UChicago MS-ADS Students Present Capstone Project at Midwest ML Symposium

When Ariel Azria, Kwaku Ofori-Atta, and Nathan Rickert walked into the 2025 Midwest Machine Learning Symposium on June 24th, they weren’t just attending. The three graduate students from the University of Chicago’s Master of Science in Applied Data Science program had been selected to share their capstone project, FinWise.AI, during the poster session.

“It was awesome,” said Azria of the experience. “We put in a lot of research, and getting to present it in a space full of other machine learning researchers was phenomenal. It really felt like the hard work paid off.”

How It Started

But the journey to the symposium started months earlier, with an idea that looked quite different.

Their first concept sounded like something out of Silicon Valley, a tool that could automatically manage someone’s finances, make smart decisions, and grow wealth with minimal user effort. But it was their capstone advisor, Justin Kurland, who steered them toward something more original. What was missing in the space, they realized, was a tool focused not on maximizing profits for the wealthy but on offering clarity, control, and trust for the people often left out of traditional banking systems.

“We wanted to focus on a tool that matched our market opportunity,” said Rickert. “That was originally scoped during the preliminary phases, understanding our market segments and where there’s opportunity for the tool. That’s where we gravitated towards and framed the entire tool.”

The result was FinWise.AI, a machine learning-powered platform designed to help unbanked and underbanked individuals navigate banking offers with confidence. The unbanked refer to people who do not have a checking or savings account, while the underbanked may have basic accounts but still rely on alternative financial services, like payday loans or check-cashing services, that often come with high fees and limited consumer protections.

Optimizing for Financial Rewards

The goal of FinWise.AI was to empower users to make banking decisions that felt both transparent and worthwhile. Many unbanked and underbanked individuals avoid traditional financial institutions due to a lack of trust, shaped by years of complex terms, hidden fees, and opaque practices.

 “We wanted to come up with a solution that helped people feel like they were in control of their financial situation… in a way that isn’t harmful to them, in a way that’s beneficial,” Azria said.

To address that, the team built a preprocessing system that could quantify dense legal language such as costs, fees, restrictions, and regulatory red tape, and simplify it. Then came the challenge of optimization: developing an algorithm that could evaluate hundreds of constantly changing banking offers and recommend those where the potential benefits outweighed the risks. 

That’s where their three optimization methods came in. Rickert started with A*, a classic pathfinding algorithm, to set a baseline for how a user might move through the “space” of banking decisions. A* helps identify the best combination of bank offers while considering deposit limits, lock-up periods, and user eligibility rules. But A* had limitations. It struggled to generalize across different income levels, especially for low- and high-earners. 

Azria and Ofori-Atta then implemented two reinforcement learning approaches: Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO). These methods allowed the model to simulate thousands of decision paths over time, adjusting strategies based on changing offers, income constraints, and long-term goals. 

“The reason we chose reinforcement learning is because the amount of offers available at any one time is changing,” said Ofori-Atta. “You have a dynamic state space, so traditional optimization techniques don’t really work. You need to dynamically optimize changing conditions.”

Ultimately, PPO emerged as the most consistent and equitable method, delivering stable performance for users across all income brackets.

From Theory to Recognition

The invitation to present at the Midwest Machine Learning Symposium, hosted at the University of Chicago’s Logan Center for the Arts, was a milestone for the team. Their project, which had begun as a classroom assignment, now stood alongside work from researchers across the region.

FinWise.AI also earned an Honorable Mention in the Winter 2025 Capstone Showcase, a recognition given to standout projects that demonstrated innovation, technical strength, and real-world potential.

Rickert, who had followed the symposium for years, said it was a natural next step. “We were proud of what we’d built, and the poster session gave us a chance to share ideas with other scholars and see where our work fit in the broader machine learning community.”

Beyond the Capstone

For Ofori-Atta, the experience has shaped how he thinks about his work. “This project has shown me that there’s a lot of space for AI and optimization problems,” Ofori-Atta said. “I have a lot of experience teaching financial literacy to kids on the Southwest Side of Chicago, and what I learned from that is that it’s very difficult to reason about money—it’s emotional, and it’s also asymmetric. It’s hard to understand what you know versus what others know. So having AI level the playing field in that way is something I want to keep working on.”