Episode 5: From Data to Action: The Evolving Role of Data Science & AI in Supply Chain

0:00:07.9 Shashank Garg: Hello everyone and welcome to the Intelligent Leader Podcast. I’m your host, Shashank Garg. Joining us today is Prateek Shrivastava, a seasoned data scientist with deep expertise in data analytics and end-to-end data delivery. He currently serves as a data scientist and analytics manager at Cummins where he plays a key role in leveraging advanced data-driven insights to enhance supply chain operations. Prior to this, he was with 84.51°, a leading retail data science company, supporting Kroger in creating personalized shopping experiences. Prateek holds a master’s in Information Systems from the University of Cincinnati. He’s based in Cincinnati today. He just talked about the temperatures that you don’t want to know and has a strong background in developing and delivering large-scale analytics projects. Prateek, welcome to the show.

0:01:00.0 Prateek Shrivastava: Thank you Shashank. It’s a pleasure to be here.

0:01:02.7 Shashank Garg: Great. So, let’s start with your journey Prateek, anything else that you would like to share about what led you in this field and you know, now that we know your background, what sparked your interest, what inspires you? Anything that you’d like to share about your background?.

0:01:17.0 Prateek Shrivastava: Yeah. Shashank, starting from childhood I was always very good at math and that’s really drove me towards a career in data science. So, I started my career in business intelligence. Then I moved into data science and started doing more data science stuff. Went to 84.51° which is, it’s called 84.51°. The reason is that that’s the longitude of Cincinnati and then it’s a company that came into the picture because Kroger was struggling with a lower retention rate for the customers. So, they were doing this work in the UK for a company called Tesco. Somebody from Kroger saw that work and then just emailed them saying can you do that same work here? And then they created a department, who came then NumPy USA it started doing this work for Kroger and it just kept on growing. Kroger did not see any decline for like, I think it was 17 straight quarters after that once they started using data. And KD4 film was really a pioneer in that regard because this is like the early 2000s, we are talking about. The data was not big, there was no big data, it was like all small data by that time.

0:02:28.4 Prateek Shrivastava: So, we built a lot of marketing analytics solutions. And back when I joined like we were building on a solution to identify those customers and serve them with the best possible coupons that we can get from our CPG Partners. So that’s what I worked on and there I moved into supply chain. So COVID hit suddenly, and we saw that our pickup numbers, pickup orders, were just growing through the roof. We used to see four, or five orders in a day at a particular Kroger store. But suddenly there are like hundreds of pickup orders. We had to convert a lot of those Kroger stores into dark stores where you cannot go and buy stuff. It’s just optimized for pickup orders. So, we rearranged every bit of it to optimize the travel time between picking up those items. That’s where all my supply chain experience came into the picture. Finally, using that experience, I came to Cummins. And Cummins, I have been able to use that supply chain experience in a variety of different ways.

0:03:26.6 Shashank Garg: Yep, great to know, great to know your background. Coming to supply chain, as you mentioned, and both in CPG retail manufacturing, given the inherent complexity of supply chains, multiple geographies, stakeholders, and systems, I’m sure there are challenges that you saw both in your earlier experience at Kroger and now at Cummins. So, what exactly? If you can talk a little bit about some of the use cases that you think have been very impactful and how you’ve addressed those challenges.

0:04:03.5 Prateek Shrivastava: Yeah, so definitely. The first one is with Kroger. So, we built a pickup order forecasting model for them. The main challenge was there were no data. So, when you are building these forecasting models you’re going all the way to two, three years back, several years back, in order to understand how many orders you might get. But when you have like five orders one day and a thousand order the next day, you can’t really do that. So, at that point in time the biggest challenge was, and that’s some of the learning that I got through playing with the data is that there are a lot of external parameters that can affect all of these things, so your model needs to be very robust to accommodate those things. So, for COVID-19, we started looking at weather parameters. We started looking at like county level parameters. So, there was a New York Times data set that was used to publish at that point in time. It would tell you like every county, how many Covid cases are there. And that was having an effect on how people are shopping. So, we needed to add those parameters.

0:05:10.9 Prateek Shrivastava: So, once we were able to build a somewhat robust model, I mean, we couldn’t do it on day one. We still had one, or two months of data. After that, once we had that data available, then we started using all those parameters and were able to solve that. When I came to Cummins, the first product project I worked on was to predict when a truck might fail. Now in this project, the bigger problem is that there is a lot of data. So, when I say there is a lot of data, is that because we have those telematic sensors installed in all those engines and then they’re sending up streaming data. Now that streaming data comes even when the truck is just running fine and there is no problem in the truck. So, parsing that data, understand what data might be important in this tons of data. So that was very important. And then mapping it with actual failures in order to create the testing data, to create, to actually train your model.

0:06:10.4 Prateek Shrivastava: So that part was very difficult, but we were able to manage that. Once we were able to manage that, we were able to create a product that helped us save a lot in our supply chain optimization. So, if a truck is going to fail, and if we see that a lot of trucks are going to fail at some point in time, then we can get those parts that might need replacement before something like this could happen. So that helps us in optimizing a lot of things. Those are two very relevant and valid. I would trade in a lot of data for any day for no data. In the field of data science, there are just far too many data science projects that we saw earlier, but not today. I think that’s a problem of the past. I think in most cases now you can always get internal, good quality internal or in some cases external data. The challenge in parsing and recognizing what’s relevant and what parameters to track, always remain and that’s where the core expertise of a data scientist comes in.

0:07:15.2 Shashank Garg: Yeah. At Infocepts, I remember we helped a national home improvement retailer, they were really struggling with inventory inaccuracies leading to frequent stockouts. And what we realized is they had not built in, they’re built in seasonality, but it wasn’t multi-level. So, you know, it took many iterations and then calling out the anomalies. But then eventually, in the earlier avatar, it would just have a lot of false inventory alarms. And then with a combination of humans and AI, you’re now finally at a point where you’ve dropped those alarms at 50%, and you’ve reduced the stockouts by close to zero, almost close to zero now.

0:08:03.5 Prateek Shrivastava: That is a problem with every industry. Like because when you have all these false alarms, then people lose trust in AI systems. So, building solutions where you constantly need to look for false positives and constantly need to reduce those things. So, I think that once you get that kind of confidence in your model only that you’ll be successful.

0:08:27.6 Shashank Garg: Another very interesting piece of work we saw just very recently, optimizing inventory allocation across stores. And again, this is a very age-old problem, man. People, merchandisers used to do it manually, you know, 20, 30 years ago then they were AI systems, who did this based on certain patterns or certain order history. But with the power of AI, you can go through 23 different parameters, and you can change those parameters through regular maintenance on what’s relevant to that model. Like you talked about the whole Covid outbreak and what it did to businesses, to be able to quickly respond and change your models so that you are agile in your approaches and the product distribution across stores can be optimized, can hugely improve your inventory turnover and operational efficiencies and sort of reduce the excess thought levels. Yeah, again, but yeah, you’re right, ensuring you’ve got the right data, ensuring there is a human in the loop with all these AI systems so that you’re always circling back to remove those false alarms and ensuring that we can all trust these systems.

0:09:40.3 Prateek Shrivastava: Yeah. Another thing that I probably like to add here is that sometimes it’s better if you put your AI system in the business processes. So, for example, a lot of times what we have done is that we have created dashboards that are going to the user, but then the user is looking at it and not taking any action. Sometimes when you have like very high confidence and then you can like really say that this is something that is working out, then when you need to put it, put it back into the business processes, by that I mean is that if the inventory is getting low, don’t wait for somebody to order it, just order it yourself, so that kind of thing. We have seen at Cummins that when we do, it helps with two particular things. One thing is that the business doesn’t need to change because the process is going to the business and that helps us a lot with change management. And then the efficiency of the system just increased significantly because of that.

0:10:36.8 Shashank Garg: I think that’s a great point. Putting it back into the business, the actions are either initiated automatically or with the business owner’s approval so that their work reduces and it’s not like they have to look at here and then implement it here. Great point. And I’m sure the example that you talked about where ordering parts, faulty parts for ensuring that the trucks keep moving must be a great example there, because unless you do it and if you just wait, your worst case is, you know, order something that will sit, but it’s still better than your other worst case where the truck will halt.

0:11:14.3 Prateek Shrivastava: Yeah, yeah. And then the cost of doing that is a lot higher. I mean, if you let the truck sit, it leads to like really the customer not being happy with your performance and that’s not good.

0:11:26.0 Shashank Garg: Great. Prateek, shifting the gears a little bit on the technology side, we do have a lot of audience which is fairly technology focused. So, you talked about how easy it has become over the years, over the last 10, 15 years to now do what a data scientist needs to do versus what it used to be 15 years ago. And would you like to talk about certain technology platforms that you are seeing the most success with? What’s been your experience there?

0:11:55.0 Prateek Shrivastava: Yeah. So back when I started my career, we did not have a lot of technologies to build data solutions. We were still working with SQL Server; we were still building models on your own computer. We were getting all the data, downloading it, and creating models in Spyder Anaconda. And once the model has been deployed, it would take forever for us to actually productionalize it. Because there were no systems that could do it, we would have to rely on a virtual computer that we ourselves would host. But that was back in the day. Technology has very rapidly moved from a static system to a very dynamic one that we see today. Now, if you look at it, there are tools such as Databricks. The Databricks is what we use at Cummins for most of our data science needs. What Databricks allows us to do is to have a data lake structure where all of our data is saved. And most of that data has already gone through a lot of processing and cleaning exercises. So, the data that we can rely on is saved in Databricks now. I mean, it’s saved in Azure where Databricks is the layer on top of it.

0:13:09.5 Prateek Shrivastava: And then what it allows us to do is to develop insights very quickly. We can build from a raw layer of data to a final model in a couple of weeks, which used to take months if not years in the older space. It also allows us to productionalize those models very quickly. So Databricks has integrated sets of tools that allow us to build the CI/CD pipeline for ML in Databricks itself. And at the core end of it, the users can use those tools very easily. And that’s the biggest advantage that we have seen with Databricks. The other part that I really like now is that a lot of data science has moved into AutoML-type features. Different companies use different tools for that, but Databricks has created a tool for anybody to use. And once you use that AutoML, you save yourself a lot of time as compared to manually building those models. The data science role has itself shifted from more model-centric to more business-centric, I would say. And that has happened because of these cloud-based technologies that I was talking about.

0:14:21.3 Shashank Garg: You brought up an interesting point. Can you comment on how you’re using AutoML versus manual modelling? What is the extent of your team?

0:14:31.8 Prateek Shrivastava: So, I see it in two ways. One is definitely a lot of productivity increase, the other is an accuracy increase. So, if I go back probably like five years ago and when we were building these forecasting models, what would happen is that somebody will create 10 different models. They’ll have like 10 different parameters. We’ll compare all those parameters, we’ll do a lot of model tuning, and we look at what new inputs we can get. Build your model again and again and again till you get to a good enough value and then even to refresh it, it will take forever because then there will be four other methods that you want to try out.

0:15:09.4 Shashank Garg: Yeah.

0:15:09.8 Prateek Shrivastava: With AutoML, what has happened is that first of all, it is in the same place as where your data is. So, you don’t have to migrate or move your data from one system to another system. You can just put your data all together in one place and try out all the models that are already there. It will tell you what feature… I mean, it will do a lot of feature engineering for you as well. So, the data science role has gone from building models to actually knowing what data is. So, if there are any external parameters that you think will help the system and that have increased a lot of productivity, I would say for data scientists. Now, we can churn out models very quickly and those models perform much better than what they used to do.

0:15:57.3 Shashank Garg: Also, the maintenance has become easier. Maintenance, productionization, all of that has become so much easier than what we’re used to.

0:16:03.5 Prateek Shrivastava: Exactly. Because once you put it in a production pipeline, the model can do relearning very fast. You don’t have to have a person making the model do what it wants, what it needs to do. It will just keep on doing that.

0:16:17.1 Shashank Garg: Yeah. So, we actually, you bring up a good point at Infocepts, we actually support all major platforms, including native Azure, AWS, Snowflake and Databricks, those are the four that we work with the most. And across our clients, different clients use technologies differently. We like to propose technologies for what they are good at. We continue to use Snowflake for a lot of structured data, data warehousing, and reporting and Databricks is one of those trends that I’ve seen for the platform is their ability to process both structured and unstructured data at scale.

0:16:54.8 Prateek Shrivastava: Yeah.

0:16:55.7 Shashank Garg: And just a collaborative workspace for data scientists and engineers together. So, the built-in ML capabilities, and rapid model experimentation.

0:17:04.5 Prateek Shrivastava: Yeah. One other shout-out that I have for Databricks is that they are very rapidly changing their platform. I mean, sometimes it’s good, sometimes it’s bad because sometimes the UI changes and then you hate it, but sometimes the features come in and most of those features have been tremendously useful for us. Even just today I had a call with Databricks and then they were showing us the GenAI system function that they have developed. And in that GenAI, it would help business users understand their data so fast and so quickly. It takes them days to do those things in Excel right now.

0:17:41.5 Shashank Garg: Yeah.

0:17:42.0 Prateek Shrivastava: So, I think it’s going to have much more impact.

0:17:46.7 Shashank Garg: That’s great. One of the things I was going to add is that at Infocepts, we’ve actually built our own platform on top of Databricks using Databricks and Azure Power BI at the top to address two, or three more things. One is explainability. So, whenever you build a model, you create a UX on top of it to ensure that the business users can experiment, change parameters, and see what the model output is going to be. So, you’re not just relying on the data scientists, but for the executives and business users they can touch and feel the model. This is an area that I see a lot of clients struggle with; a lot of the data scientists struggle with in ensuring that the business user accepts the model. It’s the lack of trust coming from not knowing what happened beneath that a lot of our clients appreciate. Another area is to ensure compliance with your own ethical frameworks. So, the whole responsible AI dimension, we build that in that UX. So, if you have your own responsible AI framework, we can bring that in. If you don’t, we can use the industry standard responsible AI framework, and you can test all your models against that.

0:18:52.2 Shashank Garg: So as the Chief Data Officer, as the CEO, as the board, you are clear and you can always in today’s world address the query that hey, are your models compliant with the regulation and ethics? And lastly, what our clients really appreciate is a feature that we built which is sort of this business value versus level of complexity to implement kind of matrix. So, in any organization like Cummins, I’m sure you’ve got hundreds of use cases where you could solve using data science. Not everything is worth solvable because it’s too complex to change the to do the chain management sometimes or the cost to make the changes. As you were talking about this earlier, unless you bake it in the business process, data science by itself is not very useful even if you put it in production. But if the output is not integrated back into the business workflows, it’s not very useful. So just a prioritization screen where you can put in the value, hey, this is the ROI I’m expecting over this time frame. And you put in production and then you can track did good or bad.

0:19:56.3 Prateek Shrivastava: Data science should not work in silo, and you should not build a product that cannot be used by business. But at least for a company like Cummins, the solutions that we are building, need to make sure that those solutions align with our business priorities. If we build something in a silo, there is like very less chance that it will actually get implemented in production and then your effort would be just your effort.

0:20:18.5 Shashank Garg: Yeah. I know Prateek, you had mentioned that at Cummins you guys have chosen to place your data scientists, and data science teams within the business…

0:20:26.8 Prateek Shrivastava: Yes.

0:20:28.2 Shashank Garg: Groups or segments itself, which is a great thing to do, a best practice because that helps bridge that gap so that siloed mentality can be reduced to much extent. And that’s good. And that’s something that we see across our clients. Clients who choose to do that are far more successful than ones who choose to run it in a very tight central organization with the latest changes in tariffs from the current administrations. And to what extent are you seeing that as a threat or an opportunity or how does that change the process you run right now?

0:21:06.5 Prateek Shrivastava: I think there are like two, three different aspects to that as well. We get a lot of products from Mexico and from Canada and a lot of plants are in Mexico. We have plants. Cummins is in 185 countries. So, there are a lot of supply chain issues that could arise with tariffs. So, from a business point of view, what we are looking at is that we are trying to have less dependence on one particular supplier or one particular country so that we are a more diverse organization in case something happens, if the tariff comes from one country, then we can always move on to the other one. So, we are building more robust supply chains in order to handle that. From a data science perspective, what we are doing is that we are doing a lot of scenario-based analysis. So, we are trying to make digital duplicates of our supply chain. So, where if something happens to one, I mean suddenly tariff goes up for one thing, then how we can robustly move suppliers to a different place. So, we are trying to build all those scenarios and there are like lots of dashboards that we are building which are doing those vortex scenarios that this happens then where do we lean on to? So that is one of the key things.

0:22:22.2 Prateek Shrivastava: And the other thing is that our future models, we are trying to make it more robust where we are putting more economic data into the model itself, so that when calculating those things, it can able to guide our business where we should ship. So, at Cummins what happens is like recently what we have encountered is that if you’re part of a trade group then you pay a certain amount of tariff for one thing, but if you’re not part of a trade group, you’re paying something else. And then there is a compliance aspect to it as well, where you cannot buy things from one place versus you cannot buy things from another. So, we are trying to build a comprehensive solution which can help us with this kind of stuff.

0:23:02.8 Shashank Garg: So, at Infocepts, we believe data-driven adaptability is the key to overcoming such supply chain disruptions. Our goal is to equip businesses with the right tools to navigate. In fact, we developed our own supply chain 360. It’s sort of a modular solution, you know, similar to what you were saying, a lot of dashboards and what we call decision science platform for the business leaders who take actions, for the chief procurement officers, for the CFO’s office, for the supply chain leadership, offering seamless data integration, modelling, real-time alerts, actionable insights. And we are just in the process of adding like what we were discussing the other day, Prateek tariffs and local regulation sort of API there so that the businesses can sort of navigate policy changes, cost fluctuations and compliance risks that come in from making certain buying decisions.

0:24:01.3 Shashank Garg: And today we are in an environment that we have to respect the administration’s stance on, they want to be clear about, where multinational organizations like yourselves should do or not do businesses on what terms. So, we have to be compliant. So be able to analyze the cost implications of sourcing from regions like China or other alternative markets, factoring in additional lead times maybe, additional quality control measures and also planning for potential disruptions. That’s one of our thoughts right now, very soon it will be out. One of the things that we often see is, and we touched upon this a little bit ago, that a lot of organizations struggle with scaling their data science and AI initiatives. And you’ve had a long-celebrated career, so maybe you can comment just to summarize. What lessons have you learned and what advice would you give to data leaders like yourselves who are looking to scale AI successfully across their organizations?

0:25:05.8 Prateek Shrivastava: There are a couple of points. The first thing is what I touched on before, which is like taking AI to the business, not the other way around. The other one is that a lot of times there are some quick wins available for data science leaders. So really need to show those wins in order to go to something big. You start from something big, and you fail, then you go nowhere. And then you lose the trust in that process. So definitely, it’s very important to do that. And the third one that I think is the most important is that you need to have AI, which is explainable as well as actionable. So, the model that we built for the engine failures was the simplest model that you would see. And the reason we made the simplest model is because we want it to be more explainable. We want to make sure that the business has a reason to believe in what we are saying. We don’t want to create a black box and put it in front of them and ask them to just trust it blindly. We want to make sure that the model that they are working with is robust and it is based on actual physics and science. So having that model explainable is helpful and definitely actionable so that once you get something out of that model, you are able to use it for getting insights and taking action into the business. So, I think these are like some key points that I would like to tell other supply chain leaders in order to build solutions.

0:26:28.4 Shashank Garg: Yeah, rightfully said. At Infocepts, we like to advise all our clients in three buckets, sort of infrastructure, governance and adoption, similar to what you laid out. By infrastructure, I mean you need to get the right data and cloud infra, the workbenches, as you talked about Databricks earlier or whatever that is. Because that’s just so key to setting up that easy collaboration and having consistent data. Without that, nothing works today. The second would be to focus on governance. So, data quality, model monitoring, and compliance. All three aspects. And we have a tool to do so. A lot of clients use it. And then lastly is the adoption. And that’s where, as you spoke, the explainability, end-user enablement and ensure that businesses can take actions directly. And you made a good point that you know, why the outputs back into the operational workflows so you don’t have to wait to order, the AI can just order for you if you’re going to run into inventory stock.

0:27:33.2 Prateek Shrivastava: Yeah. And that needs a lot of business buy-in. That business needs to know what you’re doing and then you need to involve them with every step of your work.

0:27:40.8 Shashank Garg: Yeah. Just looking ahead, I’m sure you keep track of the trends, innovations, and anything you want to share there either for data science or the future of supply chain management. Especially in supply chain management. Yeah.

0:27:54.2 Prateek Shrivastava: As I said earlier, I feel like data science as a whole is evolving right now. What data scientists used to do five years ago is not what they do anymore. There is a lot of rapid advancement in technology in this regard and then that’s what is happening there. What I see is a lot of rapid growth for LLM-based generative AI-type products, even though we are working on a lot of them right now. We are trying to make our customer’s life easier by giving them the platforms themselves. For example, the GenAI that we were talking about, we’re trying to make something out of that. For the supply chain, I see that there are so many different use cases that are not covered by the GenAI wave right now. So recently, I was talking to one of our directors who was, who leads the packaging department and when you’re thinking of packaging, what analytics you can do there.

0:28:54.0 Prateek Shrivastava: But then suddenly, he told me about so many things and he made me make up my mind to work there. You have a lot of compliance; everything needs to be green. Cummins has a goal of being a net zero company by 2050. For a diesel manufacturer, or engine manufacturer, net zero by 2050. That’s where we want to go towards. And then you need to reduce the packaging that you are putting in all of those things. Even there, there are several data science use cases. So, I feel like there are several new use cases that are coming in and several, the direction is very rapid. We are building towards a more automated future as a whole and that’s where data science and supply chain would go into.

0:29:39.8 Shashank Garg: Great. Thank you, Prateek. From my side, I see three things, sort of major trends that’ll shape the future. I think one that you rightfully said AI driven decision intelligence. So, AI won’t just recommend actions. I think we all will get used to hopefully after getting that business mind automatically executing some of the decisions which should make it very easy for all of us. I also think that historically, at least from the time I started working in the industry, there are a lot of traditional monolithic supply chain platforms. From a tech side, I’m seeing a rapid shift towards much more modular, API driven, what I call composable data ecosystems.

0:30:25.2 Shashank Garg: So, if you think about the example that we just spoke. We earlier were saying one of the things that you may want to consider is, is the impact of tariffs and compliance on supply chain models. And that could just be an API. It’s third-party data. There’s no need for enterprises to invest in that. So, I’m saying that as we move towards a much more composable data ecosystem, it makes life easy for everyone. And lastly, sustainability and resilience through AI. So increasingly, as you said, leveraging AI for carbon footprint, tracking the net zero goals. I think a lot of the organizations have not gone there yet. You know, everybody was super focused on inventory and cost. But I think it’s now time to go there, which makes all of us feel better about our work, and that we’re contributing in many ways to make this earth a better place as well.

0:31:25.3 Prateek Shrivastava: I agree with all those points.

0:31:26.8 Shashank Garg: So, thank you, Prateek, it was wonderful talking to you. Love the interaction, and your practical insights on what’s happening in the field of data science, data science technology and of course, the field of supply chain management. To our listeners, thank you for watching and listening to the Intelligent Leader Podcast. This is your host, Shashank Garg. Be sure to rate and review the podcast 5 stars wherever you’re listening, and we will see you again next time. Thank you.