Rocket Mortgage lays foundation for generative AI success

The US mortgage lender, which has a history making the most of machine learning and AI, is taking a model-agnostic approach to generative AI, boosted by a versatile data platform tuned for speed.

To succeed in the mortgage industry, efficiency and accuracy are paramount. So too is keeping your options open. That’s why Rocket Mortgage has been a vigorous implementor of machine learning and AI technologies — and why CIO Brian Woodring emphasizes a “human in the loop” AI strategy that will not be pinned down to any one generative AI model.

The Detroit-headquartered retail mortgage lender has been deploying machine learning and AI for more than a decade and is among the few pioneers that have released generative AI capabilities into the marketplace.

“We have multiple generative AI cases in production today and have for about one year,” says Woodring, noting that the company has one genAI chatbot under development, for example, designed to listen and comprehend as well as it speaks.

Another genAI assistant developed by Rocket analyzes applicants’ employer names to ensure that employers that could be entered under various names are understood to be one and the same, vastly speeding up the decision-making process. For example, most people know Google and Alphabet are the same employer. Using that human knowledge to train a genAI assistant to verify employer identity is far more efficient than building a database of parent corporate names to cross check against their subsidiaries or more common company identities, Woodring says.

Early to put generative AI into production, Rocket Mortgage did so with proper guardrails and guidelines in place to convince investors and regulators that it was implementing the technology in a safe and responsible manner, Woodring adds. The company now has several business processes fully automated with homegrown code and AI. But if any generative AI application involves a decision, such as whether to grant a mortgage loan, there is always a “human in the loop,” Woodring says.

“With genAI-powered copilots or systems, which is a lot of what we’re building, we find that, with the combination of a genAI model that knows everything posted on the internet for years, and human judgment, the accuracy of the decision is going to increase 10% to 15%, which is huge,” he says.

Analysts agree that incorporating human input to sign off on decisions and outcomes of generative AI processes is proving to be an essential driver of early genAI success.

“Generative AI is becoming the virtual knowledge worker with the ability to connect different data points, summarize and synthesize insights in seconds, allowing us to focus on more high-value-add tasks,” says Ritu Jyoti, group vice president of worldwide AI and automation market research and advisory services at IDC.

“It is transforming processes like loan underwriting, but human-in-the-loop is critical as it requires 100% accuracy without fail to be truly effective and viable, as the technology is still nascent,” Jyoti says.

Ramping up for model-agnostic AI

Rocket is as much an engineering company as it is a mortgage lender, with more than 1,000 engineers and 600 data scientists working together to build most of Rocket’s code in-house — a major advantage to its innovation efforts.

When Woodring joined the company in 2017 as CTO to lead the product engineering team, one of his top priorities was accelerating Rocket’s embrace of the cloud.

“One of the first things that I did after I joined, six months in, we declared that going forward, all of our new technology would be built in the cloud,” he says.

Today, 60% to 70% of Rocket’s workloads run on the cloud, with more than 95% of those workloads in AWS. The rest are on premises.

According to Woodring, the company’s first machine learning models were developed more than 10 years ago, to automate tasks such as marketing, lead generation pattern recognition, and loan origination processes.

But in the past five to six years, AI use at Rocket “has kicked into overdrive,” Woodring says. For example, roughly two-thirds of loan applicants’ income verification is performed 100% by machine learning models and AI technology today, he says.

“Almost every aspect of our business is now touched by ML or AI, task automation, pattern recognition, and data analysis,” says Woodring, reiterating that whenever a decision is required, a human is always part of the closing process.

Rocket’s engineers and data scientists are developing generative AI models using AWS Bedrock and Anthropic AI technology. Despite being primarily an AWS shop, Rocket has taken a model-agnostic approach to generative AI platforms. Rocket Companies CEO Varun Krishna, an experienced technology executive with stints at PayPal and Microsoft, has direct relationships with all the AI foundational model providers, including AWS, Anthropic, OpenAI, Google, and Mistral, Woodring says.

“We want to work directly with all of them because we want to know what’s coming,” Woodring says, adding there will not likely be one clear “winner” in this complex AI arms race. “It is more likely you will see these different AI models tuned for different use cases. We want to be able to plug in the right model at the right time. It’s a powerful strategy.”

One of the most valuable aspects of AWS Bedrock, Woodring says, is that it establishes a standard data platform for Rocket, which will enable the mortgage lender to get its data “very quickly” to the right AI model. In other cases, Rocket will test out various AI models and “see their efficacy in different tasks,” Woodring says. “That’s really valuable.”

The CIO maintains that AWS is of a similar mindset and “not committing itself to one winner,” he says. “That really resonates with our strategy of choosing the right AI model for the right job.”

Modernizing data operations

CIOs like Woodring know well that the quality of an AI model depends in large part on the quality of the data involved — and how that data is injected from databases, data warehouses, cloud data lakes, and the like into large language models.

As such, paramount to Rocket’s AI push is the creation of a modern data platform that incorporates 10,000 terabytes of data stored in on-prem data warehouses for more than a decade and semi-structured data stored in an AWS cloud lake. Like most enterprises, Rocket continues to operate some of its own data centers for older technology still in use.

Rocket is evolving its data lake strategy into an AWS data platform that can support structured, semi-structured, and newer unstructured data with semantics and taxonomies and an API on top to make it “significantly more discoverable and usable” for human and software consumption, Woodring says.

This will push data into repositories best ingested by AI models. Attempting to clean the entirety of Rocket’s data is unnecessary and cumbersome and will slow down the process of deploying next-generation applications, he says.

“We are a data-driven business, and the business we’re in, mortgage origination, really is a data-processing business,” Woodring says.

The company’s active generative AI engine and its next-generation data platform are being designed to deliver all forms of data quickly, curated for specific tasks and in the proper formats to advance its portfolio, the CIO says.

All it takes is the team and some time, he adds. “We prize being able to move fast here and be first to market with an idea.”

SUBSCRIBE TO OUR NEWSLETTER