CDOIQ Symposium 2024 | Krishna Valluru | Fidelity Investments

theCUBE.net

Clips
News
More from CDOIQ Symposium 2024

Krishna Valluru

Vice President / Team Leader of Advanced Strategies & Research Tech

Fidelity Investments

Mastering data governance: Insights from Fidelity Investments on scaling and emerging technologies

The field of data and analytics is rapidly growing and evolving, requiring creativity, skill and a deep understanding of emerging technologies, particularly in data governance.Advanced strategies for scaling, data governance and the balance between art and science in data processing are at the forefront of current discussions. These innovations not only enhance the efficiency and accuracy of data-driven decisions, but also present unique challenges in maintaining data quality and governance across diverse and unstructured data formats, according to Krishna Valluru (pictured), vice president and team leader of advanced strategies and research tech at Fidelity Investments

play_circle_outline Navigating AI's Impact on Data Standards and Challenges in Unstructured Data Scaling

play_circle_outline Maximizing Impact: Data Quality, Governance, and Entity Resolution in Unstructured Data

play_circle_outline Exploring Quantum Computing and the Evolving Landscape of Coding Languages for the Future

Info
Transcript

Krishna Valluru | Fidelity Investments

Krishna Valluru

Vice President / Team Leader of Advanced Strategies & Research Tech Fidelity Investments

TheCube's coverage of CDOIQ from Cambridge, Massachusetts, the 18th annual CDOIQ Conference. Sanjeev Mohan, co-host, with Krishna Velaru, Vice President and Team Leader of Advanced Strategies and Research Technologies at Fidelity Investments, discussed processing alternative data sets. They discussed the challenge of handling unstructured data and the importance of scalability. They touched on the impact of AI on scaling and the importance of data quality for unstructured data. The conversation delved into entity resolution and the challenges of privacy and g... Read more

explore Keep Exploring

What is the biggest challenge when it comes to handling unstructured data and scaling in the past couple of decades? add

What is the challenge in scaling solutions to different business use cases when dealing with unstructured data that may be in various formats and expressions? add

What statements did Jensen make that raised controversy and skepticism? add

bolt Powered by CUBE AI

Krishna Valluru | Fidelity Investments

search

>> Hi everybody, we're wrapping up day one here, theCube's coverage of CDOIQ from Cambridge, Massachusetts, the 18th annual CDOIQ Conference. It started with in Washington DC, in Crystal City, with I think about 29, maybe 30 people, up to well over 700 today. I'm here with Sanjeev Mohan, my co-host, and we're really excited to have Krishna Velaru, who's the Vice President and Team Leader of Advanced Strategies and Research Technologies at Fidelity Investments, local company. Heard of Fidelity before, so thank you so much for coming on theCube. I know you have some disclaimers you'd like to get out beforehand, Krishna? so please do so.

>> Yes. First, thank you for having me. And so these are my views only, and not Fidelity views, just I want to get it out of the way before we start, actually. Thank you for that.

>> Yeah, Fidelity, very advanced company. A lot of secret sauce that we're not going to be able to talk about, but we can generally talk about what's happening at this conference, maybe you start with your role at Fidelity. What's your purview?

>> So I lead advanced strategies and research technology team, and my team's responsibility is to process alternative data sets, and to combine with traditional data set, and provide the data ready for our analysts to process and run AI engines on it, so that they can make some investment .

>> You've been putting Fidelity for decades, yes?

>> Yeah, for 23 years.

>> So when you start... For how many years?

>> 23 years.

>> 23 years. So when you started your career, it was a lot more structured data that you were analyzing...

>> true....

>> than to probably using a lot of spreadsheets, probably still do use a lot of spreadsheets.

>> Yes, still.

>> But the amount of unstructured data has just been immense. How do you think about that? How do you handle that? Any thoughts from an advanced strategy standpoint?

>> Yeah, no, I think definitely. I think for the past decades, a couple of decades now, actually it is increasing, increasing a lot, now as we understand. And handling the unstructured is our primary responsibility. And the scale is the biggest problem. So out of that, how do you scale? You can experiment, you can do certain value. Like if you have a business use case, you can deliver the use case. However, the scaling is a biggest challenge for us. That is what we are focusing on. So first we start looking into how should we have a proper governance, proper, the quality. And then, can we scale or not? And then we can look into the use cases and then deliver that product.

>> And scaling today is not just simply throwing it in the cloud, and throwing compute and storage resources at it. The AI brings a new dimension to scaling, does it not?

>> It does. So AI definitely will help, but at the same time, it is more of an art. How do you cook that recipe, versus you have all the ingredients of, you have tools, you have data, you have unlimited of everything, you have in the cloud, we can have unlimited resources. In the data as well, you can have unlimited of data of anything. But how do you cook that recipe is a key. So for that, it is definitely AI is an aspect. Plus also, it is not just a science, but also, I say is an art as well.

>> Interesting.

>> It's interesting, because I've known Krishna for a couple of years now. I see him at CDOIQ. And the work you're doing is outstanding, but of course, we don't have permission to talk about it. So staying at a general level, when you work with unstructured data, you mentioned data quality becomes really key. How do you define data quality for unstructured data? It's very different from structured data, right? So what are the metrics?

>> Correct. And it is all fit for purpose, like depending on the use case, what is a fit for purpose? And we look for what is the business use case starts from there, and we walk backwards on how we, what are the qualities we define on the fly.

>> I see.

>> So depending on the use case. And then we create some reusable components as well. And then, so that you can scale for other business use cases. But you are right, absolutely. Actually, it is a problem and the challenge. And at the same time, actually, you can solve a problem. But now, how can you scale to different, apply to different problems is challenging. But how do you, in the unstructured data, it can be in many different formats, and people may say in many different ways. For example, you have Tim Cook, and they can say, "Timothy Cook, or Apple CEO, how can you identify that person from that structured data? Is that person you are talking to?" That is a challenge. But at the same time, now with the GenAI, that can help with, depending, given the context of the AI, you can identify that, "Oh, Apple CEO is Tim Cook, Timothy Cook, or Tim Cook MBA, all are the same individuals." So that's how we can scale with the GenAI now. That is super exciting.

>> So this is an example of entity resolution?

>> Correct.

>> Which is a structured, more of a...

>> More of, from unstructured, you can extract this-

>> So extract the entities-

>> Yeah, correct....

>> and then you use AI. But what I really liked was, what he just said was that data quality is fit for purpose. So it's the same unstructured data, like a text document could have very different data quality aspects, depending upon what the use of that is.

>> Correct, yeah, exactly. Yeah, that's right.

>> But so, you take that entity resolution example, Tim, Timothy, Timmy Cook, who has an MBA, when you do that entity resolution, you're inferring from maybe some other unstructured data, some of that, much of that could be personal information. So you have to have to worry about privacy, and governance. And obviously, Fidelity, you're in a highly regulated industry. So that's a big challenge today, is it not?

>> It is a challenge. And so some of the data that you can classify them as public, so you have to understand, what is public data. So any company CEOs, any board members are public. And for example, if personally, if Sanjeev goes in and enters some comments, or some reviews in somewhere, if you're not a public figure, your identity can still be personal. And so we have to treat, who are the public figures, and what says, who are the private that we cannot, we have to make sure that governance is in place. So absolutely right, we can scale pretty fast, but we don't want to impact the governance or the policies. We should not violate them. So that's a challenge, as well as part of the delivering a product.

>> And that starts with the data classification.

>> True.

>> Is that right?

>> Yeah, true.

>> But what is the state of data classification today? I remember when the federal rules of civil procedure changed in 2006, and email archiving was a thing. The way you would classify data was with math. You'd use support vector machines, or probabilistic latent semantic indexing. These are well-known mathematical techniques. Is that the same way today? Or is AI kind of enhanced that mathematics? Or what's the state of data classification today?

>> There are a number of communities of practice, we can say in industry, that are open source now. So there are some fiber models, which are open, that you can use, but still they're not yet ready. So the new things are evolving. As they evolve, how do you make sure that you are up to speed on that, how your models can consume that latest information, or new rules? Every day we are hearing some new policies across the globe are applied. You have European, California, New Jersey and every other state are thinking about coming up new policies. How do you inject them into your models? Make sure you are adhering to the policies and governance, that's the biggest challenge.

>> And is part of your role just understanding what's coming down the pike, in terms of new technologies? Is that right? Or is it... Yes. Okay.

>> That's right. Emerging. We look at all the emerging ones. So that is the reason I'm here today, at the CDOIQ. When I'm looking at what is emerging? How can I change any strategies internally? And how can we use them in the advanced world? So from emerging, to the advanced, that's what I look for, and implement them. You have many vector databases, and others are evolving now, which are prominent.

>> When did LLMs first hit your radar? And then what did you think when you saw the AI heard around the world with Chat GPT?

>> No, it is interesting. Boston is a great community. You have, back in 2015, 2016 at Rakuten, they were in South Station. And they had a meetup at that point, and they were talking about GPTs generating an image with the text. And then it started at the time. And then we were watching that. See, I'm watching at least since then. And all of a sudden when the application came, so we don't know how to use it, until the Chat GPT came up. So Chat GPT really said that, "Oh, this is how you can use it." So extracting information from the text, or generating content, and that is where it changed now.

>> Did all the techies in your world just light up, and buzz, and pull all-nighters, and start to huddle, "How can we use this?"? Or what was that moment like?

>> I think it is... So you cannot do that in any enterprise, if you are-

>> Yeah, I was just thinking-

>> Yeah, you cannot do that. Yeah.

>> We have so many restrictions, like-

>> Yeah.

>> When I say light up, I mean just get excited, and start talking about it, and...

>> It is good to get excited to see how it is, and then slowly slow down, actually. So I would say that, definitely, we should get that. I think otherwise, we can't take any technology to that level. So number one. Number two is that what we look for is, I think at the enterprise we have some standards. You have to use that, make sure everything is stable, and it has to go through. It should be in a generally available state. Until then, we cannot use it. But at the same time, you can see how, where other technologies are emerging. And you can identify them a little bit faster, so that either you can use them in an operational, and not at the end user level, but at the operational side, any efficiencies you can improve, that's what we can look for.

>> Exciting times now. So did the introduction of large language models and GPTs, did it change? You mentioned you have standards. Do you have to rethink those standards or extend them in different ways?

>> Yeah, so personally, I don't work directly with LLMs. We have other teams who are working on it, but I support them, actually. At the same time, we have some, I think we should have the governance in place, and proper governance. So the proper guardrails, so that when they come up with some product, they are sure that the product, they can take it to market. At that time, they don't want to have surprises, that, oh, here is a policy violation, here is some governances there. And so, instead of that, it is with the proper guardrails, then they can take that to production.

>> Okay. So you've seen a lot of technology waves.

>> Yeah.

>> You worked at Gartner, you guys have invented the hype cycle. But I don't think you developed the hype cycle.

>> No, no, no.

>> I've always amazed... It's like the buzzword cycle.

>> Yes, correct. Yes.

>> Amazing. But anyway, when you think about AI, and generative AI specifically, your opinion, how do you see it relative to other transformative technologies? The internet? A lot of people say, Michael Dell on theCUBE has said, "Oh no, this is bigger." A lot of people have said, that he's not alone.

>> Yeah.

>> What do you think? What do you guys think?

>> No, I think I have seen internet bubble. And at the time in 2000s, and I think we are at 95 Windows, Windows 95, Windows 98, at that era. And I think it can change the world, definitely. I think I feel like looking at the generative AI and what it can do, the main challenges of back then, to create a website is, it works on Internet Explorer, breaks on Netscape, if you remember.

>> Yeah.

>> Yeah, so it is the same thing. If you use one LLM model, it works. And in another LLM model, it breaks. So it depends on that, where that LLM comes from, that's a different thing. But we are at that stage of 95, 98 era, and I see that in couple of years, it will take off, like an internet. It's another internet, I would say. I see that way. Or another iPhone.

>> What do you think? You think as big? Bigger?

>> Oh, yeah. Bigger.

>> You think bigger?

>> Yes.

>> Bigger than mobile?

>> Yes. Bigger than mobile.

>> It's mobile social.

>> Yes.

>> Big data, internet, bigger than the internet.

>> That's what I -

>> I would agree, by the way.

>> Yes.

>> Because you have the internet.

>> Correct

>> And you have the cloud, so it actually should be... And you have a lot of money going in.

>> Yeah. So what's going to happen is the shine is going come off-

>> Which is starting now.

>> Which is starting already.

>> And then the work will still happen, but it'll go away from the mainstream in your face. And then in five years it'll become mainstream, just like what happened with the internet. But the question I have for you, is there something that you are looking forward to, some change? We have the transformer architecture. Transformer architecture really changed everything. But knowing technology, something new is going to come up, do you see limitations of the way Gen-AI is done today, and something that you expect will change?

>> Yeah, no, definitely. I think, so it is interesting, today there will be some changes will come. I think I'm expecting as well. So this is interesting. I have two different personal AWS accounts. This is an example, actually. I tried with a program, with a Python, wrote a Python, used Rags, sent a PDF, asked good questions in one account. Used the same model, exact model, and use the same, I copied the code.

>> I see .

>> I use the code, and then I sent the same PDF, ask a question in here, they're giving different answers.

>> Oh wow.

>> So same chunking. So the chunking parameters, I tried tuning them. Same exact here, one account A, I get like "I cannot answer." I set proper guardrails in one of the cloud vendor. And then here I set the same settings, no change, and I'm still scratching my head of why is that? So I think it is, until we reproduce the results, and that is when it can take off. There is definitely there is hype to it, and there is some technology needs to improve. And right now it is like our internet back in 95, 98. And you know how you can send something, it generates, back then it was there. You know can generate STML webpages on the fly. And then you have to, it generates like 20,000 lines. You have to cut it down and make it usable, really. That is where I see now in Gen AI world. So it takes the technologies to improve, definitely, number one. Number two is the cost, and cost should decrease a lot. And we should be able to use on any... I think they're coming now. I think I see, even in NVIDIA versus AMDs, you can train some on AMDs as well now.

>> Correct.

>> And then we have use cases. We have recipes already, and we have ingredients. And I think that tools definitely needs to come down.

>> It's interesting, the example you just gave, about everything was identical but the results are different. Is that because we don't know how it actually works? Is it still a black box, technically? It that-

>> It is a black box. In fact, I am shocked that even data scientists, hardcore neural network people, cannot tell you how exactly LLM works. Because it's got so many layers, and layers, and layers that it works. But how it works is a complete black box, even for them.

>> It is kind of like magic. That to me, the way that AWS turned the data center into an API, and democratized the access to technology for any company, small company can now have Fidelity class infrastructure, maybe not the people but the access to that infrastructure through AWS, you now have a situation where technology, LLM's turning technology into a natural language interface, and putting that at the fingertips of virtually everybody. Kind of like search did, but in a much more powerful way.

>> Yeah, correct?

>> It's, that to me is why I think it's going to be bigger, even though we're working through these problems, these 95, 98 problems.

>> Yeah.

>> So I read something very interesting. We talk about hallucination as being the problem. The output of a probabilistic model is always hallucination. Always. Now whether-

>> By definition.

>> By definition. Because it's just a statistical model that is just estimating what it should say. So it is by definition hallucination. If it's up to your liking, then it's fine. If it is not, then it's a problem.

>> Right. Or if you're trading stocks based on it, it's a problem.

>> Correct. Yeah. Yes, yeah.

>> So, interesting. Is there anything else on your radar that you can talk about, that is exciting you right now?

>> So I'm excited about any problems that I have, or any challenges. I think the biggest challenge is the data quality. I think no matter how much we do, and still it is insufficient, because if we have to make some decisions down the road. If you are answering something with this AI, we have models, we have everything. They're at good at that place. At the same time, using their output, how can you prescribe? If they prescribe something, can we use it in production or not? That's a challenge. For that, how can you have a good data quality, good data governance, and human in the loop. I think that's exciting as well. I think all these are exciting challenges now.

>> How far out do you look? Do you look at things like quantum? Or is that just too far? It's not...

>> No, I think it is there. It is also coming. I think quantum is coming as well.

>> Yeah, it's finally time to pay attention, isn't it?

>> Exactly. It is time to pay attention, and-

>> Would you take the AI and quantum come together? Is that how you see it?

>> Yeah, I think so.

>> Really?

>> I think quantum can do good calculations at some point. So I think they can be independent. They can be independent, but you can use it the same level of, I would say, I think they can go in tandem. So it is not like either or, or they can go in tandem, or some of the output from GenAI can be used in quantum, some of the quantum outputs can be used in GenAI. Some like math. It can do math pretty fast. We know that.

>> IBM will love you for that.

>> Yeah, yeah.

>> Were you at GTC?

>> No.

>> So we had a private meeting, the analysts at GTC, with Jensen.

>> Yeah.

>> And he said a couple of things that were interesting on quantum. He said we're the biggest quantum computer company that doesn't make quantum computers.

>> Oh, wow.

>> And then he said something that was controversial. He said, "Pretty much anything that you can do with quantum, you can do with AI." With the exception, he was saying, of cryptography. And a lot of people say, "That's not true." But the fact that he said it, it's Jensen, and anything Jensen says...

>> Right. Anything Jensen says, we take it as gospel. But he says a lot of things that are not necessarily going to be the fact. For example, he's saying, "Stop teaching your kids coding languages. It's finished, that era." But I don't think so.

>> No, it still there.

>> I think we still need to-

>> I hope not. I think there's logic involved.

>> Right.

>> Structure and... And the other thing he says is Jensen's Law. We have Moore's Law...

>> Yes....

>> which is the semiconductors. We had Metcalfe's Law, which is the internet, that scale. And now we have Jensen's Law, which is buy more, save more.

>> Ah, yes. Yes.

>> That's his law. Krishna, thanks so much for coming on theCUBE.

>> Thank you so much.

>> It's great to see you, man.

>> . Yeah, it's great to have you on.

>> Hope to have you back some time.

>> Thank you.

>> All right, keep it right there. We're going to wrap up day one just after. This is Dave Vellante for Sanjeev Mohan, you're watching theCUBE's coverage CDOIQ 2024. We'll be right back right after this short break. Thank you.