Nagarro SE (NA9) Earnings Call Transcript & Summary

Hello, and welcome to all of you on this episode of the Sommelier series hosted by Nagarro. I'm very happy to have a chance to talk to Dr. Venkat Padmanabhan here. He is a Managing Director at Microsoft Research, and I would consider this position a very influential one. He's been part of the research community for a long standing. He is also a recipient of Shanti Swarup Bhatnagar award, which is a prestigious award. And along with, he's also a member of the ACM, IEEE and several organizations. I have the privilege of knowing Dr. Venkat since his college days at IIT Delhi. And he went on to do his MS and PhD from Berkeley, California. And his -- I would really like him to talk a little bit about his research interest and topics that he has been working on for the last few years before we go on to the topic for the day, which is how is AI impacting or creating social impact in our world, which I think is an amazing topic generally for so many reasons that I would like to explore with him. So over to you, Dr. Venkat. I mean perhaps you can share a little bit about yourself, your research interest and other interest that you think would be relevant for our audience to know.

Yes. Thanks, Anurag. Really happy to be on this session today. I think, Anurag, you already summarized my career trajectory. What I would say is the thing that drives me is work that has an impact on users, right? So throughout my career, I've been looking at work, not just in the AI context, which is a recent phenomenon, but even in the past, on looking at ways of leveraging technology to make a difference for users in a somewhat direct and tangible manner. And I've done that in the context of mobile systems, sensing, networking and other areas that I've worked in the past. So I'm very happy today to have the opportunity to tell you about some of the work that my colleagues and I have done and other work that's in the space of AI and technology for social good and having a discussion in general.

Thank you. I'll keep calling you Venkat, it's easier, and also because I know you. Yes. Yes? So wonderful. So let me quickly introduce myself as well. It's only fair. I'm Anurag Sahay and I am the global practice leader of AI and data science at Nagarro. At Nagarro, we've invested a lot in these topics, and we are continuously trying to build our mind share and thought leadership around this topic. We see a lot of value coming from a topic like this for our clients and customers and also to the larger society we are living at. And we want to explore that in a lot more honest way, which is why we have a topic like this at the Sommelier, in the Sommelier series. Now before I start on this topic, which I think is a topic that I think at some level has been talked in a few different ways in the past. It's not a completely new topic. I mean AI for solution would been talked about in many, many different ways. But before we address the topic, I want to talk a little bit about the problem of meaning. The problem of meeting the way I look at it, Venkat, is this problem that we are all looking for meaningfulness in our lives as people. It's just a natural thing to look for, at least I look for, and I'm sure that of what I know of you all and others around us, right, it's not just always about money, status or success. Sometimes you are looking for are we doing something around in the world that we live at, right? And technology has been such a thing. It's been such a transformative element in our lifetimes. I mean if you look at what has happened in the last 100, 150 years, pretty much all the magic in our lives in the world that we look at, technology has been one of the most significant elements in this, right? So it's natural for us to think about like how could technology become even better driver of social good in our lives, right? But with this problem of meaning, the problem of like can we make our work more meaningful, would you want to share some things around this feeling and sentiment that we have usually before we go into the topic of technology and so on?

Right, right. Yes, I think that's an important point. I think the way I look at it is that I would not separate out doing good and doing what makes sense, for example, from a business report. I think the 2 are actually, in my view, sort of tied together, right? So essentially, the way I look at it, we want to do what that touches people perhaps on the other side of the world who've never seen you, who you have never met in a positive way, right? So in some cases, you do it with a commercial product. In other cases, you do it as part of a social good application. But the reality is that you're looking for hard problems that technology, and AI as part of that technology suite can help solve. And so I'll give you sort of maybe a concrete example of some of the work that -- something that has grown some of my work. So road safety is an issue that's been very close to my heart, right, because it is a pretty major problem around the world, in particular, in India. It's actually -- it's huge. In fact, it is a -- if you look at sort of the causes of fatalities and causes of death, it's in the top 10. So it's really a public health issue, right? And when I -- I worked with Microsoft Research in the U.S. for about 9 years, and then I moved to the lab in Bangalore around 2007, about 15 years ago. And one of the first things that hits you almost literally is the traffic, right, and how chaotic it is. And of course, having been in India for the first 20 years of my life, I'm quite familiar with [ the sit downs ]. I was quite familiar with [indiscernible] but when you come back after a gap, it really hits you. And so my colleagues and I said, hey, the smartphones, this is around 2007, so smartphones were just emerging, we were wondering can we use smartphones to do something about this, right? [indiscernible] phones those days didn't really have much by way of sensors and so on. So we kind of had to put together a contraption with the phone with some external sensors and so on. And we built a system, we called it Nericell, for monitoring how a driver was driving, right, things like sharp braking and speeding and things like that. So at that point, it was interesting research, but I think we were a bit too early, and we didn't really have the hardware platforms that allowed us to scale. But several years later, about 2016, I said, okay, time for a fresh look, now in the context of phones that are much more capable with cameras and other sensors and also AI models, in particular vision models, deep planning models that can do interesting things with the data that the sensors give you, right? And so that led to a project called HAMS, which I'm happy to talk about later in the conversation. And that had actual tangible social impact, right? So we're really happy that an idea that started in 2007 with an urge over time has led to something that is actually having a very positive impact on people.

That's a very important point you're making. I mean one thought you're suggesting here is that you don't really differentiate between technology being used in a certain business context versus technology being used in a social context, right? In some indirect ways, they're all related to each other. But still, we have a topic like AI for social impact, right, or social good. Somewhere we want to make this explicit this idea that technology has to be very explicitly be directed in a positive way as opposed to it being accidentally directed or not being directed. Would you think there are some elements there at play?

Yes. So actually, I would look at it as a problem, defining it. So if you pick a problem that matters to lots of people in order to make [indiscernible] and you go after it regardless of whether the business opportunity in terms of monthly returns is large or not, I think that to meet defined sort of societal good being front and center and the business aspect being perhaps [indiscernible], so it's really a question of how we select the problem. The technology that goes into then sort of creating a solution to get to the product, I think that may or may not be any different from what you would do for a commercial customer, right? It is really the selection of the problem and identification of the problem that, I would say, matters here.

Fair enough, fair enough. I think it's a very good point you're making. This point, I think the way I understood it is that, what is at the center? If the center is money or is the center the larger impact of [indiscernible]

The problem you are trying to solve. [ If solved in any way, that's good ].

The problem you are trying to solve. Right. And you could make a lot of money from it at the same time, but that was never at the center, right? And that's just an accidental outcome or a side effect.

Yes, yes, yes. And on the topic of business and so on, I think the thing that's important is that even if you're doing societal good work, it has to be self-sustaining, right? It has to be something that pays for itself, not necessarily make you a lot of money, but it has to sort of [ let you see ] go to partners. It has to make sense for that partner to then take your technology and actually know how to use this, right? So I think it's kind of important for that to be thought through because otherwise, these things will just die in many cases, right? They are not going to sustain themselves. So I think it's important. And again, I'm happy to talk about our experience in the context of this road safety, what we did.

Yes. Yes. Very good points. Wonderful. So let's go a little deeper into this idea of AI and technology. Now before again we go into AI especially because that's the topic of interest here for -- at least for me, but let's just maybe take a step back and just address technology. We have some time, so we can talk a little about it here. So technology also sometimes is -- and I think you hinted at this a little, that's start of our technology, it's how you are framing the problem. The same technology could be relevant for both ways. But sometimes people are afraid of technology. We are afraid -- we sort of look at technology as a sort of -- in a dystopian context. It's sort of like almost by default saying, I've come across a lot of people who would -- of course, a lot of people in the business we are at, technology is a pretty positive force. But a lot of people look at technology as also a negative force. What would you talk to them? How would you talk to them? Yes.

Right, right. Yes, this has probably been true for millennia, right, which is every technology has good users and users that are less desirable, right? So I think the important thing is to focus on the positives that you can have with technology, while, of course, at the same time, doing what you can to make sure that it doesn't get misused. So for example, maybe there's a good point for me to just get concrete with this project called HAMS that we did on road safety, right? So essentially, the idea was to leverage a smartphone as a way to monitor a driver and their driving, right, to improve road safety. That was the goal. Now this is exactly the flavor that you mentioned, which is, yes, you can improve safety by watching over the driver and making sure that they're doing the right thing. But the flip side is that this could become the -- like a Big Brother that's watching over the driver and including their privacy and this and that, right? So there are sort of those legitimate concerns as well. So we -- in this project, we created the technology, which is really using the sensors of the smartphone, the cameras and the accelerometers, the gyro and so on, along with machine learning algorithms and other techniques to infer various aspects. So for example, whether the driver is distracted, whether they're scanning the mirrors, but that's a very concrete example, right? So when you're driving, you should look straight ahead most of the time. I mean that's a good thing but you could also scan your mirrors because that's sort of what gives you a situational awareness, right? If there's a certain event and you need to take a [ emergency ] maneuver, you need to be aware of what's around. Now that's something that's sort of easy to say, but it's very hard to measure, right? So we use the camera of the phone along with some others to sort of figure out whether the user is scanning their mirrors [indiscernible]. So I think -- and all these things. And then we said, hey, what's the use case for this. And then we run through many use cases. One was a fleet owner might want to keep a check on their drivers, for example. But that's got this Big Brother kind of -- it's got usefulness too, but we are worried about that. And the insurance would want, and so on. Finally, just through chance, we settled on this problem of automating the driver license test, right? So it's a particular problem, I think, in India and maybe some other parts of the world where a lot of people get their driver's license not actually taking a test because it was [ a safety ] and not hard to influence. So what was nice about this is that it's a very confined thing. Now you go take your test, maybe 15, 20 minutes, in and out, right? So concerns around privacy and being watched over by the system and so on were essentially nonexistent. People don't real worry about it. On the other hand, people liked the fact that it was a machine, a piece of technology that was administering the test. And even if they failed the test, they were actually happy because they thought the machine was objective. There is an inspector that failed them, they think blame it on the bias of the inspector. Oh, the inspector didn't like me or was in a bad mood or whatever, right? So there were some positives, and the negatives were, in some sense, not that much because of the context, right, because [indiscernible]. And so for us, it was a perfect fit because it's technology for good without having to worry. And the good here is, by the way, that the system [ does it much cleaner ]. So you have -- you're doing this so you don't have the human feelings coming in the way of license testing, including simple things like making sure that the person taking the test is the same as the one they issue the license to. So I think this is a long-winded answer to your question, but certainly, it's important to worry about, to think about the downsides as well as the upsides of technology. But if you stay at it and you think through this and you are persistent, I'm sure you'll be able to find a formulation that emphasizes the positives while limiting the downside. I think that's really the lesson for that.

Excellent points. I really like the idea around how technology clearly is better off when it comes to human failings, and that's like a positive side. Clearly, that one can very easily see. Of course, there is the bias can creep in into the technology as well if you're not careful. But that's an excellent point you make. And that's just a very good segue into the next question. But before I do that, let me just once again welcome anybody or everybody who's joining late. We are talking here with Dr. Venkat Padmanabhan. He's Deputy Managing Director at Microsoft Research and very closely working with topics around social impact of technology and AI. And we are having a chat on how can AI bring more societal good in the societies we live at. Great. So let's just slide in into the area of AI then, but before I do that, I mean let me just quickly get the hype out of the AI, so to speak, right? I mean ChatGPT has changed everything, it seems like, in the last few months, literally. I've just come back from U.S. and I was in an Uber and the guy started talking, and the first thing he talks to me about is like, do you know ChatGPT, maybe because he thought that I would know something about it and so on for whatever reason. But I was just stunned by this very casual mention of ChatGPT by Uber driver in the U.S., right? But I mean obviously, each one of us here has been affected in some form through ChatGPT, right? But the larger question, Venkat, I mean what do you think is the hype in AI? And what do you think is the reality in AI, before we go into the impact of AI on social growth?

Yes. I mean that's a great question. Not easy to answer, but I'll take a stab at it. Well, I should mention, Anurag, I'm going to talk a lot about various projects at Microsoft and my colleagues and so on, but the opinions I'm expressing are my word. We are not -- I'm not speaking on behalf of Microsoft, just to be clear.

Very important. I mean these are your opinions, and I think it's really important to call it out. It's not Microsoft Research, yes.

Yes, yes. So AI, it is a field that's advancing at a stunning speed, must be very bad, right? So I think what was viewed as a distant dream just a few short years ago is now becoming reality, right? ChatGPT is a great example of that, and there are other like a [ Galil ], there are all these new age tools. I think what's happening, and frankly, I don't think even the experts understand it -- by the way, I'm not an AI researcher. I use AI in my projects where I can, but I'm not an AI researcher. But even if you talk to the AI researchers, the AI experts, people don't quite understand how these large language models, for example, are working, right? So there's this notion of as the model becomes large enough, 0.5 trillion parameters or 1 trillion parameter, or whatever the largest size are these days, there is all a matter of sort of emergent behavior that I respect and things happen. But the way I see it, with a large enough model, you can perhaps, and that's what sort of things such as ChatGPT are sort of proposing, right? You can perhaps get to the point where you have a predictive model, as a generator model that can predict the next word to utter indistinguishable from human because it's sort of learned all the relationships between past participants based on tons of data. And so given all the previous context, it is able to pull out the next word that is not that different -- not -- no different from what a human would do, right? If you're able to achieve that with a large enough model -- I am not saying that ChatGPT has achieved that, but I think they're getting to the point where -- then in some sense, you could say that, hey, that thing is sort of equivalent to a human in terms of output right now. Whether it's human intelligence in the sense that we think of ourselves as intelligent, that frankly becomes beside the point because if for the same input, you're producing output that's equal to a human and the machine is doing that, because it could play the role of a human, where it's to take a medical test or a large exam or whatever, all these things that this machine is able to do, right. So I think the hype -- in the past, people would say, hey, [ camp runners ] talk about some piece of technology, maybe there's a bubble there, right. But I think with the announcement of [indiscernible] AI, I think there's a lot of substance here. People have tried Chat. I mean you have tried it, I have tried it, we are all able to try it. This is not something that has just been hyped up. Of course, it's got imperfections. I think the people who have created it, OpenAI will necessary admit that it has [ its limits ] and it has issues. But it is something that will improve and is improving. And so I think it is something that people in every field should take seriously at this point.

Wonderful, wonderful to hear that and assuring to hear that. I'm also on the fence in some cases, right? I mean very long time ago, I came across this little code from Dr. von Neumann, who was a mathematician and was part of the quantum phenomena right, the whole theory building of quantum mechanics. And he said something like, oh, give me 4 variables and I'll make an elephant go through the eye of the needle. Like the kind of nonlinearity that is possible and the possible -- the flexibility that there is, in an equation, if you have 4 variables is sufficient for him to make an elephant go through the eye of the needle. And today, these 200 billion, 0.5 trillion parameters models, right, I mean God only knows what that really means, right, in terms of what kind of modeling is getting captured. So I'm also, at some level, completely disoriented and I'm not sure of how this -- how to really integrate it. But of course, the outcomes that we are seeing, right, with these kinds of models is also pretty remarkable. I mean on the one hand, you don't have so much accuracy, but you have the kind of breadth and general -- not understanding, but at least the ability to answer or discover answers to questions with such a remarkable precision, that's also pretty amazing. But yes, absolutely, we can't dismiss it straight off. Great. So with language models on -- and maybe we'll circle back to language models, but let's just maybe again go back to the whole question around social good. I mean it would be wonderful, Venkat, is if you could maybe walk us through a few examples of how and what kinds of projects have you have been a part of when it has come to social good. You already talked about road safety, maybe that or maybe a few other ones. And what would be even better is if you could take one and dive deeper into it as much as possible so that we can have some insight into what does a typical application which has a strong social impact typically look like? And what are the challenges, concerns and other context that we should be thinking about?

Yes. Okay. So the way I see it, where AI and technology in general get help in societal impact context is where you can either lower the cost of something, increase access to something, make up for human failings, as I mentioned in the earlier example of road safety and [indiscernible]. So I think the starting point for a lot of these things is really sort of understanding what is it that is holding back a particular application, right? And then trying to see how technology and AI can help with the cost of the access and so on and so forth, right? Yes. So let me -- I'll come back to the road safety thing because I think there are some interesting lessons to be learned there as well. But let me talk about something that's sort of much more recent. It's not my work. It's work of a colleague of mine, Pratish Kumar. On leveraging these GPT 3 and 3.5 kind of large language models for something that's very interesting, it really sort of speaks to this access problem, right? So you have this knowledge models that do very interesting things. The question was, how do you leverage that to help someone in some remote part of the world access information -- natural sort of human like way without requiring any additional effort, right. So the simple insight in this work, it's something called Bhashini bot, right? They created a WhatsApp bot that allows someone to speak Hindi or any Indian language, I think, have a natural conversation with the system about some government programs, right? And the simple observation here is that there is the interface between the user and machine. So things like language translation and so on have relevant level. And then there is a reasoning part which is in the back end, which then takes what the users spoke and figures out what is the most relevant piece of text to generate in response to the user's prompt, right. So they decouple the 2. So essentially, they have language translation and text-to-speech and speech-to-text happening at the front end using [indiscernible] and they just leveraged the existing GPT 3.5 model with the right kind of prompting. [indiscernible] so this is a model that was generated, it was not only framed on the government data, it was not certainly set in Indian context, but it had this "reasoning" ability right? So given a certain front, it can generate a response that made logical sense. And so they went from translating to essentially speech to text, text to text translation, fed that along with sort of other carefully crafted information as a prompt to the back end, get a response and then do the reverse and produce speech again in Hindi as the demo that was done. And it was really stunning, right, because essentially, this is something where you are able to leverage technology that was created without really having this user context in mind. The sort of the right decoupling of the decomposition of the problem into the user interface and the reasoning engine, sort of really sort of a [ hype ] magic happen. So that is an example of that. With regard to the road safety HAMS project I talked about, there are a couple of points I want to make, right? So AI is obviously huge that I already mentioned, right? This is reality, and I think all of us need to wake up to it. But in making any of these things work, it's an engineering problem. You can't say, hey, I want to just focus on just AI and then forget the rest, right? So I agree with a very concrete example, right? So our idea was to use the smartphone in a way to do monitoring of the driver and so on. But we were deploying it in North India where in the summer, temperatures would go to 45. And with this thing running all your AI models, the temperature of the phone could be 50 degrees, right? It would just shut down. It could just -- we had the phone just shut down because of the compute that was happening on top of the environment being really hot. So you have to be really careful about how you leverage different paradigms of competition. So for example, the best AI model is not something you might want to run all the time. You may want to be selective in using that because it's computationally expensive. And in this case, we can computational costs really manifest also as heat and then you sort of use much simpler traditional vision-based techniques. So you have to sort of make these together in a careful way. So that requires sort of understanding the constraints of the engineering problem and then sort of working out a solution. So the other point I would make about this project is that when we did this work, thinking about automation and how we'll sort of reduce cost and deal with human feeling and so on, and then the pandemic happened and there was a very unexpected benefit, right? So [ as tragedy ], in the pandemic, no driver and no inspector wanted to be in the car together, right? So the idea that we had a phone addressing the test was a bonus that everyone was very happy with, right, because they didn't want to meet in person. So the point is that the unexpected benefits happened. But anyway, so I think I made a few different points here. But my main point here is that think of these as engineering problems, AI is a big part of our toolkit but there are other things to think about as well. [indiscernible] colleagues did was really you can't just throw GPT 3.5 at the problem directly. You've got to think about [ pose it ] and do the right fit. So that's what I would say.

Wonderful answers. And thank you so much. There is so much knowledge hidden in all those things that you said. Let me bring them out a little bit for the benefit of some of our listeners. One thing that you talked about was how AI can improve social outcomes, and you give 2 indicators to it. One is that it decreases the cost for something and/or it increases the access to something, right? These are the 2 important dimensions you point out. And that those are very good mental model things to keep in mind as you're thinking. There are some things that have been said around AI generally, that it brings on the cost of prediction on an average. And so because it reduces the cost of prediction, and prediction is a very important input to decision-making, it's a really amazing thing what AI can do. So -- but that -- what you're referring to is really important. And the second thing is improving the access to something. And the example you gave to that, right, which was that how Bhashini bot -- is that Bhashini, you said?

Bhashini is a program in India. The bot was what my colleagues created as a way of -- yes.

Right. So if I think about that bot, what it's doing is it's allowing 2 people with understanding of completely 2 different language, first of all, by just using speech to text and text to speech in a sort of a translation paradigm. First of all, it allows that interaction to happen. But not only that, but it's also allowing the person to interact with the knowledge base somewhere and retrieve knowledge and get information in an automated way using language as a modality, a communication or speech as a modality of retrieval, right? And so suddenly, the access to both knowledge and communication is so much more increased because of the fact that 2 people with 2 different language paradigms can talk to each other, which is, I think, is a great interaction. I mean the amount of interaction possible now between the set of 8 billion people is suddenly gone, I don't know, to a completely different level. And then the second is being people being able to access information which is -- or knowledge which was probably not easy to access because they maybe did not know how to write or type or something, or they did not know English enough and they could not have written in English. And in this case, they could come in a language like Tamil or some native language of some local region, and if people have created enough of an AI model, that could be facilitated, right? So that's like improving the access to both knowledge and communication, which is dramatic. The third thing you pointed out, which I think is again a fascinating topic in itself and something that we should dive into is that it's not only about AI models. It's about engineering that whole thing. And that's super important because now because of the -- our attention being taken over so much by AI models, we should not just devalue the engineering that is necessary to build such solutions. And that's only going to get more complex and more challenging in the days to come because you're going to try and instrument this to our reality more and more, like this better condition in India and so on, right? So excellent points, Venkat, there. And -- but any other examples you would like to bring around social impact that we could explore in these areas, AI or maybe related?

Maybe I could mention a couple of -- I mean I mentioned one other project and then maybe throw out an idea for something that people could work on as an interesting social impact thing. So the project I want to talk about is a project that was called Dependable IoT. So there's nothing to do with AI per se. We start off with [indiscernible] AI. It's really the simple idea that the Internet of Things and sensors everywhere and that whole vision. The whole foundation here is on the sensor and the data that it's collected, right? And if that foundation is shaking, then everything else you build on it is on a shaky foundation, right? And so the idea there was can we done signatures, physics basic natures of sensor behavior that will tell us whether a sensor is working correctly or what, right? And so the -- so sort of understanding the physics of sensors and sort of understanding RC codes -- maybe Anurag, you remember like from our entry days, we had all these sort of courses on electrical engineering. And so we talked about the RC code, how things decay, there are all these sort of really fundamental [indiscernible] that are happening. And understanding those was really the foundation. Now where the ML and BI came in was then sort of in being able to create signatures that could easily be checked and so on. So certainly, it was -- and it's important that we're talking about IoT devices, very little memory footprint and so on, so also the ML model had to be really compact so it could be able to run on a really low-cost sensing device. So it is clearly important, but again, going back to my point about understanding the engineering problem, you have to understand sensors and how they work and what fundamental [ calculatons are there ] and so on. Now the societal impact that -- so this work is done sort of as really a technology-driven research. But the very interesting societal application was in the context of air pollution monitoring. There is a company called [ Respiral ]. Respiral builds these boxes. And I think we have a large number of them across India and [indiscernible] and beyond. Looking at air pollution monitoring, looking at air quality, one of the issues here is that there are these really large -- I forget what exactly what to call, I think BAM or something -- there is really a high-end air quality sensing outputs that are there, but they are very expensive. So they are like maybe you might have 2 in a city, right? We want to have -- [ Respiral ] wanted to have a much higher density of sensing nodes, but that meant going with low-cost hardware and then you have to worry about all these other sensor issues and so on. And so when they found out about Dependable IoT, worked up by some colleagues of mine, they were very interested in integrating that into the sensing box. So some of the dependable IoT techniques and the sensor signatures used sort of compact machine learning to figure out whether the sensor is working well or not, are part of the boxes that are deployed nationwide, right, to do air quality modules. So I guess the point I want to make here is that, again, AI is the flavor of the day and it's huge, but really don't lose sight of the problem you are trying to solve. Don't lose sight of all the tools available in your technologies [indiscernible], including AI, right? So really keep that in mind. The potential application that -- actually, my colleague, Pratish, who did the Bhashini bot, he was talking about creating an intelligent tutor. So let's say you're a student trying to [indiscernible] in some subject. Maybe you are in the 7th grade or something and you're trying to learn something. Can you leverage these generated AI models to create -- like for example, you can tell ChatGPT, explain something as you would to a 6-year old.