What If Property Search Actually Used the Photos?
Property Visual Finder uses AI-powered visual analysis to extract features directly from listing images, enabling search and ranking based on what properties actually look like, not just how they're described.
The Challenge
Property search has not fundamentally changed in years. Platforms let you filter by bedrooms, price, and location, but the photos, which are often the first and most influential thing a prospective renter or buyer looks at, play no role in how listings are filtered or ranked.
The result is a familiar frustration: scrolling through dozens of listings that match on paper but miss completely on feel. Renters hunting for a bright, open-plan kitchen or a south-facing garden have no way to surface those properties without opening every listing individually.
93% of home seekers form their first impression based on property images, and 94% of buyers are more likely to view a property with high-quality photos
The photos are everywhere. The intelligence to read them is not.
The Solution
Property Visual Finder uses AI-powered visual analysis to extract valuable information directly from property images and rank listings based on specific property features. The goal is not to replace existing real estate platforms but to complement them and integrate where possible. Initially focused on the UK real estate market, the solution will expand to the Greek market, with scalability to adapt to new markets and property dimensions as market needs evolve.
Key Components
- A Fine-tuned visual AI model trained to identify specific property features
- Multi-dimensional ranking model for property listings
- Tailored prompt engineering to interpret images and provide highly specific, real estate-relevant analysis
An overview of the solution is available here.
Behind the Build
I led this project as both Product Manager and AI and Technical Lead, working with a team of two engineers on implementation.
The most interesting challenge was the model itself. General-purpose vision models are trained to recognise objects broadly, they can identify a kitchen, but not whether it is open-plan, well-lit or recently refurbished. Getting meaningful, real estate-relevant outputs required fine-tuning the model on property-specific images and iterating extensively on the prompt engineering to produce assessments that felt credible to people who work in the industry every day.
The ranking layer added another layer of complexity. Different users weight features differently, a buyer prioritising natural light ranks listings differently from one focused on outdoor space. Building a ranking model flexible enough to reflect those differences, while remaining fast enough to be useful in a live search context, involved significant architectural trade-offs.
On the product side, I led the market and user research that shaped the feature set, defined and managed the development roadmap and took on the commercial side, pitching the solution to real estate agents and negotiating the terms of our first integration.
Results
The solution has been integrated into the internal system of a UK-based real estate agency, enabling agents to filter and prioritise properties more effectively and efficiently for their clients.
Agents report a 17% reduction in time spent on properties that do not meet their clients’ criteria, time that goes back into higher-value conversations.
The integration also surfaced something unexpected. Agents began using the tool’s visual assessments as a structured way to brief clients, giving them a shared language for describing what they were looking for. That was not a designed feature, it emerged from real use.
