Persistent Systems CTO Pandurang Kamat on Generative AI in the Enterprise

hi I'm James McGuire and on today's e peeks we're taking a look at generative AI in the Enterprise we'll talk about navigating some of the challenges of geni and how to use this powerful technology in a responsible manner to discuss that I'm joined by Pang camut Chief technology officer at persistent systems Pang very good to have you with us today I'm very happy to be here thanks for having me James you know and I think it would be good for to start for you to speak a little bit about what persistent systems does I I know that many people obviously know the company but for those who aren't as familiar what does persistent do yeah sure James so we are a trusted digital engineering and modernization partner for Enterprises and independent software vendors all over the world they are 33 years old publicly listed in India and today have a global footprint in about 21 countries and that's uh basically what we do we help companies design new digital experiences streamline their software and product engineering modernize and move to the cloud or reimagine and modernize the overall application stack as well as helping manage their application and infrastructure that's who we are all right that's a that's a big job and and obviously there's a huge demand for those Services um I'd like to talk to J very big topic this year obviously um what what are your thoughts about gen ai's impact on customer experiences and I'm I'm I'm wondering if there wouldn't be a certain unpredictability there and that obviously generative AI but is by its very nature does generate content and it may not be always uh you know what the company wants your your sense gen's impact in customer experien what do you say sure so see we have now been working with this technology for about a year at least uh you know starting with some of the initial EXP experimentation to actually uh building out Solutions in different scenarios so the first and foremost despite the early days where there are going to be some clunkiness to this technology some clunkiness to the experience we firmly believe that over the long term this is essentially going to provide us superpowers not just to the knowledge workers but to pretty much you know anyone who's interacting with any digital tools to do their jobs or to do their cre creative uh you know work um we see this as predominantly providing very if you do it right you know very intuitive uh interventions powered by generative AI within the workflow that the person is doing it will be very assistive in nature not necessarily fully autonomous perhaps in the very long term but not in the near term uh we expect it to be personalized so that people uh actually get not very generic responses but not very contextually relevant things not just to the task at hand but to the person doing the task and overall accelerative uh technology so that is how at least when we counil or advise customers or we build things for ourselves to consume uh this is what we keep in mind in terms of customer experiences it makes perfect sense I I think clearly there are some challenges to to J AI adoption I mean for instance sometimes employees may be entering a question that might you know suggest you know corporate strategy that has not been revealed yet maybe there's confidential information uh maybe the system does spit out an answer that isn't quite in in align with what the company wants what what do you see about handling some of the challenges of generative AI sure let me since we we're talking particularly in the Enterprise let me address this with two Dimensions quickly or two perspectives rather one is from the decision decision maker of the cios and the cxos within the Enterprise the number one concern that we are hearing or even even today a year into this is uh what is the uh scenario with data privacy right uh are is my content safe is it likely to remember things is it likely to be used to train the model and therefore create a data leak channel for me that's the number one concerns most of these concerns with Enterprise services are being addressed or already addressed but that remains the number one question the other thing uh is around hallucinations as they are known where these models can start making up things not backed by facts uh and that becomes a big concerns and there is a slew of grounding tools that have emerged and continue to get better which can be used in and a bunch of techniques actually not just tools but a series of very scientifically based techniques that you can use to ground and control these hallucinations the third challenge that we see is around consistency by the very nature of these uh you know uh models consistency by default is not a tradeit that they come with so you have to build that in the layer that you're building on top of it and that uh from a decision maker standpoint these are the biggest concerns cost of course always at the top of the mind but there's also a flip side here the end user the number one concern there is a am I having to learn something new or is this going to be intuitive enough and B is this going to make my job redundant right and those are very legitimate concerns so even as we build out these things our Focus or even majority of the you know Builders who are doing it the right way the focus should be on building very empathetically to the end user making sure that they are not uh pushed out of the equation but basically they assisted to focus more human power task and not on the mundane and that's you know that's how we navigate some of this you know I think you made a really interesting point about hallucinations and I think a lot of companies worry about hallucinations from their gen AI platforms you you mentioned a possible way to to combat that I wasn't totally sure about what that was can you elaborate a little bit how could companies deal with that problem hallucinations sure so there are techniques you know a series of techniques actually from right from crafting prompts that create a very narrow channel for the model to respond and by the way even going a step before that right uh there are ways in which if you're training your own model you could train it with like anthropic did with certain Constitution that says here are the dos and don'ts these are Paramount even as you give any responses or any inferences once you're complete right that's the first step step zero if you will but if you're using a third party uh generative model you can still build layers into the prompt engineering part of it where you giving very precise directives you can ask for it to reason out and explicitly make its Chain of Thought clear site the sources you can give instructions that grounded in the facts that are being provided to uh as part of the context during inference there are about a dozen different techniques to uh make sure that at least to some extent you're able to control the hallucinations and also provide consistency of responses interesting okay well I think the the other big issue with with artificial intelligence is responsible AI I mean for instance we want to we want to get inside the black box and know why are why are those answers being created you know that AI is speding out um sometimes maybe a loan is given or or not given and we don't know why the algorithm decided to not give that loan and you responsible AI explainable AI what what is your take on this sure um that's a great question by the way because the question of responsible AI has always been around it had become uh more in focus in the last probably five to seven years but with the Advent of generative AI that has taken a whole new dimension right and in our work in AI even before this revolution we had looked at five different pillars AI should be accountable reproducible transparent secure and private now if you look at these accountability for example ethics ethical training ethical inference bias and fairness or rather uh you know controlling or removing bias and inference and being fair these are key uh uh tenets of that in the case of generative AI this means uh preventing hallucinations addressing uh the training data has been sourced ethically and again the bias removal tools that exist with AI to begin with with tools like llm Ops we are bringing uh reproducibility in a in the AI pipeline by taking prompts from ad hoc to a structured library or by bringing industry specific ontologies we are making AI consistent and reproducible you talked about explainability that's a little bit of a tough one when it comes to the specific type of AI that powers the generation of content right because by definition neural networks that power generative AI are not very explainable however the techniques I talked about earlier some of them like asking it to site the sources that led to its inference give you traceability not necessarily full explainability as it you know with other kinds of AI but at least traceability and references that will help in making that AI a little more trustworthy uh and comfortable for people and then with security there is now you you must have heard about prompt injection attacks and things like that so making AI safe not only from attacks on the AI but from the kind of output it generates like you know things that can be damaging lead to destructive behavior and things like that uh should be controlled you know by the designers of that and finally like any other system we should be respectful of the personal identifiable information keeping the Enterprise data confidential even as you build out uh you know large scale models that are served by these hyperscalers or by any other service providers that's how we see responsibility yeah it's interesting I guess that's the very first time I've heard the phrase uh you know llm Ops meaning large language model operations so it's a I can I can see where that goes that that's going to be a hot trend of the years ahead I would imagine uh let's let's look at persistent systems itself and you talked upfront about what it does but can you give some examples of how it would would serve Enterprise the needs of Enterprise clients I mean what what is persistent do on on a day-to-day basis basis for Enterprise clients sure so particularly U you know with generative AI what we are doing is see as as I talked about we do digital engineering right so we may be we may be ending up building uh very specific bspoke Enterprise workflows or uh modern ing their Enterprise workflows to bring efficiency we are taking generative Ai and creating those workflows to be more intuitive more accelerated today with cognitive search natural intuitive interface to finding digital Assistance or building digital assistance within the Enterprise the second thing uh that we do uh is building very industry specific bepoke use cases so traditionally also one of our strength has been deep Tech uh expertise that we use used to build uh industry specific Solutions with data and AI traditional AI so today we are building generative AI Solutions powered by the distinctive data that that company has and uh solving particular needs like drug Discovery say loan underwriting being made Smarter with genv assisting the underwriter and so forth but the biggest change that we see is in changing the way we build software itself right and that becomes a very important question for us because bulk of our business comes from building software for other people the nature of how we build software itself is being completely transformed dramatically by gity and we are doing this in two ways one is uh complete Green Field uh or brown field if you will software development how you can just augment what the developer is doing by helping with code Snippets or writing test cases or even creating automatic P requests and things like that but we also do a lot of business in modernizing application and that's a whole different Journey right because your language stch before and after is very different the architecture of the system is very different so a lot of our energy and effort is being spent right now on figuring out how do we completely transform the journey for legacy modernization for engineering teams as well as for the end customers and what does that look like Power by Genia and that's what I'm one of the things I'm most excited about that that is a big question going ahead you're talking about refreshing Legacy applications applications using generative AI that that seems like a it's a it's a hard task because it's the opposite of Greenfield you got to take something that's already built and and and build it more using artificial intelligence am I understanding you correctly is that the task absolutely so so companies like us have been doing that traditionally by you know having expert engineering teams that actually study the Legacy code understand the business logic reimagine what the architecture should look like and then start writing the new architecture with the business logic that you need to bake in there right if you see here whenever it comes to understanding something that's where generative AI can speed up the process of understanding as simple as he helped explain this function to me to explain and extract complete set of business rules from these 500,000 lines of code you can try and attack you know the problem at different scales there and that is what uh you know one of the things that excites us most interesting well let's look at the the future of generative AI in the Enterprise I think that that's the big question companies are really wondering where is all where is this all going mainly because they want to get ahead of it themselves so what do you see in terms of the future of generative AI in the Enterprise and I I still have to say I think it's pretty unpredictable agree with that or not NE necessarily I you stole my line my first opening thing was the answer is I don't know yes great I'm glad you admit that I appreciate that absolutely because uh I mean if you told me a year ago uh if I knew what I knew would happen uh I I would have called it crazy right or even three months ago and last week and last weekend for example all those things so um the but the few things that we see are happening one is uh we kind of started seeing this stand even back in like April of this year where the the whole initial impression about it's going to cost you 200 million $300 million to train your foundation models to do anything uh that was going to go away and we are already seeing that people are training models in the matter of uh you know days on single gpus or few gpus and and so forth right smaller models uh not necessarily the same size and scale but they have become much more affordable so we see a future with many many uh small and large models forming an ecosystem inside an Enterprise or inside an isv these are there are going to be Foundation models that are very general and can be used for more complex planning tasks but then there are going to be very very uh fine-tuned and scaled down but tasks specific domain specific models that will do one particular task exceedingly well much better than the model and will cost much less to train as well as for inflence so that's number one number two uh in the next year or so we're going to see a lot of Enterprises move from experiments and poc's to scaling to stable resilient predictable consistent uh applications of AI and we already hearing cxos talk about that to us uh the third thing will be that a generative AI will be deeply integrated into Enterprise workflows and companies uh especially the hyperscalers are constantly if you see the releases they are putting out they're constantly aiming to make it easier and faster to integrate even going down to low code or no code ways of integrating their AI into the Enterprise workflows and the core of the Enterprise as a result is going to be much more intelligent uh and generative uh By Design right and finally uh the same Act is is going to play out in software companies that companies that build software or platforms and they are going to have deep integration across the workflows and the companies that actually make uh transformative changes to the user experience rather than incremental are really going to uh win over much larger market share interesting I I I like the idea that it's G to get cheaper which since which says it's going to be democratized and more people have access to it uh pank I think you said it a lot of fascinating stuff it's going to be a very interesting sector to follow in the years ahead very much appreciate your sharing expertise and please come back and talk with us again sometime it's a pleasure being here thank you for having me pleasure talking to you

This transcript was generated automatically from the video's captions and may contain errors.