Bedre boligannoncer med AI

Baggrund

For the past decade, danbolig has been a digital front runner among real estate agents. From the first responsive website in the industry to “Our neighborhood”, the internationally award-winning data visualization platform for home seekers and buyers.

danbolig stays ahead by constantly prioritizing innovation – with a roadmap of continuous projects in all corners of the digital space.

As with any true innovation program, many of these sandbox projects never get realized – but the most promising ones do.

danbolig has long experimented with integrating predictive AI in the home searching process, a field with enormous potential but equally enormous data requirements.

While this work continues, the recent boom in generative AI begged the obvious question: Is there any business value to be realized NOW, using this rapidly maturing technology? Or should we simply regard it as a novelty for amusement and attention-seeking?

With an organization and agency collaboration which has, for years, developed a way of working that allows for rapid innovation, iteration and deployment, danbolig immediately set out to solve business problems with OpenAI’s models.

It didn’t take long to find the glaring opportunity of content production: Real estate agents are faced with massive content demands every day, most of it repetitive, unforgiving and short lived.

With a workforce that’s – generally speaking – way more passionate about selling houses than writing, giving them a tool to produce more persuasive, compelling and effective content in next to no time seemed like a no-brainer. Or at least an ‘artificial-brain’-er.

Løsning

The solution was to create the danbolig AI Assistant for editors – a content power-tool to help danbolig employees make the most of their time and talents.

But while the “idea” of using AI for content production is hardly very innovative – and maybe barely an idea at all – the idea of embedding it deep within the daily workflow of employees is much more interesting. With no alienating technobabble or intro courses, and no disruption of ‘normal content processes’, we wanted to provide unquestionable improvement to the eventual content result, not simply a gimmick.

Our solution integrates an OpenAI-powered large language model (LLM) directly into a market leading content management solution. We used OpenAIs GPT3 – davinci 003 model through Azure as our LLM foundation and Umbraco as the CMS.

We predefined relevant AI prompts and allowed them to be triggered and customized by simple and familiar interface action points like buttons or drop down menus, essentially making a “generative AI prompt designer” available at the fingertips of the content editor.

We have integrated the LLM naturally in the authoring dialog; this allows authors / editors to not leave the CMS to finish the task. Therefore, we rewrote content components in the Umbraco CMS to be connected to the LLM endpoint at all times.

While the house listings themselves are regulated, the AI assistant adds many tools to optimizing and improving existing content. And for all other content – from neighborhood or store descriptions, to invitation emails for open house events, the Assistant helps create content from scratch.

The solution is extendable and customizable. We are now working on finetuning our version of the foundational language model, training it specifically on danbolig vocabulary and brand tone of voice, as well as specific customer segments and real estate descriptions, to produce greater variety and higher quality.

With more robust segmenting, danbolig agents will be able to create segment specific landing pages to enable true personalization at scale.

Another feature in the works is building out the content creation feature to include computer vision to create accurate listings, including neighborhoods and descriptions, simply by uploading a photo with localization data.