How can data science and AI drive sustainable innovation in a rapidly evolving world? In this episode, we speak with Dr. Satyajit Wattamwar, Data Science and Statistical Leader at Unilever, to explore how these technologies are shaping novel solutions. Unilever is a global leader in consumer goods committed to creating responsible, impactful products that improve lives. Leveraging cutting-edge research and development, the company addresses some of the world’s most pressing challenges, from reducing environmental impact to enhancing well-being.
Satya unpacks how Unilever utilizes its database to inform its models and how to determine the right amount of data needed to solve complex problems. Dr. Wattamwar explains why contextual problem-solving is vital, the notion of time constraints in data science, the system point of view of modeling, and how Unilever incorporates AI into its models. Gain insights into how AI can increase operational efficiency, exciting trends in the AI space, how AI makes experimentation accessible, and more! Tune in to learn about the power of data science and AI with Dr. Satyajit Wattamwar.
Key Points From This Episode:
Quotes:
“Around – 30 or 40 years ago, people started realizing the importance of data-driven modeling because you can never capture physics perfectly in an equation.” — Dr. Satyajit Wattamwar [0:03:10]
“Having large volumes of data which are less related with each other is a different thing than a large volume of data for one problem.” — Dr. Satyajit Wattamwar [0:09:12]
“More data [does] not always lead to good quality models. Unless it is for the same use-case.” — Dr. Satyajit Wattamwar [0:11:56]
“If somebody is looking [to] grow in their career ladder, then it's not about one's own interest.” — Dr. Satyajit Wattamwar [0:24:07]
Links Mentioned in Today’s Episode:
Dr. Satyajit Wattamwar on LinkedIn
Dr. Satyajit Wattamwar: If you really know what period of history of your underlying system has influence on your current state, that much data is required.
Rob Stevenson: Welcome to How AI Happens, a podcast where experts explain their work at the cutting edge of artificial intelligence.
Rob Stevenson: You'll hear from AI researchers, data scientists.
Rob Stevenson: And machine learning engineers as they get technical about the most exciting developments in their field and the challenges they're facing along the way. I'm your host, Rob Stevenson, and we're about to learn How AI Happens.
Rob Stevenson: How AI Happens is back in action. Hello, podcast land. How are you today? I'm so glad you are all tuning in for another classic installment of what I hope is your favorite AI ML podcast. I have another wonderful guest for you. He's had a ton of experience in our space. Currently he serves as the data science and statistical expertise leader for the nutrition and packaging center over at Unilever Co. I'm sure you all know well. Dr. Satyajit Wattamwar Wamar. Satyajit welcome to the podcast. How are you today?
Dr. Satyajit Wattamwar: Yeah, thanks Rob. I appreciate this introduction. Yes, I'm doing fine.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: Looking forward to this podcast.
Rob Stevenson: Me too. Because Unilever kind of makes everything. Everything from nutritional supplements to pens. I'm sure anyone listening to this could just sort of reach around their desk or their space and probably find something that Unilever makes. And it doesn't just happen by accident. There is all of this, amazing insight and data that goes behind this product development which you directly oversee. And that's why I was excited to speak to you. So we'll get into that. I don't want to get ahead of ourselves here. Let get to know you a little bit and then we'll earn the insight. Yeah, I would love it if you wouldn't mind sharing some of your background, maybe some of your PhD work if you feel so inclined and then how you kind of explain your current role.
Dr. Satyajit Wattamwar: Okay, yeah, sounds good. So my background is chemical engineering and then usually up to the master's degree you still are, exposed to this physics based modeling approaches or the first principle modeling that we call. But then I got an inclination to learn little bit more of the modeling and one of the good field, traditional field I would call is systems theory. Systems and control theory falls most of the time in the electrical engineering department and they look everything from the system point of view. So for example I was trying to do this data driven modeling for glass manufacturing process. It's a chemical plant or some of them were really modeling in the functioning of the heart. Some of them were doing it for the Electricity grid network.
Speaker D: Yeah, so you can imagine we all.
Dr. Satyajit Wattamwar: Were doing had different application and at the same time we were trying to contribute the core science or the algorithmic development as well. Yeah, so to be specific, now my focus was bridging the gap between the world of the first principle or physics based modeling and the data driven modelingeah. So I developed many algorithms there that kind of marries these two world, best of the both worlds. I had a startup as well for two years, after my PHSD and then later moved into the commercial role with General Electric. And then for seven years I was in the consulting field, leading various data driven modeling approaches because that also coincided with the early wave of data science when the IT world was also coming to this AI, or the data science world because there was some level of maturity in the backend operation and IT world was excited to see two opportunities at the time. One was the cloud and the power of AI. So basically was a game changer for iiety. and they upkale their game from one level to another one. So before that time you would hardly see any data scientists in those IT consulting companies. But now most of the major itult each company has greater scientists. So that's how I ended up into the world of data science.
Rob Stevenson: Yeah, yeah, yeah. Thanks for giving us the background. And you know I get to meet lots of physicists on this podcast, which is not something I expected, but it happens a lot. And I'm curious why you think there is this really common background for data folks, AI and ML folks, that is physics. Is it because math is math and the compensation is better? Is it because if you're a physicist you get sufficiently good at math but also the scientific method and conducting research which is so important also at the early stages of some of this work. What is your take on that?
Dr. Satyajit Wattamwar: Physics tries to marage these two world. Yeah, the basic sience and mathematics, that foundation of mathematics and apply it to some physically relevant problem. Yeah, that's one of the basic things in there with the physicist. But then second important thing is also, and that also had happened over the last 20, 25 years and that is towards solving this physics problem with the help of modeling and simulation approaches and the development of the algorithms. So they were very close already to this world of data science. And then the third thing is many of these physicists, they work with very large volume of data. So it is not enough for them to just know physics and mathematics, but also to have a very good know how of algorithmic science and Many of them in fact go towards the hardware learning hardware because then the best comes by understanding of the hardware as well as the algorithmic approaches.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: So take any example of understanding Skies or the climate prediction or CERN or KE or however you pronounce it in the big particle collider. Yeah. It generates terabytes and terabytes of data.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: And it's all physic is there? So they are close and slowly some of them also move and use those learning into other areas like financial science and others. So I can imagine. Yeah, you are re right. Physicists have very good affinity for this field and they are almost there already in that field. So it is just a natural gradual extension. Probably also it started paying well. So I can imagine they entered into this non physics related domains as. But then they still enjoy this same set of use of algorithms to solve the problem.
Rob Stevenson: You mentioned there. The physicists in particular have this experience with huge amounts of data and that strikes me is probably the case at Unilever. You have decades of data collection across a pretty impressive breadth of products. Is that for you like a kid in a candy shop, because there's so much data, at what point does it become overwhelming?
Dr. Satyajit Wattamwar: Very good question. Having a large volume of data which is less related with each other is a different thing than large volume of data for one problem. O for example again the same physics example, two particles colliding, it generates terabytes of data. It is one problem you want to understand versus Unilever where each product may have a limited data. again its operation has limited data. when it's being produced it has a limited data. there is a link between them, there is influence. But then if you look at together then it appears again very big data. However, I can unlock tens of values from my data at Unilever. Yeah because I can formulate tens of different use cases and each of them can be also small but still unlock some nice value versus the big one that these two particle colliding. Ah, so there the number of use cases will be fewer compared to ours. So big data is not the same in every case. It is about the context and then the industry and the value unlock it supports in that every industry and therefore at Unilever we leverage this large volume of data. But there are tens and hundreds of use cases where significant value can be unlocked even with a smaller data.
Rob Stevenson: That is really fascinating just because it confirms this burgeoning theory that I have which is that people seem to be doing more with less that it's not just about more data, bigger data, an endless arms race for who has the most. It's what can you do with it. And maybe you don't need that much to have something interesting happen. Is there a critical. I mean this is I guess sort of nebulous. But is there a critical amount? Okay. Nothing below this amount of data is meaningful.
Dr. Satyajit Wattamwar: That's really about the problem.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: So I can explain. I will give an example of a product manufacturing in let's say a.
Speaker D: Pllet plant or factories.
Dr. Satyajit Wattamwar: You can imagine we have a cloud where all our hundreds of factories data is there. And now I want to predict quality of our product that is in specific factory. I have a choice. I can look at all the data that is in our data lake in the cloud of across all the factories. But I know already.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: The operation in one factory is not influencing my operation in another factory. Yeah. So therefore even many times there are multiple lines in the same factory. ###eah, and this physical understanding of underlying setup help us to know what data that we should be looking at.
Speaker D: Ye.
Dr. Satyajit Wattamwar: Rather than doing a brute force data crunching that and most of the time you will not believe that more data is not always leading into good quality model. Almost. I will disagree. Unless it is for the same use case.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: Simply like this example. The other factor is not influencing my factory where I want to model it. So that data is totally useless for me if I use that data to model my product quality. In fact I will have poorer models simply because my models will be an average between these two factory data. So what is relevant data? What is the context of that data is more important. And then if I'm really modeling that again the quality. I will go back to my chemical principals now.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: What is the time constant of my process? Time constant is again something defin in the chemical terms.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: So that tells how much data you would need because then you will capture all the relevant dynamics of your process. But that's about the operation now coming back into the formulation area where you mix up something in the laboratory.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: And then you are measuring something. So the time constant will be different there. So how much data will be required. It is on the time constant. That's. But then that is giving. Making use of the understanding of chemical principle another domain. It will be different.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: So I would usually say the history of the data that really has influence is required. Beyond that is not required.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: Unfortunately in other fields this idea of residence time or time constant is not very Effectively used. But you can imagine also in the financial world.
Speaker D: Yah.
Dr. Satyajit Wattamwar: When the fade or all the government to the financial bodies, they change the interest rate. You can imagine how much time does it take towards influencing that.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: And then that amount of data will be useful and to adjust those interest rate to influence the economy. They will not look at the data from last 50 years because that's not the use case.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: They want to influence the economy now. And that has been the dynamics that had happened over the last three months, six months. I don't know. I'm not expert in the financial system, but the idea applies. So if you really know what period of history of your underlying system has influence on your current state, that much data is required. Yeah. And then second question would come how fast you should be sampling and things like that. But that's the second part.
Rob Stevenson: That is a great answer and it's not the one I expected, but it makes perfect sense in terms of. Okay, we must work in the context of the problem that we're experiencing. And this notion of making a time constraint, you say it's not utilized in lots of areas. Is that merely due to a lack of expertise? Why do you think that's the case?
Dr. Satyajit Wattamwar: Yeah, I would say this world of data science has been evolving. It's very relatively new in the sense it's marrying these various ideas which I explained earlier.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: So therefore many times the experts in one field are not fully aware or they don't leverage the things from there. And so I will just. It may sound a bad example, but let's say computer science.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: Idea there is as much data that we have, the better models I will get.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: But then the physical understanding usually is not embedded into the understanding of the problem. Yeah. And then the idea goes more data, the better it is. But that's not correct. So I think there is nice opportunity and this will evolve because you will see many electrical engineer, chemical engineer, mechanical engineer, they all are entering into the world of data science. So they are marrying these two things, the understanding from the classical field and leveraging the power of algorithmic algorithms in the world of data science. So again, I will give another example. In the computer science world there are now special this deep learning and other some optimization algorithm like a reinforcement learning. Extremely popular. But if you really go back into the world of mathematics, you will see there are hundreds of algorithms. Not just they belong to this class, but many other class and they all leverage the system point of view of modeling, which is classical system point of view modeling. Yeah, so more and more these things are happening and in fact things are exploding lately and it is becoming, to be honest, bit difficult even for the scientists to kind of get their heads around of all these new developments.
Rob Stevenson: Right, right, okay, yeah, that makes sense. Now Satyia, when I began this show I did not anticipate how many data science folks I'd be speaking to, which perhaps is a common that portrays vast ignorance on my part. The reason I did of course is because it's hard to get technical about AI without at some point speaking about the data and the data operations underneath it or alongside it. And so I was hoping you could kind of explain what is your connection to some of the AI subject matter experts and how you partner with those teams to layer in AI over at.
Dr. Satyajit Wattamwar: Unilever as a company Natur, we are very strong pool of data scientists but then I can say we are more closer to the problem. We understand the algorithms and things like that. But then we are still more leveraging the traditional data science approaches. Yeah, so for example there are now this deep learning or the gen AI approaches so they are relatively new. Of course we have few scientists there but then we see where we have gaps and then we connect with the outside world and in the outside world we have very strong network of partners. Yeah especially many companies who have strength in various areas. So we build an ecosystem of partners. Many of them have strength in the traditional approaches, many of them have strength in the consulting, Some of them have a nice strength in the world of computer science, some of them have very nice strength in the world of gen on the others. So we try to bring in this awareness and learning inside and not just M with respect to the know how but also the products Y so many times this new developments leads into new product innovation and depending on their maturity we also bring them. So I think we are connected and then wherever there are gaps then we try to fill those gaps. But the world is evolving so fast I must admit and yeah, so a big job although the AI is helping us to improve our efficiency of day to day operation and that we are benefiting from that. But because there is so many developments ongoing that keeps us busy just to stay up to dateors up to the state of art in the world of AI.
Rob Stevenson: This is, what is the word not puzzling, this is what is so challenging I suppose about just the AI question is that folks in your position and many I think are on one hand you are utilizing AI to make your own job better and on the other hand you are building it for the company There is a sandwich approach of you are getting it from both sides so I would love to know just well maybe we'll start with one hand and then go to the other in what ways are you leveraging it to make your job easier?
Dr. Satyajit Wattamwar: Usually I bucket usability of AI into two major groups or two major value unlo logks or two major buckets what you recall one is operational efficiency improvement and the second one is really goes bit uncommon is innovation so only those companies either any big company or small company who is in the business of R and D or the innovation they are leveraging the second bucket the first bucket is more on the operational efficiency is what most of the bu of gen AI is around yeah and then naturally the first one is kind of a stronghold of IT software company and consulting companies yeah like for example all the co pilots from Microsoft or other companies or large language models so all of them supports efficiency increase and then we have been using them a lot Ye and that's also over last two years onlyeah like I don't imagine a day when I don't use the copilotseah M and it's not just about meeting summarizing and things like that but as part of this learning and solving various problem I have also many questions yeah and then one of the first advisor for me are these SCOP pilots yeah so of course we all know they may not be fully foolprof but that gives me an idea even my data science question when I need to formulate something yeah maybe a set of requirements I have to collect my thoughts yeah and the takes time but then the starting point can be coiled as well Ye so it is I call it operational efficiency again as a result of this gen AI tools and now the innovation yeah so the innovation is as I said earlier is really specific to those who are in that business yeah you can imagine a car manufacturing coming up with a more aerodynamic shape to have less fuel consumption or maybe the batteries y how to expand the battery lifeeah all these are innovations and there there are niche gen AI approaches and niche gen AI models Ye and they are playing a big role there so we as a company we are leveraging both and for me as a personally I use the first one to take care of my own daily work but the second one is my core business which I leverage and help my company Unilever to develop a new product superior product.
Rob Stevenson: And better products it's such a reasonable way to split Things up, innovation versus efficiency. Just because right now I feel like we are in the stage of the hype cycle where the people who would be responsible for investing, continuing to invest are really pushing for roi for pushing for like, okay, is this actually working? Is this helping us? Is this saving us money or is it making us money? Is it saving us time? And so to put it in terms of time saved as an example of copilot is something that makes it easier for a cfo, for example to sign off on. And then you have this, the innovation, the more research side this is a little squishy. You have to be thinking in a little more long term of a fashion I suppose. And the appetite for that is always big in the beginning and it just slowly wanes over time. Unfortunately. I think we've seen that over the last several years.
Speaker D: Indeed.
Dr. Satyajit Wattamwar: But then there is one nice selling point now also for the innovation and that is let's say a new product development. Yeah, if you compare traditional approach, let's say somebody wants to make a better shampoo. There is a big scientist team which has to be at the cutting edge of the research. They have to follow the conferences, journal papers and they need to be up to date with the research. And then the experimental scientist then goes to the production and other areas. Now in the innovation, the gen AI models, the scientific models that can allow you to extract relevant information from thousands of literature and also to experiment the s in silico experiment with millions of combinations.
Speaker D: Ye.
Dr. Satyajit Wattamwar: And the same also applies in silica experimentation for what happens in the factory. So the now all life cycle has reduced significantly and at the same time you have scan far bigger spectrum. So therefore there is now even the innovation in my opinion is getting more support. So innovation. So Gen AI exploration activity in the innovation space is also getting nice support but I think it is going slowly in terms of the larger companies like a Unilever or major other companies who have significant R and D presence and a significant investment in R and D. They are leveraging it more. But I think that will be the norm. Yeah, and for the same reason with all the companies that we are working in this innovation space using this gen A or AI approaches, to be honest, they are doing very, very good.
Rob Stevenson: Yeah, that is very well put. And you know Satyia, I don't think we're going to find a better bookend than that this episode because we are of course here creeping up on optimal podcast length. But before I let you go though, I do want to ask you to share Some wit and wisdom here at the end. For the folks out there in podcast land. For people who are seeing someone in a role like yours at a company as large and impactful as Unilever, what advice would you give them so they can forge their career in such a way?
Dr. Satyajit Wattamwar: Oh yeah, there are actually many things.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: So it depends on what you are interested in.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: Ah, that I strongly put. So if somebody is really looking for, let's say growing in the career ladder, then it's not justifying one's own interest.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: So in fact I recently read there is current age is not only about climbing the ladder but climbing the knowledge ladder.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: Not the hierarchical ladder in the organization. That's very important.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: Because it's very clear many mundane jobs will be taken care by the year and the later robotics. even the white Collor or the service industry job. What will be valued is your knowledge, experience as well as the wisdom and even more important your ability to connect collaboratively with others as a team. So even I see very senior leader in our organization who are senior than me. They spend so much effort to understand the technology. They. They are not coming even closer to my background. They are coming far to different background. But the amount of effort they put to understand this technology is tremendous. So those who are aspiring to make a career depending on their stage of the career, there are various important thing that you need to consider. Of course those who are at early stages, they need to learn about the scills.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: And then get hold of the various understanding of the algorithm, the implementation and various knowledge from various domains.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: So I see a future where both are required unfortunately.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: Where you need to have a certain vertical specific knowledge and understanding of that specific business. Even like in the FMCG or cpg we would be operating different than our counterpart. Our databases are different. Our operation is different.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: Although at somebody who is looking from outside. Oh yeah. They are making same thing. But unless one has know how of one's own specific business, the applicability of the technology is less there.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: So therefore one needs to be also with patience y understand the business that you are engaging and have this other lake in the technology field.
Speaker D: Yeah.
Dr. Satyajit Wattamwar: To unlock some other value there and slowly in my opinion then only one can grow in one's career. Unfortunately I would see that going into another vertical will become more difficult simply because this know how of a specific business operation will be even more critical going forward. Yeah. Because you are supposed to create new value and that is possible. Only after understanding the underlying opportunities or the problem in your organization. Yeah, but then one has to keep both the activities in parallel. Understand the business, try to get hold of the data in your organization. At the same time, continue learning the technology.
Rob Stevenson: It is great advice. The only thing I would disagree with is that it's unfortunate. I think it's, exciting and it's a chance to use your whole brain and this is how it is. This is constant adaptation and development. So the advice to upskill constantly and understand the business is well taken and well given. Satyya this has been great having you on. I wish we could keep going, but I guess that just means I'll have to have you on for part two. So Dr. Satyaj Wajamamar, thank you for being here. I really love chatting with you today.
Dr. Satyajit Wattamwar: Yeah, thanks a lot Rob. I really enjoyed this podcast and yeah, I wish all the listeners a a very nice time afterwards. Thanks.
Rob Stevenson: How AI Happens is brought to you by Sama Sama's Agile data labeling and model evaluation solutions help enterprise companies maximize the return on investment for generative AI, LLM and computer vision models across retail, finance, automotive and many other industries. For more information, head to sama.com