Enterprise AI Modernization with Rakesh Ravuri

Russ Fordyce 🏆 (00:01.004) Welcome back to the Winner's Circle, everybody. Today I am here with Rakesh. He's gonna talk to us about AI and how to kind of recode the world. And he is part of Publisys Sapient. They are the makers of Sapient Slingshot, this tool that helps us do really cool things with development. And they are the latest AI Excellence Award winner. Rakesh, welcome to the Winner's Circle and congratulations on the award. Rakesh Ravuri (00:26.05) Thanks, Russ. Thanks for having us here. Russ Fordyce 🏆 (00:28.288) Absolutely. Now, before we get started and all the cool stuff you're doing, give us a little bit of background about Pueblo Sysapient and maybe a little bit about Slingshot and how you got started building it. Rakesh Ravuri (00:38.563) Yeah. Publicis Sapient is the digital transformation arm of the publicis group, which is the world's largest media and communications firm you must be aware of. Publicis acquired us around 2015-2014 time frame. They basically saw the world of media and advertising going more distal and they realized the value of having a distal transformation not just for your software systems but also all across the stack from media communication. your ad targeting, etc. Right? Like that is before the AI ML wave took off. And along with Publicis Sapient, they also acquired Epsilon, which is the world's largest data broker. So as you know, AI is only as powerful as the data and the tech systems behind it, which integrate all the other systems. So they were in an ideal place to sort of create real value for our customers. And Publicis Sapient especially is like our main purpose is to help companies transform, right? Like we are a digital transformation and we have been doing that all the way from since what we have been founded in MIT's campus back in 90s. We started off as a company helping companies adopt internet, right? Like that was the transformation trigger. And over the years, the transformation triggers changed, like transformation remained a constant, right? Like first it was internet, then it was e-commerce, then it was mobile, then it was cloud. And now AI is the latest transformation trigger. And people have always used us as a tip of the spear to enable transformation of their organizations in order to make themselves relevant and competing in the new world which we are doing. So our latest trigger right now is AI. And that is what we are helping companies use AI in agent transformation, which is, again, at the end of the day, it is transformation of their processes and company. Russ Fordyce 🏆 (02:33.494) Yeah, and it's interesting to me because we have been in all these different kind of transformations. And I read the nomination. What's really cool about this is Sapient Slingshot really started from kind of an internal tool that you built, I think. How does a tool you made for internal use become a product that now is kind of modernizing enterprises at scale? Rakesh Ravuri (02:47.213) Yes. Rakesh Ravuri (02:53.388) Yeah, so... Internally, when we first saw the chat GPT movement in November 2022, obviously everyone wanted to get hands on it. But we as a company, we work with lot of clients. So we can't risk putting the client's information into a public tool and add that to a startup at that time. No one was sure about how your data is going to be used. Later people found out that people were training using the data, et cetera, and all that. So right at the start, we wanted to let our people know the power of the chat GPT platform and what it enables so that like we can see what value we can create for the customers. But we were equally aware of the risks which were involved in it, right? How can you put customer data? How are you going to use it? How are you going to actually, if people are using it, we also need to know how people are using it so that we can create a real value and the real processes behind it. So internally we built something called PSChat, right? Like I whipped up something, now it's called wipe coding, but like at that I think back in March, right after November in March, we basically quickly put up a company portal We negotiated a deal with OpenAI so that like we they don't use our data for Training and we got indemnity clauses, etc All those good stuff and we were able to release a Like a chat GPT kind of clone for internal users, right? Like that is how it started So when we released it to the internal users, I could immediately see that there's benefit of people writing applications on top of it. So we created an API layer, right? We created an API layer and some people in our organization use those API layer to create the first Visual Studio plugin. This is back in 2023 April, right? Like long before cursors, cloud codes, none of those things existed at that time. Yes, GitHub was there. GitHub was doing only code completion at that time. They didn't have a copilot extension, but people were seeing Rakesh Ravuri (04:54.806) that okay they're going to PSChat typing something they're copying the code and putting it in Visual Studio then we said that okay if you have the API layer I can do it all from Visual Studio directly right like that is how it started that was the beginnings of the slingshot product right like so Russ Fordyce 🏆 (05:13.078) Now you guys are really focused in or the nomination was focused in kind of what I call is, you know, legacy technical debt. And it's really something that almost every telecom provider, every technology provider, pretty much every enterprise has, which is some sort of system that's sitting in the back doing its job, but it was developed 20, 30 years ago or 10 or whatever under applications that are, you know, either defunct or, you know, kind of been modernized. If you give us a kind of an executive summary, kind of that slingshot technical debt problem, give us a little bit of that and tell us a little bit about what you're doing with slingshot. Rakesh Ravuri (05:52.878) Yeah, so again When we started using AI and when we had this PS chat and slingshot interfaces, we had one of the customer reach out to us saying that, look, I have millions of lines of COBOL code, but we don't know, we want to modernize it to use AI, right? Like they wanted to add AI features to it, but they are not able to add it because the code was in COBOL and like from COBOL calling other AI services, like it was a heavy lift, right? Okay, can you help us understand what this code's business rules are so that we can implement in the new system? And can you use AI to do that? So we use that API layer, what we talked about, and we basically said that, okay, fine. So we have the APIs to the AI. So why don't we automate the process, what a human does, where it takes a chunk of code, paste it in a prompt and ask, what does this code do? And then quickly we realized that at that time, the context limit was only 4,000 tokens, right? Like if you remember, right? Now we have million tokens, but it was 4,000 tokens there. So then we realized quickly that we needed a system to intelligently break large pieces of code and tie the context and create a coherent picture. Like, so we basically said, okay, how do we break the code? Right? Like breaking code is not just at the time chunking was like, you just take chunks of file. It works for text documents, but code documents, you can't break it in middle of the function. You can't break it in the middle of the word or something like that. Everything breaks. So what we started doing is okay we needed internal intelligent chunking so we wrote parsers which can break the code into sections like which are like limited to 4,000 tokens or something and then we built a layer to tie all the chunks together create a coherent map and we built a context where we at the time it was we didn't even know that it was called context but we were building a context layer right we are building an intelligent context layer and this tool became our legacy modernization platform right when we when we saw the whole process how we did it is Rakesh Ravuri (07:55.674) we wrote the whole process on a board, right? Saying that, okay, if I was a COBOL expert and how will I modernize this, right? What will a human do? And we looked at all the processes and then introduced which and all parts we can use AI to do it. How can we automate it? What are the constraints for the AI to do that? So we broke down the whole process. We basically wrote a system and we were able to use that very successfully to understand the business rules in a 20,000 screen COBOL application, right? And obviously the next question the client asked is this awesome, right? Like I now understand what my system is doing, which no single human understood before. Now can you extend that pipeline to actually create a new application, right? Like in the new stack. Right? So initially we did code to specification part of it and then we extended that thing to us from specification to code. Right? And then we had the full life cycle. We could go from COBOL code to Java code, right? All the way. Right? And the next obvious thing which people like once they do it, they obviously want more. Right? Like, so they looked at it and say, this is great. But can I introduce my new requirement before you generate the code? Then we introduced a step between spec and spec to code called spec. specification or spec enhancement which basically allows you to enhance the specification or add new stories or add new functionalities. So this whole platform as we sort of did. as we started doing it, we learnt lot of lessons of what happens in a system. And what we realised is like the others in the market were trying to do the same thing, but they did a brute force approach and we were more successful. Their brute force approach was like they were skipping all the middle step and say, let us go from quote to quote. Rakesh Ravuri (09:45.07) In theory, it seems like a great thing, but it never works. And that's what we learned organically and that is why our platform even today, right? Like more than one and a half years later, right? Like when they pit us, most of the clients give us because there are multiple options. They basically do a POC phase. Invariably in the POC phase, the quality of output what we produce comes on the top, right? The reason why is because we went from the first principles approach of looking at the process and then fitting the AI, not taking the AI first approach and then going back, right. So that was the long answer to your question. Russ Fordyce 🏆 (10:24.598) Yeah. Well, the cobalt problem has been a problem for decades. We were talking about this earlier. The cobalt almost suffers from that kind of first movers advantage in the software world. It's decades old, but it was one of the first platforms really to help businesses automate and really kind of move into that modern kind of computerized era. And it was quickly overpassed by a million other languages, but that knowledge, all of that programming was still baked in. And now we don't We don't really have those people around. They're all retired or on to other things. Why do you think it took so long for this decades long wait to get at this? Why didn't somebody build a JavaScript COBOL translator back in the day? Rakesh Ravuri (11:10.51) Two reasons, right. The first reason which I heard from my previous companies and my previous jobs is if it is not broke, don't touch it. That was one of the reasons. It was working reliably for like a decade or more in most of the organizations and if it is not causing you a problem, there is no reason to touch it. That was one of the reasons. Then the question comes why now? Why now is basically the software systems are evolving faster than ever now. So as I told you, Now, if you have a COBOL system, is, let us say a green screen, which is accepting a user application, Like people expect now to have some AI enabled features, like an intent detection of the user, like a fraud detection, or like ability for a user to sort of like run some agents before you sort of submit the things, et cetera. Now, in order to introduce these features, which has almost become an expectation, you can't do with the old screen right like you will quickly your business will look outdated if you are using older things without having AI features in your application and that is what is forcing the companies to say okay now is the time we need to modernize because they can't modernize a COBOL application and add these features they need to first move the stack and then do it and now another reason why they are touching right now is now they are getting increasingly confident that AI is actually able to understand code better than any human can explain to me So that is another reason, Like AI is like with 98 % probability it is able to tell better than even the original coder. Like because I can tell you as a software programmer, right? Like if you wrote a program and if you ask me after two years, I will not be able to explain in that detail as what AI can explain the same program to me right now, right? Because AI can keep the whole program in the context and it can actually give you a better description than maybe my skills of describing. Rakesh Ravuri (13:16.336) are not that good as skills of programming, right? Russ Fordyce 🏆 (13:19.67) Yeah, now you mentioned context a lot and I think the real kind of core differentiator here is that you guys built this enterprise context graph that has all the kind of details about the software and kind of how it relates. Give us a little kind of background on what it is and why does giving AI a map of the system before it acts, why does that change everything? Rakesh Ravuri (13:39.564) Yeah, so context as you know is basically a set of information like a reference which you give to your AI model saying that, okay, use this as a reference, not just use your trained data, but use this additional context and then answer my question, right? Like that is the basic definition of context, what you sort of paste in your prompt or your system prompt or whatever it is. Now that is quickly limited by the model window which you have, right? Like you can't paste too much context. So people have built rag systems, like so that like you can retrieve the context and then paste, but after retrieving, you're still pasting it into the context window, right? You still have the limitations of context window, but at least you are able to query something. What we quickly realized is when I did the software, as I talked to you about the legacy modernization, right? Like we went all the way from the old code into new code with the new stories being introduced. What we realized was the AI world is very siloed. There were a lot of tools for individual faces separately, right? Like coders have Visual Studio plugins or... whatever coding agent was there using, right? The people who are the product managers had their own tools embedded into maybe their Jira or Atlassian ecosystem, right? The designers, Figma and others had their own AI tools inside that, right? Your strategy people might be using ChatGPT or like whatever it is. What was... being is that the context was getting recreated every time or they are reusing pieces and chunks of context, but they're all in silos. So as a developer, I see a story saying that, okay, we need to add a Google Pay button to the e-commerce cart. Like let us say that is the specification of the story. Rakesh Ravuri (15:30.418) AI will basically say what is the context that requires, how to connect to Google Pay, what I need to do, maybe some context on what are the test cases, what are the frameworks I using in this library and use it. That is all what I call it as local context, that is not enterprise context. The real context is why am I adding the pay button? Because that context was there by the product manager when he created the story. The product manager said that, I'm using a Google Pay because this set of maybe Zen Z audience prefer Google Pay over credit card. That is the reason why I want to add it. So the context is, OK, if the audience is Zen Z, maybe the Google Pay button should be the default button and then the credit card becomes later. more context a good developer will do that. And the context is, okay, it has to always work in this particular scenarios because that was the hypothesis. The hypothesis context is not there for the coder because coder is just pulling a story context or something like that. Something has to tie what was the context used while creating the story? What was the context used by the developer while writing the code? What was the context the tester used while testing the story? And what is the context which is required when you saw a production defect? Now, all these things have to be tied. Right. that tying across the enterprise different phases is what we call enterprise context. So that it is not context island, but this context or like a graph which ties your enterprise process. So your AI processes, when they are working in any particular node of the processes, have access to the context graph of what happened before, which makes it intelligent. Humans indirectly have it, right? Like, so we don't write it, but that's what makes a good team member from a bad team member. The guy who has context is able Rakesh Ravuri (17:20.698) to sort of do things more appropriate. Russ Fordyce 🏆 (17:25.962) Yeah, and I've seen it time and time again, exactly that context layer, especially with developers is if they're not involved in that conversation for building a product, for understanding a product, then they can't build that context around it. And so it was really important in my kind of career growing up and building products of making sure that you had that team member in so that they understood all of that. And you're right, without... Rakesh Ravuri (17:49.304) Thank you. Russ Fordyce 🏆 (17:50.944) Without that context, you're really flying blind. And when something comes up, you just don't know what to do. As a developer, would imagine you're like, well, do they want this or that? They didn't tell me. Because we always leave something out as product managers, right? Rakesh Ravuri (18:00.768) Exactly. And that is why you call developers to a user experience meeting, you call developers to a product discussion, right? Or you get them to go on the field and experience how the users are. That is building context. Russ Fordyce 🏆 (18:14.836) Yeah, and it's that expertise layer of, understand what this is doing, and AI really doesn't have that. can guess, and I've seen it myself a million times. You described Slingshot as kind of that GPS for modernization. What goes wrong with AI when you don't have that kind of map? Rakesh Ravuri (18:32.034) See when you don't have that map one common problem which everyone knows is hallucination. As you said, because it has to fill the gaps of missing context, basically hallucinates. Actually, it's a very bad, sometimes I feel sorry for AI when we say it's hallucinating, but it is doing its best to fill the gaps because you have not given it the right context. So it is your job to fill the context in order to do that. And the GPS is the same thing. Why do we rely on GPS? There are many roads to go to the same destination. Which road do you take? So, that is why we sort of colloquially refer to as GPS. Like I there multiple ways to modernize. Which path will you choose? Like why did you sort of choose a particular path because of the wider knowledge of the spec and the context and the provenance of the code and where the code has come from you need to do a particular decision. That is why it is called the GPS of modernization. It is not just about reaching the destination that is code to code. It is about how you reach the destination. What is the spec you generated? What is the acceptance criteria you generated? And that determine what, how did you write code so that it can be tested for that code because you can have a code which is working but if it cannot be tested it I call it like a wheel which you can't take out what will you do when you have a puncture right like so. So you need to have ability to build robust systems to build one of the ways we test link shot. There are so many coding agents, right? People ask like, okay, what makes it yours better? I can do the same thing with others. Yes, you can do the same thing, but that is not the criteria. We use an LLM judge and say how maintainable is the coding agent thing that because that determines because that is the problem we are solving. Why are we modernizing things? Because they are not maintainable. So if you produce new code, which is not Rakesh Ravuri (20:22.582) maintainable you are you are making another problem for the future right like so how maintainable your problem another feature which we have patented which is very interesting you will and you will appreciate that is we call it a provenance feature right like we applied for a patent it is right now in the process provenance feature is when when a person uses a coding agent and changes the code we have markers at every line level which says that whether this line was written by human or it was written by AI if it is written by AI what was the context used by the AI, when it was written, what model it was used, right? Now that's very important. Lots of people I ask them, open your code base and tell what parts were written by AI. That is really important because tomorrow if you have a problem or if you have an audit requirement especially for large banks and insurance firms and medical firms and all this will quickly become a problem. And we built that into the system from day one whether we have modernized the code and why it was modernized which is there or if a human basically prompted and generated the code who generated when they generated what AI model was used. So tomorrow like how you have product recall someone basically says this product basically had a it didn't happen but tomorrow if someone mentions that it had a security issue and all the code generated by this code has to be revoked. What will you revoke? Russ Fordyce 🏆 (21:43.414) Yeah, I imagine that's extremely important, especially as we go forward. Rakesh Ravuri (21:46.67) And people are not thinking about it too much, right? Like these are liability concern for companies if they can't tell which was AI generated and which was human generated. Russ Fordyce 🏆 (21:50.625) No. Russ Fordyce 🏆 (21:56.47) Well, and you guys have a lot of customers in kind of regulated environments. I know you're big into healthcare and finance. And so you guys have built this kind of governance layer into the platform. What does that mean in practice for your clients? Rakesh Ravuri (22:11.854) So for clients, firstly, the two parts, even before we go into the governance and audit part of it, very simple management is right now not possible because right now CIOs are spending money on all these coding tools. They can't even tell if people have used the coding tools for their wipe coding projects or their actual projects. Because I can't look at the code and tell whether this code was generated using the tool which you bought. Right? So the only way they can measure is, okay, I'll only measure saying that whether you're doing more work, but that is not a good measure. Right? Like, if I was the CIO, I will want to know how much of my code which was checked in was generated by an agent. Right? And when it was generated by agent, like what are the tools which were used? That's the next level. Right? Like people are using MCP tools right now, like left and right. What tools were used? Right? What context was used? What documents was used? Right? When it was done, what was the prompt? was this prompt and a shared prompt which the whole team is using. Things like that, right? initially first is to just get the visibility, the traceability and what we call it as AI assisted agile, where we said that in agile, there used to be a principle saying that working code over documentation, right? With a very famous statement where everyone stopped documenting after that. But we basically said that, okay, we need to evolve thing. I said explainable code over working code. It's the next level. Documentation is easy to create right now. I can create the spec from any AI tool. So it's okay you don't write documentation. I can ask AI to explain it. But the whole point is explainability is not just about what the code does, why the code was written. When was it written? Why it was written? What prompt generated this code? So explainability of code when AI is generating 80 to 90 percent of the code, explainability becomes really important. Russ Fordyce 🏆 (24:08.194) Well, and I imagine that's where, you know, if the hallucinations happen and errors happen, you go back to that context layer and start to understand it. Yeah. Now in the healthcare cases, you guys have been modernizing these COBOL screens, you know, without the need to pull these guys, the old COBOL programmers off the golf course and out of the retirement homes. How did you kind of operationalize that? Rakesh Ravuri (24:13.836) Exactly. Rakesh Ravuri (24:30.263) Yeah, so. I will not say that without pulling the people from the golf courses, without pulling more people. Because what happens is we made it scalable. Because there are millions and millions of lines of code. What we did is we enabled to have one expert to verify the specs, the structure, what it is generating, the sampling of what it is doing. So that like we still have a human in the loop verify it but we didn't need to have like like we didn't we didn't need 50 experts one person could do 50 % job because the AI followed the structure the specification the acceptance criteria and the relationships and if there are any learnings if that one person said it could be very quickly incorporated and regenerated the whole thing right so it is it is faster to sort of take the learnings and also make sure that it is applied everywhere. So it is not completely eliminating human, it is using humans where they are really add most value and scaling their value. That is how I will put it. Russ Fordyce 🏆 (25:43.222) Yeah, it's getting that expertise and making it. Yeah. Rakesh Ravuri (25:45.134) Yeah, and how do you take this expertise and weave it back into your process and make it repeatable. One of the biggest problem with expertise is it is not replicatable. So the problem which we try to solve is take the expert's knowledge and make it a self-improving continuous process and repeatable. Russ Fordyce 🏆 (26:01.184) Yeah. Now you guys are looking at these older systems, these 20, 30 year old systems and rebuilding them in what seems like an astounding amount of time, you know, in days. How is that even possible? You've got millions lines of code and I mean, you, how is that possible? Rakesh Ravuri (26:20.27) See two things right because see agents can be parallelized so the moment you have an agent doing it see some most of the process if you see in modernization is a repetitive process. like if there is a class, if the class functionality has to follow a certain design pattern and get implemented in a different language, now you need to do that pattern maybe thousand times across the code base. Now you can do all the thousand parallel rather than sequentially, right? Like that is where you get the speed from. And second thing is in terms of consistency, right? Like it is easy with AI to verify the consistency and if something is wrong, ask it to redo. Redoing the work is a lot faster. So you can iterate multiple times and you can parallelize things. If you put both of them together, what seems like a large amount of work can be compressed into a short time. Russ Fordyce 🏆 (27:15.724) Well, I imagine as if I was your customer, you come to me and I've got this problem of, this thing's been in there for 25 years and Bob, the developer, is gone. We have zero documentation. When you tell them that we'll have this done in a week or a few days, what's their initial reaction and what's their reaction after those few days? Rakesh Ravuri (27:37.922) So obviously it is disbelief. Obviously they're like, OK, why should we trust you? So our standard thing is, OK, let us take a small module and show it can be done. So we generally do a POC phase. And what we do is we keep something constant so that it's measurable. Like, for example, if the application is most of the applications write to a database, like when they do some transaction or read from a database. So we'll say that, we'll not touch the database layer firstly, we'll just change the application layer and to show that it is the same behavior or we'll replicate the database, create a test copy of the database, we'll write the new process running parallel to it, writing the data and say you compare the data. If the operation results in the same data in both the databases, that means we have replicated it exactly the same. Like that we have actually we also have a patent process called DBRE right like it's like a how you measure the business results and the business logic of making sure that you have done this right thing how do you measure the old system versus new system and compare them right so we have multiple techniques we developed over what you can say engagements, we apply all of them and we give them those options saying that okay which option convinces you most or like which you will trust it more even if a human is doing how will you verify right like we start from that then when they say this is how we verify and say okay we can replicate that. Russ Fordyce 🏆 (29:10.754) Well, let's walk through that a little bit because these are enterprise systems that are in charge of our finances, in charge of our health. In my experience, you get into these AI modernizations. I've been working on it a little bit on my programs. There are these hallucinations that happen, but then there are beyond hallucinations when the AI is just confidently wrong. They're sticking by their guns. This is the way to do it. How do you build trust with clients whose system just simply can't afford that? Rakesh Ravuri (29:44.066) Yeah, so see, two things, right? Like when you ask AI for an opinion, that is when you get this. But if you tell AI a guardrail saying that, it has to pass this test case, there is no way for it to be confidently wrong. It can't confidently say, I'll not pass this test case. Either it will say, I can or I failed. So, what we do is like the whole test driven strategy, what is your constraints you put on the system, what are the guardrails you put on the system, what is the verification and evaluation loops you put on that. And those things have to be deterministic. You don't need your AI process to be deterministic. You need your evaluations and your tests to be deterministic and you need to be sure about that. Like for example, the example which I gave you, Like a particular use case will result in a database structure in a particular way. You can't change that. That is the end result because other systems depend upon the data. So you can't change that. We can write tests for all the edge cases for that and say this test has to pass, no matter what code you generate, the test suit doesn't pass then you are right, you need to redo it again. Russ Fordyce 🏆 (30:53.93) Yeah, now if I'm a CTO, I'm one of your clients or a soon to be client, and I'm sitting on all this deprioritized legacy debt, what's the first kind of internal conversation that I need to have with my team to kind of get ready for a modernization or even to start thinking about it? Rakesh Ravuri (31:11.768) Two things, right? Firstly, most of the companies, they already see the power of it because most of the companies where we have sort of actually engaged it, they already had some internal thing because people could see the thing that, yeah, I can actually, because people were already trying it. They were taking the old code base and asking AI saying that explain it to me. So there were already internal pilots going on in order to use AI to understand the code or to test the code or to sort of modify small chunks of it that were already happening. What we sort of when we go there we tell them how we did it at scale, right? Like yes, one thing is to do POCs and like experiment and convince yourself that most of the people are already convinced that AI can do it. What they are really stuck at a CIO level is how do I do it at scale? Yes, my team has showed me that they can do it for one file, two files, but I have three million lines of code, right? Russ Fordyce 🏆 (32:04.886) Well, and I imagine that they start like I did. They open up ChatGPT and say, develop me this. And you've got Copilot, Cursor, and dozens of other, millions of other platforms now. What does Singshot do that none of them can, and why does that matter? Rakesh Ravuri (32:10.016) Yeah, exactly. Yeah. Rakesh Ravuri (32:23.534) As I already talked about, I referred to some of them, right? one of the things is like the provenance feature. I have not seen that in any one of the tools which you mentioned, right? Like we, because the reason why like they have not implemented is firstly, they are like focusing only on the engineer. Secondly, what we are doing is tying the context across all phases of SDLC, right? So because as a company, we call it a speed acronym, strategy, product, engineering, experience and data, right? So we used to operate because when you want to transform a company, one of our earlier things before AI was like there was handoff problems. Like there used to be a strategy consulting firm, then there used to be a design agency, then there's a development company and then there's a maintenance company. Everyone used to blame everyone else but like the problem is your transformation is not done right like it doesn't matter whose problem is it right so we basically have all the capabilities and we used to do end-to-end transformation see we used to measure ourselves saying that final outcome is transformation it doesn't matter who succeeds in like there's no point in engineering succeeding or like the strategy succeeding right everyone has to succeed So when we did that, when AI came in and everyone was using their AI tools to do that, we found the same problem, right? Again, we are creating silos and saying that my AI tool is working, but it didn't solve the problem. So what we said is we need to tie all of them, right? So the biggest learning for us is if you don't tie the context, that is the feature which other tools don't have because they all operate in a single space. Russ Fordyce 🏆 (33:46.732) Right. Rakesh Ravuri (33:57.902) Slingshot is actually a platform and it has multiple tools like we have Visual Studio plugin, have a Jira plugin, have a tool, a backlog manager tool, we have a legacy modernization tool and we have a sustain.ai tool. So these are all the tools working on the same platform which is powered by an enterprise context graph. So now you can understand, right? Like the platform versus tool differentiation, that is our biggest differentiator and we built for enterprise first use case. We didn't build for a consumer. Most of the tools are built for NB2C kind of customers. When I say enterprise first, you can ask me what does that mean? That means like, what are the governance procedures? What is the audit mechanism, right? Like if you are using a tool, how do you ensure that the person who is calling the agents context is passed to the tool also so that like the tool is using the same authentication context so that like the MCP tool is not accessing a database because right now and using someone else's credentials and giving you the data. and putting it in your code, right? So that is a serious security problem. So we need to, we build this platform to respect the enterprise guidelines, to respect the enterprise governance procedures, to give explainability and traceability to companies, provide extension points, no lock-ins, so that you can change the model providers you want, you can change the infrastructure that you want, you can choose to run your agents in your prem rather than running on the cloud, right? options. that is our differentiators, Giving flexibility and enterprise control to... Right now agents are black box if you're running on a cloud because you can't tell really what agent did, whether the agent posted your data to some other place. There's no way to tell that. Russ Fordyce 🏆 (35:43.424) Yeah. Now you've been sitting in this role for a few years. You have massive experience of helping other companies with AI and some pretty big companies. I see the shift of everybody kind of using it as a task accelerator, a task manager, do this, do that. You guys have really made it kind of this full life cycle partner and it's changing how teams work together. How does that look like? You what does that look like from the inside as you're, as you're building and kind of managing this team? Rakesh Ravuri (36:11.798) Yeah, that's a great question, right? And this relates to the previous question also, right? One of the things what we built is once you use an AI tool, this is something which I asked my partners when they sort of come in. I said, OK, you used XYZ tool. So who is your star developer, right? Like they'll say, this person, he does amazing work with this coding agent. I said, like, how is the person sitting next to him benefiting? They don't have an answer for that. They are like, no, he teaches them or he sends his prompts to them. But I said like, that's the manuals, what you can say, tedious way of doing it. So I tell to CIOs saying that, companies are built with teams. It is not about individuals. You need to raise the team, not an individual. Because if you're a writer, an individual, you're making one link of the chain better. The same thing as my saying that, OK, your engineers are producing code faster but your other SDLC is slow, how does it matter? So what we built into this is a hierarchical concept of like something like I'll give you one example. We have something called a prompt library. Now this prompt library can be maintained at a team level, at a product level, at an enterprise level and at an industry level or something like that. And I'm as an enterprise architect, I can set a guideline saying that, I want every payment thing to be audited to this audit database. Now. As a team developer, I don't need to know about that. I basically say that I'm working on a payment project and I'm implementing this functionality to accept this payment. It immediately knows that guideline and before it generates the code, it puts the audit logs because enterprise architects guidelines take precedence, which get appended to his query. Right. And a team lead might say, I want you to use log for J for logging. Right. Like that's the next level. Okay. You want me to log the team lead. Rakesh Ravuri (38:08.342) said you need to use this library the user actually wants to do this functionality I'm now adding all three together now the whole team is benefiting so someone is setting standards now you can see how it gets propagated right So that is how an enterprise and a team benefits from having a collaborative context sharing, collaborative best practices sharing. And this is just not related to prompts. Prompts are the lowest level of sharing, but it applies to your context sharing. applies to your, and the guidelines on the context, who has access to what context. If you just ingest all of them into a vector database, there might be some sensitive financial algorithms. People can just ask the context and say, okay, tell me how this, how our order. say fraud check works that suppose not to be accessible to anyone only to them. So you need to, when you ingest the enterprise context this should be guardrails which who has access to what context and how can they use. You should be able to know the fraud interface guidelines so that you can call it in your program but you should not know the algorithm because that is sensitive, that is trade secret, maybe whatever it is. So things like that you need a platform to enable that, individual tools will not be able to do that. That is one of the ways where we help the teams collaborate better. Russ Fordyce 🏆 (39:27.53) Yeah, and I think as I talk to people, you see that a lot of this kind of disparate work going on that doesn't have that commonality layer or that context layer or that shared libraries that you're talking about. And you have all these guys kind of going off on their own trying to do things. And as this agentic world develops and matures, you've got to have that. What do you think this looks like in five years if we keep accelerating at this pace? Rakesh Ravuri (39:56.067) Yeah. In five years, what I feel is in addition to sharing the common context and all of those, it is also sharing behavior. See, when people talk about enterprise context, tell them, like the, see, we already had data warehouses and data lakes and all, they have the enterprise data, right? Like, so what is difference between an enterprise context graph and a data warehouse and a data lake? Enterprise context, primary thing is to create links of the data, which is that in your data warehouse with behavioral links, right? Why is this behavior important? important is if you want an agent to do the job of yours, like maybe you get an email, someone asking you saying that I want to sort of do a podcast, you might say yes, you might say no. And that is captured because in the reply I can see whether you said yes or no. But I don't know why you said yes or why you know that, is that in your head? Like you don't really document like every time you said no, you don't take another page and say I'll write it down why I said no or something like that. Now you can build agentic transformation where agent will basically go through initially it will be very simple. It will use historical data which is basically like AIML and I'll say based on your previous behavior, I decided that you should decline this request and accept this. you will look at the data and say no no no I want you to accept this also and decline this then the AI will ask you why right and you will give them some maybe one sentence feedback saying that no I decided that I will not do any podcast anymore on ERP systems I decided that I want to do some podcasts on AI transformation you should remember that. Rakesh Ravuri (41:36.878) So next time when the agent runs, it will do that. So it's capturing your behavioral data. And tomorrow, you are not there. You are on vacation. Your assistant basically comes and says that, I am not accepting this. Your AI will say that, no, based on what Russ told me last time. he wants to do AI transformation work. You might want to consider like accepting this or at least check with us on that. That is behavioral capture. Right? It is not just automation. It is behavioral capture. Now, just scale the simple example at an enterprise level. Why does the supply chain order went through? Why was the merchandising rejected? Like why was this code review accepted when the functionality works, but still the architect basically asked me to rewrite this? Right? Why was that bug occur even though acceptance criteria met. Now this is institutional knowledge and you need to build upon that. One of the other issues you are talking about individual tools is they all talk about learning from their memory but they learning individuals behavior they are not learning enterprise behavior. You need to know because the team people will come and go you need to know how that team as a collective made decisions and why was one product the team produced by one team is better than the product produced by other. Right now no one can explain it. They will say they had better people. Russ Fordyce 🏆 (42:36.353) All Rakesh Ravuri (42:53.229) Okay. Russ Fordyce 🏆 (42:53.28) Yeah, yeah. Now, Puglisys has been at this a while. What do you see as the horizon, you know, kind of for the next step of kind of the slingshot tool? Rakesh Ravuri (43:03.662) See, as the name suggests, like Slingshot is what we feel is if you take the analogy of Slingshot, it will basically give you that acceleration to take on problems which we thought were like, just people thought that we better not touch it because it was too hard or too complex, right? So the legacy modernization, we never worked in COBOL before, right? Like as a company, we were always working on latest tech stacks and all those things. We were not like a scale company which used to manage COBOL. But that didn't stop us from now taking COBOL code and doing it. So what I think what it will enable Publicis Sapient is to really pursue our goal of transformation of companies without having the barriers of limitations of what can be done and what cannot be done. I think AI will enable us to reach the true transformation potential which we can offer to our clients. Russ Fordyce 🏆 (43:59.382) Yeah, now you've been through all these transformations. I remember that digital transformation has been thrown around for so long of helping companies basically get online. This is such a different beast and you guys are doing such an amazing job at it. So I really thank you for all your time and all the success that you've had. Congratulations. Rakesh Ravuri (44:21.047) Thanks for your time. It was fun talking to you.