Contributions: Henrique Lopes, Rohan Ramanath, Vitor Olivier, Hiroto Udagawa, Ariel Fontes, Arissa Yoshida, Cinthia Tanaka, Felipe Almeida, Pamela Perez

At a recent internal event for Nubank’s data science team, Henrique Lopes, Rohan Ramanath, and Vitor Olivier shared their perspectives on how artificial intelligence (AI) is reshaping the way we build products, organize teams, and make decisions at scale.

The conversation touched on a wide range of topics, from the practical challenges of moving from prototypes to production systems, to the growing importance of platformization, and the evolving role of Data Scientists and Machine Learning Engineers in this new context.

Lopes leads Nubank’s Data Science function and Data Products business area. Ramanath heads AI Core, the team born from the Hyperplane acquisition, focused on the intersection of AI, infrastructure and modeling. Olivier, Nubank’s CTO, brings over a decade of experience driving the company’s technology strategy and building systems that scale.

The exchange is part of Purple MinDS, a series that explores how our data teams act to do their best in the challenging and high-opportunity scenario of artificial intelligence. 

How do you use AI in your routine?

Lopes: I use gen-AI tools for many random questions, often replacing commonly used search engines. On the job, AI has been powerful for coding between meetings, especially with the recent launch of agent-based coding platforms, which has helped me do more testing and proof-of-concepts.

Ramanath: I use AI primarily for planning and summarization. I use agent-based coding platforms to understand codebases written by others, treating it like asking a teammate to walk me through the code. I also use AI for travel planning and ideation, helping to refine vague ideas, articulate them, and find technical terms for searching related literature.

Olivier: My daily use of AI is similar to what Lopes and Ramanath described; I mainly use it for search, to interact with codebases and do proof-of-concepts.

What human task do you want to be completely replaced by AI in a few years’ time?

Lopes: While current AI usage is impressive, it falls short quickly. For tasks like interacting with code, it often gets stuck in “rabbit holes” as context increases, sometimes requiring a restart. I tried to compile something, and the tool quickly started changing my computer’s Java configuration without solving the problem, going deeper and deeper without success. This is improving, but it still requires significant human intervention to pull it out of unhelpful loops. I think this reliability issue needs more development for large-scale production use.

Ramanath: I want to see a shift from AI providers charging by usage (licenses or API calls) to charging by outcome. I’d like to see a world where companies charge based on successfully achieving the end state, such as paying a provider for every library successfully migrated. Currently, using AI tools still requires significant mental engagement. There’s initial novelty and productivity gains, but the happiness from good suggestions is a decreasing function of time; it becomes just another tool. 

I’d like to see AI reach a state where, for simple tasks like applying for a visa, I can ask what documents are needed, get an answer instantly, and not have to verify the output. Today, recommendations are fast, but I still spend significant time verifying the information. 

Olivier: I’m more excited about creating systems that know how to use AI effectively to achieve outcomes, rather than just expecting AI to become more intelligent on its own. For example, home renovation involves hundreds of payments to different places in various ways. I wish I could send all payment requests to one spot and have them paid automatically. I don’t think there’s any technical barrier to this; it’s about processing unstructured data at scale and transforming it into a useful process through “plumbing” and engineering. Even if AI development stopped now, there’s still a decade of features that could be built on top of existing AI by focusing on product management and good “plumbing”.

Ramanath (following up to Olivier): While AI can easily list plumbers or electricians, the human still has to call each one, find out their availability, prices, and compare offerings. AI companies claim “tool calling” is solved so I don’t know why we don’t yet have the experience where we can call all these people, get feedback, provide a comparison, and facilitate a decision. Orchestrating this reliably to reach a final decision is a “magical product experience” that is still not built.

What role will AI play in the future of Fintech?

Olivier: I see AI as just another tool for the core promise of fintech and the internet: more access, better quality, and lower price for everything. AI adds to the toolkit to achieve this. This includes achieving very fast, high-quality response times in customer service, enabling true hyper-personalization beyond discrete segmentations, and in the investment space, helping identify user patterns and risk profiles to provide confidence in decision-making. I think AI will permeate everything in fintech, just like mobile did, becoming a given capability that fundamentally improves the user experience with higher quality, lower cost, and more accessibility.

Ramanath: I’m excited about newer interfaces enabled by AI. With good engineering abstractions for banking concepts (like accounts, cards, transactions), AI can enable users who aren’t experts to create apps. I see a future where “personal software is free,” allowing consumers to customize their banking app interface. Instead of companies creating many interfaces, consumers could pick and choose how their banking app interacts with them, whether they prefer detailed expense breakdowns, high-level overviews, or interaction via voice command. This would solve the problem of discovery, allowing each customer to design their ideal interface without the company having to push different versions. I believe technology will drive the evolution of these newer experiences and interfaces.

With AI evolving so quickly, how can Data Scientists and ML Engineers stay ahead and remain relevant?

Lopes: I’d advise against trying to keep up with everything happening due to the fear of missing out, as there is too much information. While being an early adopter shows innovation, you also see many things that don’t work. My recommendation is to focus on areas relevant to your interests and work, such as data engineering, MLOps and applications that align with your background. There will be many sub-areas of AI, and no one person will understand all of them deeply. The key is to focus and avoid the trap of trying to be an expert in everything.

Ramanath: Data science is a broad field. I suggest picking a niche (like model explainability, evaluation, MLOps, serving infrastructure, causal models) to have a better chance at staying close to the cutting edge. Specializing takes years to ramp up on literature. Individuals need to decide whether to be a generalist or a specialist; specializing increases the chance of keeping up with trends and predicting the industry’s direction. The goal is not just to use the coolest thing, but to build intuition to forecast future directions, enabling systems to be built such that adopting new technology has near-zero cost. 

It’s also important to listen to the right people. In the noisy AI landscape, identifying a few experts who are opinionated and have a track record of predicting the future in a specific niche helps cut through the noise. These experts might write about productionization challenges and the gap between “AI toys” and “AI systems”. Finally, nothing beats hands-on experience of playing around in free time.

What challenges does Nubank face in implementing AI across its operations?

Lopes: One major challenge is the leap from an “AI toy” (something promising) to an “AI system” (something that actually works at scale). Everything at Nubank is at scale. The difficulty lies in balancing risk-taking and dedicating time to make new AI systems work reliably while also maintaining traditional systems.

We are on a learning curve: what works for this model won’t necessarily work for the next one. Prompt engineering was the hot topic six months ago, but that doesn’t mean it will remain relevant as models evolve. Finding the difference between what is fundamental and what is specific to the moment is particularly challenging when putting things into production with AI.

How do you see the DS function interacting with other functions (engineers, product managers etc.) that are also introducing AI-based tools and workflows?

Lopes: While generative AI has a new shape, in terms of how it’s generating text, images and videos, the underlying concepts of how it’s built and the pitfalls (like leakage, data quality issues) are similar to the ones we face using ML. These pitfalls are still present in AI but often hidden behind a user-friendly interface. The data science function has a key role in helping the organization navigate these pitfalls, both when building models and when other functions are using AI tools. For less technical people, AI can feel like magic, and the data science team has the responsibility to show the underlying complexities. They can help ensure that AI usage is done quickly and bring innovation, but in a responsible way. 

As people who work in a field that didn’t exist in its current form recently, data scientists and ML engineers are on the front line of testing and understanding where AI works and doesn’t, playing a leading role in guiding others on the directions to be explored with AI.

Very few major AI-first financial institutions exist (if at all). Why is Nubank well-positioned to become one?

Olivier: Nubank is well-positioned because, since our earliest days, we recognized data science as a key competitive advantage. We intentionally invested in technology, became cloud-native, and built tools for data extraction, analytics, modeling and so on. We also invested in bringing in skilled people to the company. This 11-12 year bet has positioned us well to continue playing the “infinite game”, in which what we earn is the right to continue playing, and we earned it.

Building an AI-first financial institution is hard for everyone. Incumbents face difficulty migrating, while new entrants struggle with regulatory burdens, credit skills, distribution, and data volume. Still, we need to double down on AI because otherwise someone else might be able to figure out an angle to gain a competitive edge. While there are disadvantages to early adoption, I think adaptability and being anti-fragile are what makes us ultimately win. We are now in a platform shift so we will have to navigate through this change, and be well-positioned to capture it.

What’s your vision for the future of the Data Science function? Will roles like Data Scientists and ML Engineers start coaching other teams on how to use AI effectively?

Lopes: There are a lot of new names for things data scientists have been doing for a long time, but the job is also changing. We have new roles as AI engineers in many companies, but this is not what matters the most. I believe we will see a shift of software engineers, MLEs, and data scientists focusing on integrating AI systems and agents into their activities. This is more of an adaptation in how software is built rather than a fundamental change in the role itself.

Data will become more present in every system, flowing through transactional environments, without the historical separation between analytical and transactional environments. Being able to use this data at scale for live decisions will become more important, which might change a bit the roles of data engineers, software engineers, data scientists, and ML engineers.

Ramanath: I see an opportunity for the data science group to influence the “how”. While conversations across the company might focus on what new AI tools or vendors to use, the data science function understands how these models are trained, their assumptions, and the impact of violating those assumptions. There is an opportunity to educate others across the company on how to use AI tools, code using agent-based coding platforms, train models, implement systems like RAG, focusing on what questions might break assumptions on those systems. This can be done through examples, prototypes, and collaborations.

What parts of the DS/MLE workflow will AI substantially change and what will be harder to replace?

Lopes is excited about AI improving the “busy work” parts of coding and workflows.

Lopes: Think about it: tasks like getting AWS certificates, refreshing tokens, or running staging tests to check costs and performance. AI should be able to guide these workflows, automating those repetitive tasks that currently demand we remember specific commands.

This means data scientists can finally focus more on the core business problem. This doesn’t just mean AI writing code or queries in a streamlined way, but also automating validations, running unit tests, and even validating changes as part of the overall workflow. Anticipation is high for platform changes that will enhance the user experience by automating these kinds of tasks.

Ramanath complemented the discussion by sharing his beliefs on the aspects of the job that will be more challenging for AI to replace.

Ramanath: The part of the job that will likely be hardest for AI to substantially replace is formulating the problem that needs to be solved. Mapping a business problem into something an AI model can address requires judgment, intuition, and an understanding of the interactions between existing systems and humans.

While data scientists have traditionally dedicated significant time to implementing and rolling out solutions after defining the problem, AI productivity is expected to drastically reduce that implementation time, potentially from months to weeks.

Not just smarter, but more useful

The future of AI at Nubank will be defined by how we integrate AI technologies into scalable, reliable systems that deliver real value to our customers. As Lopes, Ramanath, and Olivier made clear, this shift requires more than technical innovation: it demands new ways of thinking about infrastructure, ownership, and collaboration across functions.

From automating repetitive workflows to redefining how we design products, AI is becoming a foundational layer in everything we build. And while the tools will keep evolving, one thing remains constant: the role of Data Scientists and ML Engineers as critical partners in building the Purple Future.

To learn more about how Nubank is advancing the field of artificial intelligence, explore our AI Research section, where we share ideas, experiments, and insights from the teams building tomorrow’s systems today.

Check our job opportunities