Upstart

We know there's no substitute for connecting in person, so we complement Upstart's digital-first approach with intentional gatherings. We love spending time together, so we prioritize company on-sites. In 2024, we kicked off the year by bringing Upstarters to Austin for what we call Upstart Live.

We reconnected with our mission, spent time with fellow Upstarters and looked ahead to a year of innovation. Our offices also connect our employees. Upstart offers quarterly, in-person orientations at our office in San Mateo, California.

The two-day program includes time with our leadership team, product overviews and more. We empower leaders to decide how and when to bring their teams together. This could be a structured onsite for strategy sessions and team-building, coworking with local Upstarters, or gathering to solve a tough problem.

Our fantastic office management team supports it all. We also make a point to connect remotely. On Fridays, we host a TGIF all-hands meeting to hear updates, celebrate wins, and engage with our leadership team during Q&A.

Our nine employee resource groups offer remote and in-person programming. These communities enable Upstarters to build connections with each other that help us work better together.

When we made the decision to be digital first, a key driver was access to a national talent pool and the flexibility to allow people to work in the way that allows them to be their most productive. Finding talent with the right skills and experience, regardless of location, provides opportunities to candidates from various socioeconomic backgrounds who may not be able to relocate for a job.

It also creates opportunities for people with disabilities who might find commuting or working in a physical office challenging.

At Upstart, ML and AI models are the heart of our product. In fact, you could argue that these models are our product. For example, our underwriting models drive the loan offers we provide on our platform and are designed to evaluate the credit risk of any given applicant.

That is, they answer the question, "What is the likelihood this applicant pays back the loan?" For our product to succeed, we must do this with a high degree of accuracy. To achieve this, our underwriting models utilize state-of-the-art ML techniques trained using a vast amount of payment data and alternative credit variables that many lenders overlook.

Beyond underwriting, we use ML models to optimize marketing spend and detect financial fraud. A major aspect of our product development is working on improving these models. This can take the form of integrating new data sources or developing new modeling techniques.

We continuously update our models to keep up with economic trends and improve performance. The process involves ML researchers, ML engineers, risk reviewers and, of course, product managers. This results in higher approval rates and a seamless experience, with funds often reaching customers in a few days.

First, we hire bright and highly motivated people. One of our core values is "Be smart, but know you might be wrong." Our people are encouraged to try new things, such as new modeling techniques, implementing new training infrastructures and thinking outside of the box about how the product works.

This encourages an environment of continual learning that allows us to keep our team's knowledge at the forefront of AI and ML. Next, we're provided with access to the best computational resources. Modern ML and AI can be constrained by the amount of compute required to implement it, both in training the models and serving them in production.

This requires investing in compute resources to allow for fundamental research that strives to continuously improve our models. Finally, we have a dedicated set of engineers and risk professionals who support our AI and ML researchers in building robust frameworks for training, maintaining and deploying models in a cost-effective manner.

Upstart takes a unique approach to underwriting where we rely primarily on sophisticated ML models to make decisions on how to price loans. Alternative approaches include using a human to manually review credit applications and apply predefined rules or applying a more traditional statistical model to assess credit risk.

The first approach is clearly costly since it directly involves a trained person to review each loan application. It's simply not scalable. The second approach, while closer to ours, doesn't achieve the accuracy required to meaningfully expand access to credit to underserved consumers.

By utilizing better, more accurate models, we can approve more consumers at lower rates all while maintaining loan performance.