see what the business case looks like.

The salon owner sleeps.The agents work.