I built an AI agent that answers my villa booking emails — Gemini + Google ADK

#agents #ai #automation #gemini

I host villas in Phuket. So I built an AI agent that answers my booking emails — with Gemini and Google ADK
published: true

tags: googlecloud, ai, agents, gemini

This article was created for the purposes of entering the **All Things Agentic Hackathon.

The chore

I'm a vacation-rental host in Phuket. Every week the same chain eats my hours:

read a guest email → decode vague dates ("around Christmas", "del 29 de diciembre al 5 de enero") → check the calendar → do seasonal price math with discounts → write a warm reply → remember to place a hold → follow up.

Every step is trivial. The chain is not. And one missed email is a lost $3,000+ booking.

That's exactly the "messy, multi-step chore" the hackathon's Taskmaster track describes. So I built the agent I needed myself.

What HostPilot does

A raw guest inquiry goes in — email or chat text, any language — and the agent handles it end-to-end:

Extracts dates (including cross-year stays like Dec 29 → Jan 5), party size, guest name.
Checks real availability against the bookings store.
Builds a quote: high/low season rates, 10% long-stay discount, cleaning fee, min-stay and capacity validation.
If the guest clearly wants to book — places a 48-hour hold in Firestore. A real state change, not just text.
Replies in the guest's own language (English, Spanish, Chinese in the demo) with a full price breakdown.

My favorite moment: when requested dates conflict with an existing booking, the agent doesn't just say "unavailable" — it inspects the calendar, finds the nearest free windows on both sides, verifies and prices each one, and offers the guest two concrete alternatives.

Live demo: https://hostpilot-406193234542.asia-southeast1.run.app
Code: https://github.com/Barmaley26/hostpilot
Video: https://youtu.be/LGkHqfLt3mY

The architecture

Google ADK 2.6 defines the agent: one Agent, five plain-Python tools (get_property_info, check_availability, quote_price, create_booking, list_bookings).
Gemini 3.5 Flash (via Vertex AI) does the reasoning, multilingual parsing and reply generation through function calling.
Cloud Run hosts the whole app (FastAPI + a one-screen "host inbox" UI), deployed straight from source.
Firestore persists bookings across instances and restarts.

guest text → FastAPI → ADK Agent (gemini-3.5-flash)
├─ get_property_info
├─ check_availability ──► Firestore
├─ quote_price
└─ create_booking (48h hold) ──► Firestore
← reply in guest's language + full action log

Three things that bit me

1. The model has no clock. "December 20" was parsed as 2024. One line fixed year resolution, including cross-year stays:


python
instruction = f"Today's date is {date.today().isoformat()}. ..."

2. Free tiers are not for agents. One agent turn = 5–6 generations (planning + tool calls + final reply). The Gemini API free tier (20 requests/day) died mid-testing. Switching to Vertex AI with ADC on Cloud Run solved it properly: no API keys in the container at all, billing through the project.

3. Autonomy needs guardrails in code, not in prompts. Every tool returns explicit ok/reason dicts; create_booking re-validates availability server-side. The model physically cannot double-book, no matter what it decides.

The lesson

Agents become reliable when the LLM plans and the tools decide. Everything that must be true — availability, price, capacity — is enforced in deterministic, testable Python. Everything that must be human — tone, language, judgment about intent — is left to Gemini.

The result: a guest writing in Chinese gets a correct, warm, fully-priced booking hold in Chinese — while the host sleeps.

Built solo with Gemini 3.5 Flash, Google ADK, Cloud Run and Firestore for the All Things Agentic Hackathon.

DEV Community