Case Studies

A production AI concierge across voice, SMS, and web

Stay in Lakeland

Visit the live site

Role

Founder & Builder

Timeline

2026 - Present

Key Skills

Voice AI
AI Agents
LLM Tool-Calling
Product Strategy

Voice

SMS

Web chat

One AI concierge

voice, SMS, and web

Grounded in live data: Google Places, events, and weather.

Project Overview

STR guests ask the same questions at all hours: WiFi password, gate code, where to eat that is still open, how to book another night. A host cannot answer them all, and a generic chatbot that invents an answer is worse than silence when a guest is at a locked door.

So I built one concierge that answers on the phone, over text, and on the web from the same brain, grounded in real data rather than guesses. I founded Stay in Lakeland, a local guide for short-term-rental guests in Lakeland, Florida, and built its core: an AI concierge (persona "Heather") that answers guests from a single brain. Live at stayinlakeland.com, a solo build from PRD to production.

Key Challenges

  • Every channel (voice, SMS, web) has different constraints, but users expect one consistent assistant.
  • A wrong answer about a gate code or a restaurant that just closed erodes trust instantly.
  • Private property details (WiFi, gate code) must never leak to the wrong guest.
  • The whole thing has to pay for itself without taking a cut of bookings.

Solution & Approach

One brain, three channels

A single OpenAI-compatible concierge endpoint serves voice (Retell), SMS (Twilio), and web chat, with output styled per channel (spoken vs texted vs on-screen). Build the logic once, not three times.

Grounded, not guessing

Real tool calls fetch live restaurant hours from Google Places, local events, and weather at request time, with guardrails so the concierge never invents a venue and never claims a booking is confirmed. It runs on Vercel AI Gateway with a fast primary model and an automatic fallback.

Private data, gated by booking

The number a guest dials or texts maps to their property, and that home's private details (WiFi, gate code) are only unlocked while the booking is active. Access is a verified boundary, not a prompt instruction.

Results & Impact

  • One AI concierge serving voice, SMS, and web from a single brain

  • Live tool calls (Google Places, events, weather) instead of guessed answers

  • Private property info gated to active bookings only, a real security boundary

  • An advertiser-funded model: local businesses pay to be recommended, instead of a cut of every booking

Key Learnings

Trust is the product

For a guest at a locked door, a confident wrong answer is worse than a handoff. The concierge is designed to say what it does not know.

Own the model, own the tradeoffs

Running it on a gateway with routing and fallbacks (not a single vendor wrapper) is what let me tune cost, latency, and reliability myself.