MongoDB .local NYC 2024 | Ashvin Parmar | Capgemini

Clips
More from MongoDB .local NYC 2024

Ashvin Parmar

VP, FS GenAI Center of Excellence

Capgemini

play_circle_outline Introduction to the interview with Ashvin Parmar from Capgemini

play_circle_outline Capgemini's $2 Billion Investment Explores GenAI: Surge in AI Expertise and GenAI Training

play_circle_outline The role of AI in job creation

play_circle_outline Building industry-specific solutions using MongoDB

Info

Ashvin Parmar | Capgemini

Ashvin Parmar

VP, FS GenAI Center of Excellence Capgemini

search

>> Hey, welcome back to theCUBE, everybody. We're here in New York City. Come on inside. We are wrapping up with day-long deep dive into MongoDB, MongoDB's ecosystem, and we're going to wrap it up where the rubber meets the road. Ashvin Parmar is here. He is the Vice President for Financial Services for GenAI Center of Excellence at Capgemini. Ashvin, thanks so much for coming to theCUBE. Great to see you.

Ashvin Parmar

>> Thank you having me, Dave.

>> Yeah, you're very welcome. So you're in the heart of it all, New York City. It's the money capital of the world. London would argue with that, but we'll take New York City. And so, we've been just talking all day about customers, how they're applying AI, how they could get more ROI out of it, because they're struggling right now. A lot of experimentation. But start with your role, your focus on FS, share with the audience what you .Absolutely.

Ashvin Parmar

>> So I'm doing a couple of roles. One job is to lead the data and AI practice globally for financial services for Capgemini. And then secondly, about a year ago I took an additional role to lead the Center of Excellence around generative AI for financial services globally. And in that capacity we are looking at how can we generate value, how can we drive efficiency and productivity gains using GenAI across the value chain for our clients, both on the banking as well as on the insurance side.

>> So you're at the heart of it. You guys have announced this $2 billion investment in GenAI. Give us the details around that. What are

Ashvin Parmar

>> the specifics?We are only one of the three global SIs who gave made such a big commitment towards AI and GenAI. So over the next three years, we'll be investing two billion and we are using it primarily for three or four different things. Number one, we are going to be doubling our capacity in terms of data and AI experts globally from around 33,000 to 60,000 plus over the next three years. Secondly, we were looking at a number of value based solutions, which are focused on these sectors or the verticals and how we can help clients reap the benefits of Generative AI, but in the domain context, in the context of their business. And then, lastly, we are also investing heavily with our partners like MongoDB and others where we are jointly co-developing solutions and taking it to the market. One other thing, we are also forward-looking. So we have an AI lab set up, which is looking at six years down the road in terms of what is coming down the pike and how we can prepare now to reap the benefits when the moment arrives.

>> So you're doubling the number of AI experts at your firm from 33,000 to 60,000 plus over the next three years? Did I hear that correctly?

Ashvin Parmar

>> That is absolutely correct. Now, in addition to that, we are training about 100,000 people globally on AI and GenAI to make sure that all AI and GenAI literate so that we can apply it effectively for the values and risk.

>> That's in addition too, or is that included?

Ashvin Parmar

>> Yes. 33,000 to 60,000 are the AI practices and then we are doing a mass enablement for the AI and GenAI technologies globally so that everybody's AI literate within our company.

>> The obvious question, I thought AI was going to kill all these jobs, but you're doubling the number of headcount that's focused on AI. Explain to people why people are still needed.

Ashvin Parmar

>> Yeah. If you look back at internet, when internet came, people though that was going to kill the jobs. Now, if internet goes down, there will be significant impact across the board. So I think AI and GenAI will actually add value. It will allow the talent to focus on the things which can be more creative, more generating value as opposed to doing the mundane routine task, which can then be outsourced and augmented with AI and GenAI. We are not saying it's adversely impacting in terms of our business or our client's business. It will increase the throughput, it will increase the productivity and the efficiency, and it will help us transform some of the inefficient business processes.

>> I'm going to show you a picture because the second thing you talked about was focus on specific sectors, domain specificity. So this is a picture, and I'll describe it. Well, you can grok it. I'll describe it for the audience. There's a picture of a power law. A power law, imagine just a right angle. In some industries, power law, you have a couple of, maybe a handful of dominant players and you have this long tail of smaller players. What we did is we took liberties with the power law we had the vertical axis was the size of the model. So you got the big giant model players, Open I and Tropic, et cetera. And then the long tail is domain specificity, and that's kind of what you were talking about before. Now, the torso is not a hard right angle because of open source and third parties pulling up that torso. So it's a very interesting dynamic. First of all, we did this shortly after ChatGPT came out. So this is how we see it. I think it's evolving this way. Do you see it that way? Are customers actually starting to build along that long tail?

Ashvin Parmar

>> Yeah. The way we look at the AI and GenAI use cases is in two dimensions, as I was mentioning. One side is the value dimension. The other side is the operation efficiency dimension. On the value side, you can apply AI and GenAI in two ways. One is you can do the expert augmentation or you can have AI and GenAI run autonomously. And then if you look at the X domains, similar to the chart that you were showing me, you can start with the large models which are developed on an internet available corpus of data. And then if you go up, you become more specialized. Then you train with industry specific data. Then you train with your own company's specific data or your line of business in the specific data. So as you move up the chain, your specialization becomes more tighter, your differentiation becomes more profound, and that is how you can also control some of the side effects of GenAI hallucinations and so forth and really control the outcome and really make an impact from the value generation perspective.

>> Are customers doing this on-prem? It seems like there's a lot of action in the cloud. We've talked to customers who have said, "We're a little nervous about that data leaking. You can do VPC and stuff in the cloud, but we don't necessarily want to move the data. We have a lot of data on-prem. We got a lot of data in the cloud. We want to bring the AI to the data." I'm not saying repatriation, I'm not a repatriate, but are you seeing that activity on-prem?

Ashvin Parmar

>> Yeah, so we see both extremes. We see some clients who are adopting cloud and GenAI in a big way and having production-ready applications which are customer-facing on the internet. And then there is other more conservative organizations who tend to be really concerned about the data leakage, the IP leakage to some of the external models and they're looking more for an on-prem type of deployment. We see both the spectrum of use cases. For people who are concerned about the safety and soundness of AI and GenAI, we tend to recommend having a human in the loop. That seems to control the risk at the same time you'll reap the benefits out of it. But you do need a strong governance process in place whether you are on-site or on the cloud.

>> We were talking to somebody from Meta because last year, actually late last year, who inferring the downloads, this was a Lama too, they inferred that close to half of the downloads very well could be companies doing things on-prem. So that sort of validates what you're saying there. Where does MongoDB all this? Where do they fit into your 2 billion investment. Over three years, by the way. That's pretty good. A lot of times people say $2 billion over 15 years and so three years, that's really compressing it. Of course, a lot of that's people, but where does MongoDB fit?

Ashvin Parmar

>> We have been working with MongoDB for quite some time, primarily on the data modernization type of programs. But over the last year or so, since the GenAI became all the rage, we have developed a number of industry-focused solutions which are built on top of the MongoDB. In fact, a couple of hours ago, I was presenting one such use case, we call it GenYoda, which is a solution which is built specific for the insurance industry where we are using it to make the underwriting more efficient, claim processing, more efficient in the contact center, more efficient, but in the context of the domain. And under the hood is MongoDB with the collections and the vector search. So it is been working quite nicely. We have been able to ingest a large corpus of data in a short amount of time. So the scalability and the performance has been working out quite nicely. So in addition to GenYoda, that I mentioned, there are five or six other industry solutions that we have built in partnership with Mongo and we will continue to invest as more and more use cases surface.

>> What are you seeing that are the main blockers for AI, GenAI specifically adoption? We're seeing a lot of experimentation. We see a lot of rag. Actually, pretty easy to do. Maybe not at scale, maybe not governed, but you can build a proof of concept pretty easily. But many customers are saying, "We're still really in search of that big ROI use case." Not so much ROI, but the NPV. The size of the benefit is not there. You can get quick hits, definitely productivity, but they're small little hits. What are you seeing as the adoption barriers or the blockers to get bigger ROI?

Ashvin Parmar

>> Yeah, I think the primary apprehension is about the safety and soundness, like losing your IP, losing the data or having the solutions do unintended things, which are valid concerns. A lot of the times we also see in the prototypes and proof of concepts, things work spectacularly well. But as you scale that up with thousands of unstructured documents, then things start to hallucinate. So having the adequate guardrail in terms of your training data, having the adequate guardrails for your model, and then also on the outcome side, how are you using the output coming out from some of this model are extremely important. The regulatory landscape is also evolving With the EU AI Act being one of the first one, which is bringing the risk perspective in terms of the high risk activities are monitored or prohibited, and the low risk you go allowed. So finding the right balance of how much risk exposure that you have, having the adequate governance and oversight is critical. We tend to use the expert in the loop for anything which is consequential in terms of decision-making. But things like the operational efficiency made on the software engineering side or on the creative side in the marketing space, those things, we see a lot of products and ready use cases.

>> So you see a lot. You have a big wide observation space. You guys are consulting with a lot of organizations. I'm interested in your thoughts on two questions. So there's a lot of companies in the database business, a couple actually, developing LLMs. We saw Databricks and Snowflake developed their own LLM. Mongo's not doing that. And Dave has said, "We're not doing that. We're going to put our R&D elsewhere." So I wonder your thoughts on that. The second part of that question is, I've had some financial services companies tell me that they're going to build their own LLMs. I was initially surprised to hear that they're actually working with semiconductor firms, whether it's their specific use case, try to drive low latency, maybe they're worried about open source licenses, et cetera. Remember the early days of the cloud?Banks used to tell me, "We're going to build our own cloud. We're better at it than the hyperscalers.", Well that didn't work out. Is it a similar situation here or are there really valid reasons why end customers will actually build their own LLMs?

Ashvin Parmar

>> Yeah. So there are three models in terms of leveraging the LLMs. You can buy off the shelf and use it as is that you can buy, customize it. And then third is to build your own. To be honest, I've not come across a lot of clients who are inclined to build their own models just because it is a massive undertaking. It requires substantial amount of investment in terms of the capital, in terms of the people and the infrastructure to support that. And on the outside of that, the return is not really clear that it will generate. So most of the clients that we work with globally tend to be falling in the second bucket where they would take an LLM and then train them with their corpus of data within their sector. The second trend that we start to see is there's a trend away from the large models to more smaller purpose built models in terms of the performance. They tend to be doing quite well. So it doesn't make sense to build your own model in most cases. So we don't see that a whole lot.

>> What do you tell customers about that other risk that I mentioned that was mentioned to me about the licensing terms in open source models? We're worried they're going to pull the rug out from underneath us. How do you manage that risk?

Ashvin Parmar

>> We have actually built an experiment platform called Genesis and it allows customers to look at the use case, be there on the technology side or on a value use case, be it marketing, sales, underwriting, risk, whatever. And depending on the use case and what they're trying to do, we have a library of models that can do that effectively. And using Genesis in a blind test, we allow them to check for accuracy, completeness, costs, etc. And then they select, okay, what makes sense, as opposed to being biased by one versus the other. And then we run these experiments with a large set of people at the different capability and then use that to make an informed decision. So that works quite nicely.

>> Very nice. What's next for you?

Ashvin Parmar

>> What's next is to continue driving more value-driven use cases and then see how we can use this to really transform some of the inefficient processes produce new service lines. That's what we are thinking about.

>> Well, congratulations. Really impressed with the multi-billion dollar investment. Would love to have you back to talk about the progress and the learnings and the results you're seeing with clients. Ashvin, thanks so much for coming

Ashvin Parmar

>> on theCUBE.Thank you.

>> Appreciate it. All right, thank you for watching. That's a wrap from MongoDB Local wall-to-wall coverage. Go to siliconangle.com. You'll see all the news. Go to thecuberesearch.com, which is our research site. And, of course, thecube.net is where you'll find all these assets on demand. Don't forget, go to thecubeai.com. Sign up for our public beta. Thank you so much for watching. This is Dave Vellante, for theCUBE and the entire team, great job and thank you to MongoDB. We'll see you next time.

Ashvin Parmar | Capgemini

search

>> Hey, welcome back to theCUBE, everybody. We're here in New York City. Come on inside. We are wrapping up with day-long deep dive into MongoDB, MongoDB's ecosystem, and we're going to wrap it up where the rubber meets the road. Ashvin Parmar is here. He is the Vice President for Financial Services for GenAI Center of Excellence at Capgemini. Ashvin, thanks so much for coming to theCUBE. Great to see you.

Ashvin Parmar

>> Thank you having me, Dave.

>> Yeah, you're very welcome. So you're in the heart of it all, New York City. It's the money capital of the world. London would argue with that, but we'll take New York City. And so, we've been just talking all day about customers, how they're applying AI, how they could get more ROI out of it, because they're struggling right now. A lot of experimentation. But start with your role, your focus on FS, share with the audience what you .Absolutely.

Ashvin Parmar

>> So I'm doing a couple of roles. One job is to lead the data and AI practice globally for financial services for Capgemini. And then secondly, about a year ago I took an additional role to lead the Center of Excellence around generative AI for financial services globally. And in that capacity we are looking at how can we generate value, how can we drive efficiency and productivity gains using GenAI across the value chain for our clients, both on the banking as well as on the insurance side.

>> So you're at the heart of it. You guys have announced this $2 billion investment in GenAI. Give us the details around that. What are

Ashvin Parmar

>> the specifics?We are only one of the three global SIs who gave made such a big commitment towards AI and GenAI. So over the next three years, we'll be investing two billion and we are using it primarily for three or four different things. Number one, we are going to be doubling our capacity in terms of data and AI experts globally from around 33,000 to 60,000 plus over the next three years. Secondly, we were looking at a number of value based solutions, which are focused on these sectors or the verticals and how we can help clients reap the benefits of Generative AI, but in the domain context, in the context of their business. And then, lastly, we are also investing heavily with our partners like MongoDB and others where we are jointly co-developing solutions and taking it to the market. One other thing, we are also forward-looking. So we have an AI lab set up, which is looking at six years down the road in terms of what is coming down the pike and how we can prepare now to reap the benefits when the moment arrives.

>> So you're doubling the number of AI experts at your firm from 33,000 to 60,000 plus over the next three years? Did I hear that correctly?

Ashvin Parmar

>> That is absolutely correct. Now, in addition to that, we are training about 100,000 people globally on AI and GenAI to make sure that all AI and GenAI literate so that we can apply it effectively for the values and risk.

>> That's in addition too, or is that included?

Ashvin Parmar

>> Yes. 33,000 to 60,000 are the AI practices and then we are doing a mass enablement for the AI and GenAI technologies globally so that everybody's AI literate within our company.

>> The obvious question, I thought AI was going to kill all these jobs, but you're doubling the number of headcount that's focused on AI. Explain to people why people are still needed.

Ashvin Parmar

>> Yeah. If you look back at internet, when internet came, people though that was going to kill the jobs. Now, if internet goes down, there will be significant impact across the board. So I think AI and GenAI will actually add value. It will allow the talent to focus on the things which can be more creative, more generating value as opposed to doing the mundane routine task, which can then be outsourced and augmented with AI and GenAI. We are not saying it's adversely impacting in terms of our business or our client's business. It will increase the throughput, it will increase the productivity and the efficiency, and it will help us transform some of the inefficient business processes.

>> I'm going to show you a picture because the second thing you talked about was focus on specific sectors, domain specificity. So this is a picture, and I'll describe it. Well, you can grok it. I'll describe it for the audience. There's a picture of a power law. A power law, imagine just a right angle. In some industries, power law, you have a couple of, maybe a handful of dominant players and you have this long tail of smaller players. What we did is we took liberties with the power law we had the vertical axis was the size of the model. So you got the big giant model players, Open I and Tropic, et cetera. And then the long tail is domain specificity, and that's kind of what you were talking about before. Now, the torso is not a hard right angle because of open source and third parties pulling up that torso. So it's a very interesting dynamic. First of all, we did this shortly after ChatGPT came out. So this is how we see it. I think it's evolving this way. Do you see it that way? Are customers actually starting to build along that long tail?

Ashvin Parmar

>> Yeah. The way we look at the AI and GenAI use cases is in two dimensions, as I was mentioning. One side is the value dimension. The other side is the operation efficiency dimension. On the value side, you can apply AI and GenAI in two ways. One is you can do the expert augmentation or you can have AI and GenAI run autonomously. And then if you look at the X domains, similar to the chart that you were showing me, you can start with the large models which are developed on an internet available corpus of data. And then if you go up, you become more specialized. Then you train with industry specific data. Then you train with your own company's specific data or your line of business in the specific data. So as you move up the chain, your specialization becomes more tighter, your differentiation becomes more profound, and that is how you can also control some of the side effects of GenAI hallucinations and so forth and really control the outcome and really make an impact from the value generation perspective.

>> Are customers doing this on-prem? It seems like there's a lot of action in the cloud. We've talked to customers who have said, "We're a little nervous about that data leaking. You can do VPC and stuff in the cloud, but we don't necessarily want to move the data. We have a lot of data on-prem. We got a lot of data in the cloud. We want to bring the AI to the data." I'm not saying repatriation, I'm not a repatriate, but are you seeing that activity on-prem?

Ashvin Parmar

>> Yeah, so we see both extremes. We see some clients who are adopting cloud and GenAI in a big way and having production-ready applications which are customer-facing on the internet. And then there is other more conservative organizations who tend to be really concerned about the data leakage, the IP leakage to some of the external models and they're looking more for an on-prem type of deployment. We see both the spectrum of use cases. For people who are concerned about the safety and soundness of AI and GenAI, we tend to recommend having a human in the loop. That seems to control the risk at the same time you'll reap the benefits out of it. But you do need a strong governance process in place whether you are on-site or on the cloud.

>> We were talking to somebody from Meta because last year, actually late last year, who inferring the downloads, this was a Lama too, they inferred that close to half of the downloads very well could be companies doing things on-prem. So that sort of validates what you're saying there. Where does MongoDB all this? Where do they fit into your 2 billion investment. Over three years, by the way. That's pretty good. A lot of times people say $2 billion over 15 years and so three years, that's really compressing it. Of course, a lot of that's people, but where does MongoDB fit?

Ashvin Parmar

>> We have been working with MongoDB for quite some time, primarily on the data modernization type of programs. But over the last year or so, since the GenAI became all the rage, we have developed a number of industry-focused solutions which are built on top of the MongoDB. In fact, a couple of hours ago, I was presenting one such use case, we call it GenYoda, which is a solution which is built specific for the insurance industry where we are using it to make the underwriting more efficient, claim processing, more efficient in the contact center, more efficient, but in the context of the domain. And under the hood is MongoDB with the collections and the vector search. So it is been working quite nicely. We have been able to ingest a large corpus of data in a short amount of time. So the scalability and the performance has been working out quite nicely. So in addition to GenYoda, that I mentioned, there are five or six other industry solutions that we have built in partnership with Mongo and we will continue to invest as more and more use cases surface.

>> What are you seeing that are the main blockers for AI, GenAI specifically adoption? We're seeing a lot of experimentation. We see a lot of rag. Actually, pretty easy to do. Maybe not at scale, maybe not governed, but you can build a proof of concept pretty easily. But many customers are saying, "We're still really in search of that big ROI use case." Not so much ROI, but the NPV. The size of the benefit is not there. You can get quick hits, definitely productivity, but they're small little hits. What are you seeing as the adoption barriers or the blockers to get bigger ROI?

Ashvin Parmar

>> Yeah, I think the primary apprehension is about the safety and soundness, like losing your IP, losing the data or having the solutions do unintended things, which are valid concerns. A lot of the times we also see in the prototypes and proof of concepts, things work spectacularly well. But as you scale that up with thousands of unstructured documents, then things start to hallucinate. So having the adequate guardrail in terms of your training data, having the adequate guardrails for your model, and then also on the outcome side, how are you using the output coming out from some of this model are extremely important. The regulatory landscape is also evolving With the EU AI Act being one of the first one, which is bringing the risk perspective in terms of the high risk activities are monitored or prohibited, and the low risk you go allowed. So finding the right balance of how much risk exposure that you have, having the adequate governance and oversight is critical. We tend to use the expert in the loop for anything which is consequential in terms of decision-making. But things like the operational efficiency made on the software engineering side or on the creative side in the marketing space, those things, we see a lot of products and ready use cases.

>> So you see a lot. You have a big wide observation space. You guys are consulting with a lot of organizations. I'm interested in your thoughts on two questions. So there's a lot of companies in the database business, a couple actually, developing LLMs. We saw Databricks and Snowflake developed their own LLM. Mongo's not doing that. And Dave has said, "We're not doing that. We're going to put our R&D elsewhere." So I wonder your thoughts on that. The second part of that question is, I've had some financial services companies tell me that they're going to build their own LLMs. I was initially surprised to hear that they're actually working with semiconductor firms, whether it's their specific use case, try to drive low latency, maybe they're worried about open source licenses, et cetera. Remember the early days of the cloud?Banks used to tell me, "We're going to build our own cloud. We're better at it than the hyperscalers.", Well that didn't work out. Is it a similar situation here or are there really valid reasons why end customers will actually build their own LLMs?

Ashvin Parmar

>> Yeah. So there are three models in terms of leveraging the LLMs. You can buy off the shelf and use it as is that you can buy, customize it. And then third is to build your own. To be honest, I've not come across a lot of clients who are inclined to build their own models just because it is a massive undertaking. It requires substantial amount of investment in terms of the capital, in terms of the people and the infrastructure to support that. And on the outside of that, the return is not really clear that it will generate. So most of the clients that we work with globally tend to be falling in the second bucket where they would take an LLM and then train them with their corpus of data within their sector. The second trend that we start to see is there's a trend away from the large models to more smaller purpose built models in terms of the performance. They tend to be doing quite well. So it doesn't make sense to build your own model in most cases. So we don't see that a whole lot.

>> What do you tell customers about that other risk that I mentioned that was mentioned to me about the licensing terms in open source models? We're worried they're going to pull the rug out from underneath us. How do you manage that risk?

Ashvin Parmar

>> We have actually built an experiment platform called Genesis and it allows customers to look at the use case, be there on the technology side or on a value use case, be it marketing, sales, underwriting, risk, whatever. And depending on the use case and what they're trying to do, we have a library of models that can do that effectively. And using Genesis in a blind test, we allow them to check for accuracy, completeness, costs, etc. And then they select, okay, what makes sense, as opposed to being biased by one versus the other. And then we run these experiments with a large set of people at the different capability and then use that to make an informed decision. So that works quite nicely.

>> Very nice. What's next for you?

Ashvin Parmar

>> What's next is to continue driving more value-driven use cases and then see how we can use this to really transform some of the inefficient processes produce new service lines. That's what we are thinking about.

>> Well, congratulations. Really impressed with the multi-billion dollar investment. Would love to have you back to talk about the progress and the learnings and the results you're seeing with clients. Ashvin, thanks so much for coming

Ashvin Parmar

>> on theCUBE.Thank you.

>> Appreciate it. All right, thank you for watching. That's a wrap from MongoDB Local wall-to-wall coverage. Go to siliconangle.com. You'll see all the news. Go to thecuberesearch.com, which is our research site. And, of course, thecube.net is where you'll find all these assets on demand. Don't forget, go to thecubeai.com. Sign up for our public beta. Thank you so much for watching. This is Dave Vellante, for theCUBE and the entire team, great job and thank you to MongoDB. We'll see you next time.