When was the last time you opened a finance app and really thought about what’s going on behind the scenes? Most people don't. You might log in to view your credit score if you are getting mortgage-ready or on the lookout for the best loan or credit card offer, but within a few seconds, you'll move on with your day.
In this article, we look behind the scenes at the technology powering those moments for more than 140 million people.
Wan Agus, CTO at Credit Karma, shares his insight into how the company is modernizing its data platform, blending traditional artificial intelligence (AI) with generative AI, and managing the complexity of scaling AI-first experiences.
Key Takeaways
Credit Karma runs more than 60 billion daily AI predictions to guide member decisions.
Intuit and Credit Karma unified tech stacks for seamless identity and data integration.
Balancing speed and trust is essential when AI influences financial lives.
A culture of mission, autonomy, and self-critical debriefs drives engineering resilience.
Agent-to-agent technologies could bring open banking-style transparency to US consumers.
From Retail to Fintech
Wan's career path did not begin in finance. He spent a decade in e-commerce at Walmart.com and Macys.com before joining Credit Karma. For him, the leap to fintech was less of a pivot and more of a logical progression.
Agus shared with me:
"In e-commerce, if you're selling AA batteries, everybody needs one. But when you're selling a sofa set, you really need to contextualize it. When you think about fintech, it's the same thing. Credit cards are like AA batteries; everyone needs them. But a home loan? That's a major decision, and you have to make sure it's the right thing for them as well."
That connection between context and trust is what pulled him into fintech. "In e-commerce, if you send more than one item, it's not the end of the world. In fintech, there is no room for mistakes. You have to protect members' privacy and data while still moving fast," Agus said.
The Challenge of Two Tech Stacks
Credit Karma's acquisition by Intuit in 2020 presented a new challenge: integrating two large yet distinct technology ecosystems. Credit Karma runs primarily on Google Cloud, while Intuit is built on AWS and its own internal systems.
"These are two different tech ecosystems," Wan Agus said. "Integrating them is not easy at all, but the moment we realized how much value you can unlock, it's all worth it."
He compared the process to a strategy game. "I look at it as playing two board games connected. And it's even more fun when you're really trying to figure it out," he said.
But this was far more than an engineering puzzle. It was a foundational task for the member experience. Agus explained:
"We first had to federate the identity systems. You don't want members relogging in and having to complete two-factor authentication every time they switch between products. During tax season, we saw that about 70% of Credit Karma members could use TurboTax without even logging in. That's huge."
The other big lift was data rationalization. "We know Neil over there in Intuit is the same Neil over here in CK. We have thousands of attributes, and we need to determine which ones are canonical. If we don't get that right, everything we build on top won't resonate," he said.
More Than 60 Billion Predictions a Day
The scale of Credit Karma's AI operation is difficult to overstate. The company now runs more than 60 billion daily AI predictions to guide personalized financial decisions.
Agus said:
"When I joined, it was around 60 billion predictions a day. Now it's more than that. And what do we do with these predictions? We determine what's best for you in your current financial situation. Is your credit card the best one for you? Can we lower your loan interest rate?
What about a mortgage? We don't want you clicking on offers just for the sake of it. If you apply and are not approved, your credit score will decrease. That's a lose-lose."
This isn't just a Credit Karma problem. Poor personalization can cause real financial harm. A misaligned recommendation can negatively impact someone's credit score, restrict their access to future opportunities, or result in wasted time and money.
Maintaining the scale of 60+ billion predictions requires continuous reinvention of infrastructure.
"Right now, we're migrating to NVIDIA Triton servers," Wan said. "You have to be on the cutting edge, knowing what's cutting edge today probably won't scale tomorrow. But you start early, so by the time it's ready for production, you can meet the demand." That constant iteration reflects a broader truth across the fintech industry. Speed matters, but stability matters more.
Behind the scenes, those 60 billion daily predictions run through a layered architecture that combines classic machine learning with newer generative models. Credit Karma's systems take in vast amounts of member data, apply thousands of attributes, and generate probability scores that help match people with the right financial products.
Wan described it as 'a continuous journey' rather than a one-off build. “When you have a massive amount of data, you have to rationalize what you are using and why, and then evolve the infrastructure underneath. That's why we are moving to NVIDIA Triton servers, so we can handle the scale while keeping latency low and costs manageable,” he said.
It is an operation that functions much like a recommendation engine, but with stakes that reach into members' credit histories and long-term financial well-being.
Walking the Line Between Innovation & Trust
In most areas of tech, speed is king. Move fast, break things, fix later. In fintech, the stakes are different. "That's giving me a lot of white hair," Wan laughed. "If you're in e-commerce, AI makes a mistake, you refund it. In fintech, it's people's financial lives."
To illustrate the challenge, he referenced Andrej Karpathy's framework of Software 1.0 (traditional code), 2.0 (machine learning), and 3.0 (generative AI).
"When I started, we were already rocking and rolling with Software 2.0, more than 60 billion predictions daily. Then GPT-3.5 was released, and within three months, we developed a chatbot that could discuss your financial life. The goal was to learn what members actually needed this technology for," Agus said.
Technology doesn't scale without culture. At Credit Karma, mission alignment is the anchor. "In our engagement surveys, mission is always number one. People feel it," he added. "I'm an immigrant. The first time I learned about credit scores was through Credit Karma. Millions of people have that same experience."
That sense of mission drives autonomy, but with clear boundaries, Agus explained:
"Autonomy isn't autonomy for its own sake. You need clear guidelines on what's non-negotiable, like privacy and safety. And you need to contextualize decisions. If people don't know why, they can't do their best work.”
The company has borrowed lessons from outside industries to strengthen this approach. "We had Blue Angels pilots come in, and they talked about how after every performance, they debriefed by being self-critical. It's not finger-pointing; it's pointing the finger at yourself. That's the culture we want," he said. It's this philosophy that shapes how Credit Karma experiments.
Not every AI project works, but the goal is to move quickly, learn, and adjust. "Most experiments fail. That's normal. But when they fail, you need observability, debriefs, and a celebration of learning. Otherwise, autonomy becomes chaos," Agus said.
Credit Karma’s CTO is clear that building responsibly isn't about slowing down. It's about knowing when and how to move: "You have to understand it's a risk-reward equation. If you take on more risk, you also need to make sure the observability is there so that you can react very quickly."
That mindset has guided Credit Karma's adoption of third-party AI tools. Agus said:
"If we had used our previous process to vet third-party tools, we would not have onboarded Copilot or Cursor. We had to adjust. Otherwise, we'd fall behind. But we balanced that by working with legal and security teams, making sure we could move fast without creating blind spots."
Looking Ahead: Agent-to-Agent Futures
When I asked Wan Agus what excites him about the future, he pointed immediately to agent-to-agent (A2A) technologies. Intuit is part of the working group behind the A2A specification, which could help create an open banking environment in the US. similar to the one already existing in the UK.
"In the US, we're still lagging," Wan admitted. "But if the critical mass of fintech players adopt A2A, that's a faster way to create an open banking-like environment. Technology isn't the barrier, consensus is."
Agus also cautioned against hype:
"There is no such thing as a fully autonomous agent. Many agents are essentially glorified workflow engines. The truth is somewhere in between. But when you mix autonomy with real utility, like QuickBooks' bookkeeping or payments agents, that's when it gets interesting."
The Bottom Line
Credit Karma's scale is staggering, with more than 60 billion predictions made daily across millions of members spread across various markets, and an ever-expanding set of AI-driven services available. But what stands out in speaking with Wan Agus is that scale alone does not define success.
What defines it is the balance between speed and trust, autonomy and oversight, innovation and responsibility.
FAQs
Why is balancing speed and trust important in fintech AI?
In fintech, AI decisions impact people’s financial lives. Moving fast without trust can lead to privacy risks and harmful recommendations, so balance is essential.
What are agent-to-agent (A2A) technologies in fintech?
A2A technologies allow secure, automated data sharing between financial services, paving the way for open banking in the US, like that in the UK.