Hotel AI Concierge

The problem

Hotel guests ask two very different kinds of question. Policy questions — “Is breakfast included?” — live in documents. Commercial questions — “What would three nights in a sea-view room cost?” — need a number, and a wrong number is worse than no answer. An assistant that improvises both is a liability.

What I built

• Document answers with numbered, clickable citations from Azure AI Search, behind a relevance gate that prefers “I don’t know” to a weak match.
• Dated quotes and earliest-availability searches computed by code from the rate engine; the commercial answer is written from those same facts with no model call.
• Reservation requests stored only with the guest’s consent, for staff to confirm — never a booking, a hold or a payment.
• A protected staff panel with editable rates and availability, the request queue, and AI summaries and reply drafts that never see a guest’s identity.
• Live voice in Turkish and English over WebRTC with Azure OpenAI Realtime, server-side turn routing and a server-side allow-list of what can be searched.
• A dedicated Azure SQL database guarded at startup: the app verifies the exact server, database and engine it connects to, and never changes a schema in production.

Architecture

1. • Guest browser

     text chat · WebRTC voice

2. • Blazor host on Azure App Service

     .NET 10 · turn router · tool allow-list · answer composer

3. • Azure AI Search

     hotel documents, citations

   • Rate engine

     deterministic quotes

   • Azure OpenAI

     language and realtime voice

   • Azure SQL

     requests, rates, staff data

Key decisions

• No model touches a price

  A dated quote is arithmetic. The answer text is composed from the same facts the quote card shows, so the two can never disagree.

• Requests, not bookings

  Nothing in the request path can reserve, hold or charge — and an automated test fails the build if that ever changes.

• The server decides what is searched

  The browser sends a question; the server routes the turn and picks the index from an allow-list, so a public assistant cannot be pointed at private data.