AI Leader Diego Represas on What’s Different About Digit
Four years after joining Digit as its first data practitioner, Diego Represas now helps lead the Data Science/AI team. Below, he explains what drew him to the 10-person company (now 80 people), discusses what’s different about Digit’s business model and vision, and shares some of the unique technical challenges the team is tackling with cutting-edge ML and AI.
For those who aren’t familiar, what does Digit do?
Our mission is to make financial health effortless for everyone. That starts with helping people save; Digit analyzes your spending and sets aside just the right amount, without making major changes to your lifestyle, thereby allowing you to focus on other important things. From there, we’re working toward a larger vision that includes automating all financial decisions, from how much to save for retirement to how much to pay off on a credit card bill. Right now, that kind of service is only available to those who can afford their own personal financial manager, which the vast majority of people can’t. By making decisions automatic, we can make financial health accessible to virtually anyone who has an internet connection. This is especially critical work considering that nearly half of Americans struggle to come up with $400 in an emergency.
Why did you join the team?
When I joined Digit four years ago, we were about 10 people. I was the first data practitioner on the team, which in and of itself was exciting because I knew realizing the company’s long-term vision was going to require a lot of AI and data science. But I was also drawn to the opportunity to work on a problem of such gravity. Financial well-being is the number-one source of stress for Americans, and it affects everything from physical health and life expectancy to community and family cohesion. So in terms of the problem we’re solving, it doesn’t get much more impactful. During my interview process, I got to see the types of data the team already had — including the strong customer retention curves, which showed an incredibly sticky product. It was clear people loved Digit. I decided I was in.
What sets Digit apart from other fintech companies?
One big difference is that our incentives are aligned with our customers’ incentives. We made a conscious decision that we are not interested in profiting from inefficiencies in the financial system at the expense of our members. This fact not only differentiates us from many fintech companies, it also serves as the reason we chose a flat subscription model. A lot of fintech companies make money by taking a percentage of the assets they manage, or by charging fees on each transfer. Instead, we charge a flat rate: $5 per month. That’s it. We design everything we do to provide value to our members.
“We made a conscious decision that we are not interested in profiting from inefficiencies in the financial system at the expense of our members.”
And while there are other companies that also want to make people more financially healthy, we’re different from them, too. Those systems tend to assume humans are perfectly rational — that all you need is the right information, and you’ll make the right financial decisions. We understand that personal finance is about behavior, not just education. Even if you give someone all the right information about what to do with their money, that doesn’t mean they’ll do it. Humans — and their lives — are more complex than that. But with Digit, financial responsibility happens automatically. This is a massive exercise in personalization technology, requiring some of the most complex solutions out there.
Give us an example of those technical challenges.
You can think of any artificial intelligence as a system that is managing four distinct processes — sensing, perception, decision-making, and acting. For example, a robot navigating a forest has to see a tree, understand what it is, decide how to avoid it, and then make the turn. Of course, the sensing phase is standard by now. And plenty of companies are digging into the perception phase, but very few do it as deeply as we do. Because financial behavior evolves over time, we engage in a particularly challenging flavor of ML called human-in-the-loop, active machine learning.
“We’re pursuing some of the most cutting-edge technology and research and applying it to our problems.”
The challenges inherent to this type of machine learning have made us develop some pretty impressive pieces of technology. One example is our dataset management platform called Bernard — named after the character from HBO’s Westworld — which allows us to “automagically” create and iterate on human-annotated machine learning datasets. Bernard creates new datasets, adjusts for quality, flags and excludes any suspicious annotations, and tests each new model against its predecessors with minimal intervention on our part. Beyond Bernard, we’re currently pursuing some of the most cutting-edge technology and research published in this field and applying it to our problems.
What about the next phase, decision-making? Tell us about the challenges there.
Planning and managing a real-world budget to achieve long-term objectives fits within what is commonly known as the reinforcement learning problem, which is an area of ML most engineering teams don’t get to tackle. It’s like a massive game of chess or Go — there are so many variables to consider; you’re trying to deliver exactly the right experience to the right person at the right time, and you’re catering to an almost infinite breadth of situations and goals. One member might have four dependents and want to buy a house; another is a college student just trying to save; the next is in their 60s and hoping to retire soon. Before you modify your algorithm, you need to know how it’s going to affect each of those people. You don’t want your change to make the experience better for one member but worse for another.
“Planning and managing a real-world budget to achieve long-term objectives fits within what is commonly known as the reinforcement learning problem, which is an area of ML most engineering teams don’t get to tackle.”
This problem represents the largest area for future growth of AI at Digit. Today, we have an in-house simulation system to help us iterate on different money-management policies or algorithms. However, one of our 2020 road map projects is to create a platform that will make the process interactive. Every time an engineer or product manager wants to make a change, they’ll first click “Play” and get a complete visualization of what would happen for all members, getting nearly real-time feedback on how their proposed changes will affect our members. No one else in consumer finance has anything like it, and it’s just one of the reasons I’m so excited about Digit’s future.
You’ve spent the last four years of your career at Digit. What do you think the next four years look like for you?
I get this question all the time! In terms of what my long-term future holds, I’ve wanted to be an entrepreneur since high school, so at some point I know I will take what I’ve learned at Digit and apply it to a new problem. But I’m not done learning here. This is an awesome place for someone who’s deeply technical and cares about making a positive impact on the world. The magnitude of our vision and the related technical challenges we need to solve in order to achieve our mission make me excited to come to work every single day. So does the honor of working alongside such an incredible and diverse team. And the strength of our culture here is palpable — so much so that in our most recent engagement survey, 100% of our employees would recommend Digit as a great place to work.