It’s a cliche but remains true nonetheless: the three rules of real estate are: 1) location; 2) location; 3) location. Even the most sophisticated real estate or property tech firms have found that their AI is only as good as it predicts granular location dynamics.
Two years ago we started working with a leading REIT on AI-based dynamic pricing. Their data scientists had successfully created a model that generated hundreds of basis points in incremental revenue growth in several test locations. However they faced two major obstacles to rolling it out nationally to their thousands of locations:
- Dynamic pricing wasn’t a single model but really 7 chained models interacting with each other. Updating one model could lead to unintended consequences to the rest of the chain, which meant they needed a simple way of validating model updates before rolling out fully into production.
- As a REIT with thousands of locations with dozens of different unit types, each with their own distinct local demand and supply dynamics, there was no one national model they could apply.
When you combined (1) and (2) together, that meant their single dynamic pricing model would need to turn into tens of thousands of distinct models (the 7 chained models multiplied by the thousands of locations each requiring their own permutation). And this would apply not just to dynamic pricing, but to just about all the use cases this REIT was considering for AI. If their AI investments were to pay off, they would need to scale their ML model operations in order to apply to each location.
Standard MLOps practices usually require one ML engineer dedicated to each model in production given the amount of manual oversight required to deploy and then monitor the performance on an ongoing basis. At the time, however, their data science team consisted of 5 data scientists with only a single machine learning engineer.
But instead of thinking about any one specific model getting into production, the Chief Data & Analytics officer sought to industrialize his team’s AI processes so that even a small team with few machine learning engineering resources could deploy, monitor, manage, and continually optimize hundreds or even thousands of machine learning models across dozens of business units and thousands of locations (thinking of models as cattle, not pets).
Working with Wallaroo, he sought to rebuild his team’s ML deployment and model operations from the ground up with a different approach – one that relied on automation, standardization/repeatability, and centralized control (something we call no-ops / low-ops for ML model operations). Instead of ad hoc processes for each model, his data scientists would be able to deploy a model into a production or testing with a single line of code even as he applied AI to different use cases. Validating models before full rollout was simplified with Wallaroo’s advanced support for various testing frameworks. And his data scientists would be able to monitor all models in testing and production from a single UI, with automated alerts flagging anomalies or when the models started to drift out of preset baselines.
What this all enabled was for the CDAO to move AI beyond a few isolated proof of concepts to instead become a full-fledged program with scalable operations. This way, his data scientists and ML engineers could continue focusing on making a business impact and not on server operations.
In addition to this client, we work with other real estate enterprises, including RealPage, the world’s largest provider of property management software, on renter acquisition, retention, and marketing use cases. If you are in real estate or proptech, reach out to us to learn more about our capabilities in time series, demand forecasting, natural language processing, computer vision, and other scenarios applicable to your industry.
