In The Arena by TechArena - Anusha Nerella of State Street: AI Infrastructure for FinTech

Episode Date: July 2, 2025

Anusha Nerella of State Street Bank shares how AI copilots, LLMs, and agentic computing are transforming financial services—securely, responsibly, and at scale.

Welcome to the Tech Arena featuring authentic discussions between tech's leading innovators and our host, Allison Klein. Now let's step into the arena. Welcome in the arena. My name is Allison Klein and I am so delighted to be joined today by Anusha Narela, Senior Software Engineer at State Street Bank. Welcome to the program, Anusha. I'm so glad to have you on the show.

Thank you, Anusha, for the warm welcome. I'm feeling excited to be here. So this interview is part of a series of interviews that we're doing ahead of AI Inverse Summit coming up in September. What a fantastic event in lineup. It's offering for everyone. You will be on hand as a speaker at the event. Is that correct? Yes. Let's dive into it then. You know, I think that the reason why you're speaking is that your organization has leaned into AI adoption. What ways do you see the financial sector utilizing AI today?

I would say we are still in the explorative phase. And currently we are operating in such a way that everyone to lean towards in understanding and migrating towards with parallel learning. And we introduced to co-piloting at various phases right now so that we can reduce the manual efforts and increase the efficiencies of whoever are working in the part of the organization. So we are providing enterprise level licenses for the people to understand and adapt

to the AI based tools and frameworks so that they can come up with framework utilization and how they are effectively utilizing the frameworks to enhance our infrastructure level or at the development level. So this way, we would be able to embed strategically across various components, and also we are streamlining the process in the form of trade flows or in the process of compliance.

And we will be abiding by the regulatory standards for sure. But we want to ensure that we do not want to lose the client trust just by not following the things. So that's why we are trying to align with the standards for sure. With the maturity of large language models and enterprises looking at large language model integration, how do you see these technologies enhancing AI-driven solutions in the financial sector?

Yeah, I would say they are benefiting in such a huge manner because LLMs work in various ways. Like there are local LLMs and LLMs that doesn't require Internet to be utilized. So right now, Fintechs require a majority of local LLMs because for the development sake, developments require a majority of local LLMs to be beneficially utilized and we act on complex data. And as we know, when we are in the fintech sector, every minute to minute, we deal with petabytes of data

and billions and trillions of dollars in trades or in the consumer banking, in the retail banking. So there are various components that we deal with. And I would say it is very crucial complex data that we deal with. And we are actively exploring the deployment of localized NLMs and we are in the early stage of AI co-pilots.

Like I said earlier, we are trying to assist internal teams in order to reduce manual efforts. And we have pretty much difficult documentation to understand. So we are ensuring the agents do understand the documentation first in order to perform the task in a seamless manner. So the inputs or the prompts that we try to provide should be understood by the agents. So while these initiatives are in introductory phase,

so the impact of productivity and tradition making already they're looking promising. We are having a domain specific adaptation and secure deployment are the key focus areas that we are looking into. Now, I know that agentic computing has also gotten a lot of attention and is actually getting some deployment. How has this differed from other AI utilization? Yeah, it's more of the differentiation always has to be between how reactive and proactive they are. So, the agent computing always shifts between these two. So, I would say like across the index base, firms are adopting multi-level AI agents and channels, whether it is for the client

support or in order to escalate to the operational management. There are many traditional models that are present, I would say. Along with the traditional models, we are trying to adopt these agentic systems in order to make context-based decisions. We cannot go high level, so we always have to follow the proper context. So the context-based decisions, and we have to learn from what outcomes that we have received.

So always lean towards the outcomes and also act independently. So within what the boundaries that we have defined, we have to act accordingly. This represents, I would say, like a significant leap towards self-optimizing finance systems. And I also think that it's a significant leap in terms of giving AI the ability to act in multi-steps on data versus providing data that a human would act on. Can you talk a little bit about some of the key challenges you face when looking at deploying

AI agents in a financial service context? Yeah, sure. Because it's more about domain, even in fintech. So we have a specific domain expertise that is required here. Domain knowledge plays a major role in order to operate anything, whether it is a standard workflow, complex workflow, whichever it is. The rules we define for our domain plays a major role, and we have to feed into the systems the same thing.

So any automated processes or agents, they need to understand the rules, the domain expertise. So integrating with legacy systems like us, because we are very legacy, I would say. And learning with these kinds of agents and with these kinds of legacy data, they also have to abide by the regulatory standards along with this alignment is critical. So these are the challenges that we are trying to face,

and we are trying to get better in making this seamless. So I would say, agents must not only perform accurately, but also they have to explain the decisions transparently. So I would say model transparency should be there, just like how we explain a way that we use with AI. There is an explainable model transparency present in the AI framework that has to be present and it has to be transparent according to the domain.

So I would say governance, traceability, and ethical oversight, these are the essential factors in our mission critical environments like us. Now, when you describe that, one thing that comes to mind is clearance, because I know that in the financial service industry, given the tight control on security, that clearance is really important. Human employees, does that extend to agents,

and how do you actually manage that from a standpoint of ensuring that agents are prepared for their roles? Yeah, definitely. I would definitely agree to the point. In order to embed or incorporate any new things into the system, we need so many clearances. That plays a major role. So we will definitely get the clearances for the AI agents because we will explain to the organization in such a way that they act like human analysts and they reduce the human efforts in simulations. Like in the structured manner, the structured simulations plays a major role that can be

done by the AI agents. So we limit the clearance levels because we don't want to explore and expand in the vast manner because we wanted to experience them in the bounded level. Like, we don't want to expand the clearance levels to enlarged version because considering the real-world validations and the scenarios that we are facing right now, so much fraudulence is happening and we don't want to face that in the initial phase itself. So those kind of fraudulent scenarios we need to feed into the system saying that these are the patterns and these kinds of predictions our agents will

be doing even in the human absence. So that kind of data and scenarios that we are trying to feed, proper anomaly detection will be done by the agents. What I would say is context prediction and what kind of prompts that we are feeding into the system are they domain specific or not. So we will make sure our agents are prepared for managing all these roles in the progressive manner. Even though it is automated, that has to be in the progressive manner.

So we have to think about the broader scope when we are trying to embed something new and something we are trying to do is a supervised learning, giving a limited controls to the agents and giving limited rules and responsibilities. So the behavioral monitoring will be happening by giving these controls and we will ensure the auditing mechanisms are ensuring the accountability and the compliance at every step of the process that we are trying to achieve. Now, you mentioned governance, so I need to ask you about it. There's been so much attention on data governance of late.

What do you think the important steps are to ensure that the automated systems are operating within a context of governance? That is foundational because without governance, we cannot bring anything into picture. Any fintech sector, we are subjected to the Fed standards. So any fintechs are establishing this ethics board. Every quarterly, we get some kind of trainings to the employees and we give them to follow these ethics and we have to follow this ethics board that is in place so that nobody, whether it is knowingly or unknowingly, nobody will fall into that picture where they did not follow the rules or they did not follow the ethics and they fall under the flag structure.

So we have to ensure that trust scores and tracking all the model decisions are in real time. That's why we use zero trust infrastructure here. Like any other fintech space, we have to maintain clear logs for the compliance. Otherwise, our system will be flagged and we will be worried and keep answering to the organization saying that our AI systems are not functioning well. We don't want to end up in such situations, right? We wanted to ensure those function well and they function responsibly and securely. Now we are going to AI Info Summit, so I need to talk to you about infrastructure. As you look forward, what role does infrastructure play in making these advancements possible?

And what do you need from the underlying systems in terms of innovation from the industry to move forward effectively? Yeah, that's a great question, actually. So I would always think a step ahead as we are still learning and evolving towards AI. I would say scalability and resilience is the first and foremost thing. Everyone gives all their efforts to make the application reliable and scalable, and they are most production ready. So we are trying to ensure, as we are following the event-driven mechanisms, our infrastructure or the applications are more of multi-cloud and they're even driven. So what we need to ensure is our environments

has to be GPU accelerated so that they are production ready. And even in the production ready also, we need AI assisted production ready environments so that we can ensure the real time inference and also the high throughput and compliance even in the production-ready environments. So, Fintech needs this kind of infrastructure that can handle both the velocity and verifiability across the domain. Otherwise, how we can handle this kind of complex environments and infrastructure

ensuring the client's trust. All right. So we talked that you're speaking at the conference before. What are you hoping to see at the summit? And how do you think the conversations there will shape the future of AI infrastructure? Yeah, I'm literally eager to hear about more of orchestration around this agentic involvement and also LLM observability. The observability plays a major role, and how that is being done with LLMs is the part that I really want to hear about. And the conversations,

I know that it's kind of a powerhouse for sure, and I really want to be part of it and get excited to hear about all sorts of conversations and all perspectives towards the next gen of this evolvement. So how the scaling is happening and whether it is responsibly happening and how responsibly this is being done. That is the major part that I wanted to see towards this effort and compliance and regulatory, how they are doing it with agentic systems, how they are trying to achieve it.

So these kinds of conversations I'm looking forward to. That's awesome. So it was so wonderful talking to you, Anusha. The fact that you're really on the forefront of this massive innovation for financial services, I can't wait to hear more from you. I am excited to attend your session at AI Infra. I'm sure that folks who are listening online today also want to

connect with you. Where can they reach out and connect and continue the dialogue? Yeah, I would always be available on LinkedIn and I'm also a part of Hope Technology Council. So I submit articles and I ensure my thought leadership runs towards publishing these kind of articles and answering expert panel queries. So I'm available on various mediums like ADP as well for continuous mentorship. So yeah, people can reach out to me through any of these mediums. I would always be available and I would encourage them to interact in such a way so that the thoughts can be shared. Awesome. Thank you so much.

It's been a pleasure talking to you. I can't wait to see you in September. Thank you. Me too. I love to be there. And, yeah, excited to hear about that powerhouse conversation. Thanks for joining the Tech Arena. Subscribe and engage at our website, thetecharena.net. All content is copyrighted by the Tech Arena.