RealPage Deploying Gen AI to Speed Leasing Process

Proptech players are rushing to form tech partnerships with chatbot makers, licensing deals that give them access to a generic LLM (large language model) and the fundamentals on training the bots for domain-specific tasks.

RealPage, the revenue management platform, has licensing agreements with Microsoft, Google and other LLM makers to use their Gen AI technology, which the company is training for a variety of use cases that focus on speeding the leasing process.

Evan Davies, RealPage’s chief technology officer, told GlobeSt. the LLM deals are similar to API licenses: RealPage can use the technology to make proprietary chatbots.

“We can incorporate (LLMs) into our applications stack and surround it with our applications,” he said. “LLMs are becoming a utility.”

RealPage has organized an advanced technologies team to train its bots, a process Davies described as “tuning down” a generic LLM and focusing the prompts—the questions you ask the bot—on the precise domain-specific queries that need to be answered.

“What you release into production is a tuned, trained bot solution. I don’t need it to give me a description of poetry, I just need it to focus on the problem we’re trying to solve,” Davies told us.

Davies said RealPage is finishing the “maturing process” for several bots the will be put to work as support services to dramatically speed the leasing process. With Gen AI, “dramatic” can mean delivering information that usually takes weeks for humans to aggregate in a matter of seconds.

“We believe that a lot of things GPT brings to the table [are applicable to] the leasing process. Our objective is to make it more efficient, to reduce the friction of the process and to centralize its activities,” Davies said.

“It’s a funnel. If you want to do something with your property, it’s a funnel process that you go through in order to attract market to address, make decisions and convert,” the CTO explained.

“I think anywhere in that process large language models can really optimize and reduce the amount of inaccuracies humans provide, by having the domain of information that’s already known about the process and about the specific assets,” Davies said.

“It can help address turnover issues, it can help address inaccuracies, it can help address a number of different areas specifically around repetitive tasks, moving somebody quickly through the leasing process and getting them into a unit,” he said.

“That’s where I see the biggest bang for the buck from a value proposition from this technology right now,” Davies added.

According to Davies, the tuning-down process is critical to controlling issues associated with early versions of ChatGPT, including inaccuracies and “hallucinations”—basically where the bot makes stuff up.

“What you do with a large language model is you feed it a corpus. It’s a corpus of information it consumes, and it can predict what a good response is based on the underlying corpus,” he said.” You get better results out of this if you take these things into domains and then tune the domains.”

“You have to set the context of what you’re trying to do—you’ve got to set the tone, the context—and give it a hint of which questions might produce hallucinations,” Davies said.

With the intellectual prowess of bots increasing exponentially—a consortium of tech giants is building a neural network consisting of 22K Nvidia AI chips in San Francisco—we asked the RealPage CTO how long it will be before advanced LLMs are making real estate investment decisions—assuming that all of the issues with inaccuracies and hallucinations are dealt with in coming months.

“I’d be very surprised—and I’ve been proven wrong in the past—if we let it make decisions about economics in our domain for the foreseeable future,” he said.

“Right now, it doesn’t have the wherewithal—and it might someday—it doesn’t have the wherewithal to collect all this information and do the necessary calculations to figure out what it should be demanding,” Davies noted.