NerdWallet Uses Machine Learning on AWS to Power Recommendations Platform

NerdWallet is a personal finance startup that provides tools and advice that make it easy for customers to pay off debt, choose the best financial products and services, and tackle major life goals like buying a house or saving for retirement. The company relies heavily on data science and machine learning (ML) to connect customers with personalized financial products. “We realized early on that data science was going to be essential to building a more personal product and user experience,” says Ryan Kirkman, senior engineering manager at NerdWallet.

As the company’s engineering team began to deploy its first ML models to production, Kirkman and his team found that the process took much longer than expected. He says, “It would take months to go from prototype to production and there were many inefficiencies along the way.”

At the time, NerdWallet data scientists used a largely manual approach to managing ML libraries, which wasn’t optimal from a cost or workflow perspective. “Our data scientists had to install things by hand and deal with whatever environment the last data scientist had left on the machine,” says Sharadh Krishnamurthy, staff software engineer at NerdWallet.

The company needed to solve its engineering plumbing problems so its data scientists could train ML models more effectively, speed up the process from concept to delivery, and focus more on high-value projects. “The faster we could ship models to production, the faster our data science team could iterate on those models and the better we could make our product experience,” says Kirkman. “Reducing the feedback loop would significantly improve our ability to execute on data science projects.”

As a startup, NerdWallet didn’t have the resources required to reinvent the wheel. “The key question for a startup is ‘how do we add business value fastest?’ We wanted a machine learning platform like some of the big companies had, but we weren't in a position to invest much in it,” says Krishnamurthy.

NerdWallet was already using a number of Amazon Web Services (AWS) solutions, including Amazon Elastic Compute Cloud (Amazon EC2) and Amazon Elastic Container Service (Amazon ECS). The team decided to add Amazon SageMaker to the mix. With this fully managed service, the company could leverage underlying Amazon EC2 instances, including Amazon EC2 P3 instances with NVIDIA V100 Tensor Core GPUs, and their existing Amazon ECS image-building pipeline to reduce the time it takes to train ML models. “Amazon SageMaker basically provided us with machine learning as a service,” says Kirkman.

Adopting Amazon SageMaker enabled NerdWallet to quickly modernize its data science engineering practices. “We essentially unlocked business value in two months,” says Kirkman. “That wouldn’t have been possible otherwise.”

The new solution also helped the company remove roadblocks and speed time-to-delivery. “Amazon SageMaker makes it easy for our data scientists to become the core owners and drivers of their work, rather than having multiple handoffs and having to re-implement everything,” says Kirkman. “We’re providing a guided path that makes solving these surrounding infrastructure problems easier from a platform and engineering perspective, while also accelerating the work of our data scientists. It’s a win-win.”

NerdWallet’s first project leveraging the new approach was a recommendations platform powered by TensorFlow. Previously, NerdWallet provided customers with a list of potential credit cards, but it had no way to forecast the likelihood of acceptance. Now, using Amazon SageMaker and machine learning, the company can more effectively match customers to the right financial products for them.

The use of Amazon SageMaker and Amazon EC2 P3 instances with NVIDIA V100 Tensor Core GPUs has also improved NerdWallet’s flexibility and performance and has reduced the time required for data scientists to train ML models. “It used to take us months to launch and iterate on models; now it only takes days,” says Kirkman.

Amazon SageMaker has enabled NerdWallet’s data scientists to spend more time on strategic pursuits. Kirkman says, “We can now focus more energy where our competitive advantage is—our insights into the problems we're solving for our users.”

For example, NerdWallet is now building an inference storage system that will allow teams to easily access and use predictions data scientists are producing in other areas of the company. “I don't think we could have made the business case for that system without having the streamlined workflow for training that our data scientists now have,” says Kirkman. “It would have been too expensive, too risky. Streamlining the workflow and enabling standardization unlocks so much more value for us.”

Amazon SageMaker also helps NerdWallet keep costs in check. Because the company can pay per use rather than pay for infrastructure to run endlessly, expenses are only incurred when compute resources are needed. “We've been able to reduce our training costs by around 75 percent, even while increasing the number of models trained,” says Kirkman. “We can do so because we’ve moved from a model of running a single large Amazon EC2 instance 24/7 to running whatever instance type we need via Amazon SageMaker on demand.”

Building its ML platform on Amazon SageMaker also means NerdWallet’s small IT team can immediately take advantage of industry advancements. “From an infrastructure and technology perspective, using Amazon SageMaker helps us significantly advance our machine learning platform efforts, features, and functionality,” says Kirkman. “And by working with AWS, we effectively get to stand on the shoulders of giants.”

The use of machine learning and Amazon SageMaker represents a shift in how NerdWallet is leveraging technology to differentiate itself in a crowded and competitive financial services market. Going forward, Kirkman and his team plan to continue to use technology to offer value-added services. “Helping customers with any sort of financial decision they have and being aligned with the consumer are key differentiators for us,” says Kirkman. “Using data science and machine learning helps us double down on that effort.”

To learn more, visit aws.amazon.com/sagemaker.