Aurelia Properties
Conversational property advisorA property assistant that knows the portfolio it is selling.
Aurelia Properties is a fictional business, created by Valtair to show how the system works.
At a glance
- Type
- Demonstration build, publicly reachable
- Sector
- Real estate (Dubai residential)
- Channels
- Web chat, embedded in the site and in each listing
- Grounding
- Structured listing and community records, queried per answer
- AI approach
- Grounded assistant with tool access to portfolio data
- Oversight
- Defers to a human advisor outside its scope
- Model providers
- Multiple, provider-agnostic
- Deployment
- Cloud, server-rendered with streamed responses
- Business
- Fictional, built to demonstrate the system
Aurelia Properties is a fictional Dubai estate agency we built to test one question: what does a property assistant need to know before it is worth talking to? Most site chat widgets are a contact form with a cursor. Ayla answers against the actual listing set, holds the requirements a visitor has already given, and hands over with the context intact.
The scenario
A buyer relocating to Dubai has a budget and very little sense of what it buys. Agency websites answer that with a filter and a results page, so the visitor is left to work out whether Business Bay or Dubai Marina suits them, what a service charge adds, and which listings are genuinely available. The advisors who could answer that work office hours, and a serious enquiry at 11pm on a Sunday waits until Monday or goes elsewhere.
What it demonstrates
Answers grounded in the live listing set
Ayla answers from the same records the page renders: reference, community, price, beds, baths, area, and handover quarter for off-plan releases. Ask what is available under a budget in a given community and the answer is the portfolio, not a plausible-sounding invention.
Contextual entry from any listing
Every listing card carries its own Ask Ayla button, so the conversation opens already attached to that property. The visitor never has to describe which one they mean, which is where a generic widget loses most of its enquiries.
Qualification before handover
Budget, community, property type, and whether the visitor is buying, renting, or looking at off-plan are gathered in conversation rather than as a form, and travel with the enquiry when an advisor takes over.
Cover outside office hours
The agency publishes Monday to Saturday, 9:00 to 18:00. The assistant covers the rest, which is the practical case for it: the enquiry is captured and qualified while the office is shut instead of arriving cold on Monday.
System architecture
AI and data components
- Assistant grounded in structured listing and community records
- Tool calls for listing lookup, filtering, and availability
- Requirement capture held across the conversation
- Refusal and handover behaviour when a question is outside what the data supports
- Provider-agnostic model orchestration
Backend and integrations
- Streaming chat endpoint with per-conversation state
- Listing and community data as the single source for both page and assistant
- Enquiry capture with the conversation attached
- Handover to a human advisor during office hours
Engineering decisions
The page and the assistant read the same records
A separate index for the assistant is the fastest way to get an agent quoting a price the page no longer shows. One source removes the drift instead of monitoring it.
Listing facts are looked up, never recalled
Prices, availability, and handover dates come back from a query on every answer. A model asked to remember a portfolio will confidently produce a number that was true last month, and in property that is a complaint rather than an error.
Deferring is a designed behaviour, not a failure
Legal questions, mortgage advice, and anything about a property outside the portfolio route to a human rather than getting an answer. An assistant that will not guess is what makes the answers it does give worth trusting.
Entry from the listing, not only from the corner
A launcher in the corner starts every conversation from nothing. Opening from a listing card means the first message already has its subject, which is most of the qualification done before the visitor has typed.
What this demo does not do
- The agency, its advisors, listings, communities, and client quotes are fictional. No property described is real and no figure is a market rate.
- The assistant is a single web surface. There is no WhatsApp, voice, or email channel in this build.
- There is no CRM behind it. Enquiries demonstrate capture and handover; they do not enter a real sales pipeline.
- Viewings are not actually scheduled. Arranging one is represented, not booked.
RAG and Knowledge Systems
For products that must retrieve and reason over private or domain-specific information with citations.
Explore capabilityAgents and Automation
For multi-step workflows that require tool use, decision logic, and human approval.
Explore capabilityProduct Engineering
For teams building complete AI-enabled products, from discovery through production monitoring.
Explore capabilityWant one of these for your business?
This was built to be used, not watched. Open it, try to break it, then tell us what you are building and we will show you how we would approach it.