Contributions (in alphabetical order): Arissa Yoshida, Arthur Scardua, Cinthia Tanaka, Edesio Alcobaça, Felipe Almeida, Hiroto Udagawa, Isabela Piccinini,  Matheus Fagundes and Victor Goraieb.

At Nubank, innovation, collaboration, and challenging the status quo are at the heart of everything we do. Purple MinDS is a blog series highlighting the unique challenges and experiences faced by data scientists and machine learning engineers at Nubank, capturing our dynamic and creative work environment. Recently, we hosted our first company-wide hackathon focused on Large Language Models (LLMs). 

Over two exciting days, teams across various chapters—including Business Analysts, Engineers, and Product Managers—identified business problems, proposed innovative generative AI solutions, and pitched their ideas for the top prize. For Nubank, Artificial Intelligence is one of the core foundations of the business; we are always working with cutting-edge technology.

Remarkably, of the 114 participants, engineers outnumbered data scientists, highlighting the widespread enthusiasm for generative AI and underscoring our commitment to building strong and diverse teams. The organizers, who are from different Business Units and include Data Scientists and Machine Learning Engineers (Arthur Scardua, Edesio Alcobaça, Isabela Piccinini, and Matheus Fagundes), shared an inside look at what they learned from organizing this exciting event and how it’s influencing our AI-first approach

Q: What was the hackathon’s main goal?

A: First of all, to have fun! But also, we wanted to democratize the use of generative AI across several teams. The way we see it, the hackathon was a way to plant a “knowledge seed” that would last even after the hackathon finished. We hoped the teams would start to identify when to use and how to adopt LLM solutions inside their domains. Generative AI is a new technology that can do so much – text completion, images, and voice recognition – and there are also agents! Everyone was curious about it, and they were waiting for an opportunity to build a project using this technology.

This is one of the reasons why we provided masterclasses about LLM before the hackathon. The masterclasses aimed to spread knowledge through demos and other means, providing an overview of the platform, while the hackathon was an excellent hands-on activity. We think the two complemented each other well. The masterclasses introduced concepts along with practical use cases and generated excitement for the hackathon, helping to increase the event’s reach, especially with prominent names like OpenAI bolstering its appeal. The masterclass attracted a bigger audience because it required less commitment than participating in the hackathon. Even for those who only attended the masterclasses, the knowledge gained will likely continue to grow and influence their understanding of LLMs.

It’s impressive because, although it feels like generative AI has been around forever, it’s still quite new. Not everyone has the privilege of a Ph.D. or specialized study in generative AI. While many of us know how to use tools like ChatGPT, truly understanding the underlying processes is complex. The masterclass was essential for educating people on the power of generative AI, and reflects our commitment to transparency and the importance we place on respect for our people by equipping them with the right knowledge and resources. With OpenAI being the leading company in the field, it added significant value and drew more attention to the hackathon. More people from the chapter became interested in participating in the hackathon because of this.

Check our job opportunities

Q: How did you come up with the idea for the hackathon?

A: In the past, we had an initiative inside the chapter focused on learning in which we organized mini-hackathons. We received good feedback about it, and people were truly engaged with it! Thinking about spreading the knowledge within our chapter to other people (regardless of their chapters), we came up with the idea of creating this big company-wide hackathon about LLMs as a kickstart to boost AI-First strategy. 

This strongly resonates with our value of inspiring purpose, reinforcing our ambition to continuously reinvent and make history by placing innovation at the core of our strategies. We thought of it as a way to enable people to generate value by creating quick prototypes for solving their business problems.

We chose LLMs because it is a hot topic and, plus, we have a great infrastructure to support it. For example, we have several tools built for internal usage and there is no limitation regarding number of tokens or number of requests.

Q: How was the process of organizing the hackathon and masterclasses?

A: Sponsors played an essential role, providing crucial support and technical assistance, particularly from the LLM Platform and individuals like Bruno Simioni [Director of LLM Platform] and Henrique Lopes [Data Science Function Lead], whose contributions were indispensable to the hackathon’s success. Originally planned as a small event, the hackathon evolved into a company-wide initiative, thanks to suggestions from Henrique Lopes, so everyone from the company could benefit from it.

The expansion required careful planning to mitigate any disruptions to company operations. When you invite the entire company, you must consider things like efficiency impact, because you are going to take some time from software engineers, data scientists, etc., so part of the company stops working for doing the hackathon. Without M-Team sponsorship, for instance, it wouldn’t have been possible. One of the challenges we had was that initially we thought a three-day hackathon would be ideal, but leadership pushed for a shorter one to avoid halting company operations, so we settled at a two-day event.

Q: What were the main challenges during the Hackathon execution itself?

A: There were several challenges, but we were able to learn a lot from them. For starters, scheduling an event on a date that suits everyone involved is impossible. There are many issues such as holidays, availability of external speakers, vacations, leaves of absence, proximity to other important events, etc.

During the hackathon itself, we learned that it’s hard to manage a large number of people in an online event. We need a large organization team to troubleshoot tooling problems, access to files and assets etc. The time we set apart for the hackathon was quite short and unfortunately, we had to skip some parts we had previously planned.

At the final stage, when attendees presented the results to a panel of experts, we also had some additional challenges and learnings: it was hard to compare peoples’ projects with each other, as projects had very different scopes, levels of complexity and completeness — some were rough research pieces and others were fully working prototypes.

Q:  What would you change if you had to do it over again?

A: Although the event was largely a success – as measured by feedback from the participants – there are some things we’d do differently if we had to do it over again:

  • More people in the organization: The organization team had around 5 people. Although most problems were solved and the team worked hard at those, having a few more people would help us navigate some times where there were multiple problems happening at the same time.
  • Additional days for the hackathon: We only had 1 ½ days for the actual hands-on part of the hackathon. We think at least 2 full days are needed for people to experiment a little bit and have enough time to build a working prototype.
  • Enable solo registrations: Some people wanted to participate but didn’t find partners who were interested in the same topics. We could have made arrangements for solo teams to take part in the hackathon and in the contest.
  • Fairer competition rules: At the end of the hackathon, participant groups presented their results to a panel of judges. Having stricter rules as to what constitutes a working prototype and adding some limits to the team sizes would help make the competition among teams fairer and would make it easier to compare projects among each other.
  • Require working prototypes: We should make it clearer at the beginning of the hackathon the criteria of what it means to have a “working prototype” at the end, to increase the likelihood that working products come out of the hackathon.

Q: And finally, which solution won the hackathon? How did this solution stand out from others?

A: In the final presentation, teams had 5 minutes to present their solutions and were evaluated on 3 dimensions: business impact, innovation and feasibility. The winning solution excelled in these 3 dimensions by proposing an AI agent to scale product negotiations like loans, late debt or insurance, for example.

The winning team of 5 people built a quick proof-of-concept that leverages AI to improve customer service using a Jupyter Notebook, LangGraph and the React framework. By using these tools, their agent was empowered to consult customer information and act independently to provide a human-like and personalized experience. The highlight of the solution was how it could improve our customer service in a seamless, scalable and personalized way.

Other interesting solutions:

  • Early identification of fraud crises.
  • Metric suggestion for randomized control trials and A/B tests on our experimentation platform.
  • AI-enabled parser for legal PDF documents.
  • AI answers in the in-app Help Center.

Conclusion

The hackathon was a great way for spreading knowledge about LLMs, reinforcing Nubank’s commitment to continuous learning and innovation. 

We see many teams evolving their projects into tangible products designed to impact our customers’ lives. Even teams that didn’t directly participate are leveraging the hackathon materials, embodying our values like challenging the status quo, acting as owners, and striving for smart efficiency. 

This collaborative spirit and shared purpose not only drive innovation but also foster an environment where diverse teams can truly make the extraordinary happen.

Check our job opportunities