PropTech · Agentic AI · Mobile App
Scout is a conversational way to find a home to rent in Dubai, explored for Property Finder. You say what you want in your own words. Scout keeps a short memory of what it understood, marks what it guessed, and shows only the homes that match. It runs on a real model. Nothing is mocked.
A prototype, made as a design exploration. Not affiliated with or endorsed by Property Finder.
The challenge
Search with filters is easy to read. What you asked for sits in one place, you set every part of it, and the results are clearly built from it.
A chat loses that. What you want is spread across messages: some said, some implied, some changed three messages later. And you can’t see what the results are based on.
Filters
One place. You set every part of it, and the results come from exactly this.
Chat
The same wishes, spread over four messages. Some were implied, and some changed along the way.
The challenge: keep a memory of what you asked for, easy to change and clear about what the results are based on, without a filter panel.
The design
Two rules hold it together
Nothing changes silently.
What you see in the memory is exactly what the results were built from.
Filters do one thing well: they’re always there. I kept that and dropped the panel. One line above the keyboard, in thumb reach, written in words instead of controls.
No quick-reply chips either. Chips are filters by another name.
An agent that only does what it’s told is a form. Scout reads between the lines: a dog means a pet-friendly building, kids mean schools nearby probably matter.
What must be true filters. What’s only likely ranks. Every guess is marked as one.
AED 88,000Bloom Towers, Jumeirah Village CirclePets allowed
AED 92,000Uptown Mirdif, MirdifUptown International School · 4 minAED 138,000Executive Towers, Business Bay640m from metroThe memory opens as a sentence you can disagree with, phrase by phrase. Taps and words both go through the conversation, so the memory and the chat never disagree.
“Just for now” sits on top of what you said before and never replaces it. Drop it, and you’re back where you were.
Saying “no results” is easy. A smart agent works out why, and what would change it.
Scout first widens to what you’d likely accept anyway: the neighbouring areas. If still nothing fits, code tests every requirement, finds the one blocking the most homes, and Scout names it with what you’d get. It never drops a requirement without asking. You decide what to give up.
AED 138,000Executive Towers, Business Bay
5 homes, from AED 72kHow the agent works
Understanding a person is fuzzy work. Matching homes is exact: a budget is a number, and a building either allows pets or it doesn’t. So each half goes to what does it best. The memory on screen is where they meet.
The memory on screen is the single source of truth. Results are built from it and nothing else, and every change to it is said out loud once.
Technical choices for this prototype
Building it
I decided what Scout should do, how it should talk, and where the model stops and code takes over. Claude Code wrote the interface, the prompts, the matching code and the tests. I judged every answer.
The prompts are design work. Scout’s voice lives in the reply prompt: short sentences, plain international English for a city of expats, no idioms, always the area, the price and the distance.
The test it has to pass
A script replays a seven-turn conversation against the live model. It fails if Scout…
Most business customers on Tradeling ordered through sales agents, not the app. I redesigned it so each type of business gets its own home page. Conversion rose around 38%, and GMV from the app around 30%.
QFunder is Tradeling’s new fintech vertical. Banks rarely lend to small businesses, and when they do, it takes weeks. We built an MVP to test a faster way: apply online in minutes, get funded in 24 hours. We built and tested it in two months.
