>> Live, from Times Square, in New York City, it's theCUBE covering IBM's Change the Game: Winning with AI. Brought to you by IBM. >> Welcome back to the Big Apple, everybody. I'm Dave Vellante, and you're watching theCUBE, the leader in live tech coverage, and we're here covering a special presentation of IBM's Change the Game: Winning with AI. IBM's got an analyst event going on here at the Westin today in the theater district. They've got 50-60 analysts here. They've got a partner summit going on, and then tonight, at Terminal 5 of the West Side Highway, they've got a customer event, a lot of customers there. We've talked earlier today about the hard news. Seth Dobern is here. He's the Chief Data Officer of IBM Analytics, and he's joined by Shreesha Rao who is the Senior Manager of IT Applications at California-based Niagara Bottling. Gentlemen, welcome to theCUBE. Thanks so much for coming on. >> Thank you, Dave. >> Well, thanks Dave for having us. >> Yes, always a pleasure Seth. We've known each other for a while now. I think we met in the snowstorm in Boston, sparked something a couple years ago. >> Yep. When we were both trapped there. >> Yep, and at that time, we spent a lot of time talking about your internal role as the Chief Data Officer, working closely with Inderpal Bhandari, and you guys are doing inside of IBM. I want to talk a little bit more about your other half which is working with clients and the Data Science Elite Team, and we'll get into what you're doing with Niagara Bottling, but let's start there, in terms of that side of your role, give us the update. >> Yeah, like you said, we spent a lot of time talking about how IBM is implementing the CTO role. While we were doing that internally, I spent quite a bit of time flying around the world, talking to our clients over the last 18 months since I joined IBM, and we found a consistent theme with all the clients, in that, they needed help learning how to implement data science, AI, machine learning, whatever you want to call it, in their enterprise. There's a fundamental difference between doing these things at a university or as part of a Kaggle competition than in an enterprise, so we felt really strongly that it was important for the future of IBM that all of our clients become successful at it because what we don't want to do is we don't want in two years for them to go "Oh my God, this whole data science thing was a scam. We haven't made any money from it." And it's not because the data science thing is a scam. It's because the way they're doing it is not conducive to business, and so we set up this team we call the Data Science Elite Team, and what this team does is we sit with clients around a specific use case for 30, 60, 90 days, it's really about 3 or 4 sprints, depending on the material, the client, and how long it takes, and we help them learn through this use case, how to use Python, R, Scala in our platform obviously, because we're here to make money too, to implement these projects in their enterprise. Now, because it's written in completely open-source, if they're not happy with what the product looks like, they can take their toys and go home afterwards. It's on us to prove the value as part of this, but there's a key point here. My team is not measured on sales. They're measured on adoption of AI in the enterprise, and so it creates a different behavior for them. So they're really about "Make the enterprise successful," right, not "Sell this software." >> Yeah, compensation drives behavior. >> Yeah, yeah. >> So, at this point, I ask, "Well, do you have any examples?" so Shreesha, let's turn to you. (laughing softly) Niagara Bottling -- >> As a matter of fact, Dave, we do. (laughing) >> Yeah, so you're not a bank with a trillion dollars in assets under management. Tell us about Niagara Bottling and your role. >> Well, Niagara Bottling is the biggest private label bottled water manufacturing company in the U.S. We make bottled water for Costcos, Walmarts, major national grocery retailers. These are our customers whom we service, and as with all large customers, they're demanding, and we provide bottled water at relatively low cost and high quality. >> Yeah, so I used to have a CIO consultancy. We worked with every CIO up and down the East Coast. I always observed, really got into a lot of organizations. I was always observed that it was really the heads of Application that drove AI because they were the glue between the business and IT, and that's really where you sit in the organization, right? >> Yes. My role is to support the business and business analytics as well as I support some of the distribution technologies and planning technologies at Niagara Bottling. >> So take us the through the project if you will. What were the drivers? What were the outcomes you envisioned? And we can kind of go through the case study. >> So the current project that we leveraged IBM's help was with a stretch wrapper project. Each pallet that we produce--- we produce obviously cases of bottled water. These are stacked into pallets and then shrink wrapped or stretch wrapped with a stretch wrapper, and this project is to be able to save money by trying to optimize the amount of stretch wrap that goes around a pallet. We need to be able to maintain the structural stability of the pallet while it's transported from the manufacturing location to our customer's location where it's unwrapped and then the cases are used. >> And over breakfast we were talking. You guys produce 2833 bottles of water per second. >> Wow. (everyone laughs) >> It's enormous. The manufacturing line is a high speed manufacturing line, and we have a lights-out policy where everything runs in an automated fashion with raw materials coming in from one end and the finished goods, pallets of water, going out. It's called pellets to pallets. Pellets of plastic coming in through one end and pallets of water going out through the other end. >> Are you sitting on top of an aquifer? Or are you guys using sort of some other techniques? >> Yes, in fact, we do bore wells and extract water from the aquifer. >> Okay, so the goal was to minimize the amount of material that you used but maintain its stability? Is that right? >> Yes, during transportation, yes. So if we use too much plastic, we're not optimally, I mean, we're wasting material, and cost goes up. We produce almost 16 million pallets of water every single year, so that's a lot of shrink wrap that goes around those, so what we can save in terms of maybe 15-20% of shrink wrap costs will amount to quite a bit. >> So, how does machine learning fit into all of this? >> So, machine learning is way to understand what kind of profile, if we can measure what is happening as we wrap the pallets, whether we are wrapping it too tight or by stretching it, that results in either a conservative way of wrapping the pallets or an aggressive way of wrapping the pallets. >> I.e. too much material, right? >> Too much material is conservative, and aggressive is too little material, and so we can achieve some savings if we were to alternate between the profiles. >> So, too little material means you lose product, right? >> Yes, and there's a risk of breakage, so essentially, while the pallet is being wrapped, if you are stretching it too much there's a breakage, and then it interrupts production, so we want to try and avoid that. We want a continuous production, at the same time, we want the pallet to be stable while saving material costs. >> Okay, so you're trying to find that ideal balance, and how much variability is in there? Is it a function of distance and how many touches it has? Maybe you can share with that. >> Yes, so each pallet takes about 16-18 wraps of the stretch wrapper going around it, and that's how much material is laid out. About 250 grams of plastic that goes on there. So we're trying to optimize the gram weight which is the amount of plastic that goes around each of the pallet. >> So it's about predicting how much plastic is enough without having breakage and disrupting your line. So they had labeled data that was, "if we stretch it this much, it breaks. If we don't stretch it this much, it doesn't break, but then it was about predicting what's good enough, avoiding both of those extremes, right? >> Yes. >> So it's a truly predictive and iterative model that we've built with them. >> And, you're obviously injecting data in terms of the trip to the store as well, right? You're taking that into consideration in the model, right? >> Yeah that's mainly to make sure that the pallets are stable during transportation. >> Right. >> And that is already determined how much containment force is required when your stretch and wrap each pallet. So that's one of the variables that is measured, but the inputs and outputs are-- the input is the amount of material that is being used in terms of gram weight. We are trying to minimize that. So that's what the whole machine learning exercise was. >> And the data comes from where? Is it observation, maybe instrumented? >> Yeah, the instruments. Our stretch-wrapper machines have an ignition platform, which is a Scada platform that allows us to measure all of these variables. We would be able to get machine variable information from those machines and then be able to hopefully, one day, automate that process, so the feedback loop that says "On this profile, we've not had any breaks. We can continue," or if there have been frequent breaks on a certain profile or machine setting, then we can change that dynamically as the product is moving through the manufacturing process. >> Yeah, so think of it as, it's kind of a traditional manufacturing production line optimization and prediction problem right? It's minimizing waste, right, while maximizing the output and then throughput of the production line. When you optimize a production line, the first step is to predict what's going to go wrong, and then the next step would be to include precision optimization to say "How do we maximize? Using the constraints that the predictive models give us, how do we maximize the output of the production line?" This is not a unique situation. It's a unique material that we haven't really worked with, but they had some really good data on this material, how it behaves, and that's key, as you know, Dave, and probable most of the people watching this know, labeled data is the hardest part of doing machine learning, and building those features from that labeled data, and they had some great data for us to start with. >> Okay, so you're collecting data at the edge essentially, then you're using that to feed the models, which is running, I don't know, where's it running, your data center? Your cloud? >> Yeah, in our data center, there's an instance of DSX Local. >> Okay. >> That we stood up. Most of the data is running through that. We build the models there. And then our goal is to be able to deploy to the edge where we can complete the loop in terms of the feedback that happens. >> And iterate. (Shreesha nods) >> And DSX Local, is Data Science Experience Local? >> Yes. >> Slash Watson Studio, so they're the same thing. >> Okay now, what role did IBM and the Data Science Elite Team play? You could take us through that. >> So, as we discussed earlier, adopting data science is not that easy. It requires subject matter, expertise. It requires understanding of data science itself, the tools and techniques, and IBM brought that as a part of the Data Science Elite Team. They brought both the tools and the expertise so that we could get on that journey towards AI. >> And it's not a "do the work for them." It's a "teach to fish," and so my team sat side by side with the Niagara Bottling team, and we walked them through the process, so it's not a consulting engagement in the traditional sense. It's how do we help them learn how to do it? So it's side by side with their team. Our team sat there and walked them through it. >> For how many weeks? >> We've had about two sprints already, and we're entering the third sprint. It's been about 30-45 days between sprints. >> And you have your own data science team. >> Yes. Our team is coming up to speed using this project. They've been trained but they needed help with people who have done this, been there, and have handled some of the challenges of modeling and data science. >> So it accelerates that time to --- >> Value. >> Outcome and value and is a knowledge transfer component -- >> Yes, absolutely. >> It's occurring now, and I guess it's ongoing, right? >> Yes. The engagement is unique in the sense that IBM's team came to our factory, understood what that process, the stretch-wrap process looks like so they had an understanding of the physical process and how it's modeled with the help of the variables and understand the data science modeling piece as well. Once they know both side of the equation, they can help put the physical problem and the digital equivalent together, and then be able to correlate why things are happening with the appropriate data that supports the behavior. >> Yeah and then the constraints of the one use case and up to 90 days, there's no charge for those two. Like I said, it's paramount that our clients like Niagara know how to do this successfully in their enterprise. >> It's a freebie? >> No, it's no charge. Free makes it sound too cheap. (everybody laughs) >> But it's part of obviously a broader arrangement with buying hardware and software, or whatever it is. >> Yeah, its a strategy for us to help make sure our clients are successful, and I want it to minimize the activation energy to do that, so there's no charge, and the only requirements from the client is it's a real use case, they at least match the resources I put on the ground, and they sit with us and do things like this and act as a reference and talk about the team and our offerings and their experiences. >> So you've got to have skin in the game obviously, an IBM customer. There's got to be some commitment for some kind of business relationship. How big was the collective team for each, if you will? >> So IBM had 2-3 data scientists. (Dave takes notes) Niagara matched that, 2-3 analysts. There were some working with the machines who were familiar with the machines and others who were more familiar with the data acquisition and data modeling. >> So each of these engagements, they cost us about $250,000 all in, so they're quite an investment we're making in our clients. >> I bet. I mean, 2-3 weeks over many, many weeks of super geeks time. So you're bringing in hardcore data scientists, math wizzes, stat wiz, data hackers, developer--- >> Data viz people, yeah, the whole stack. >> And the level of skills that Niagara has? >> We've got actual employees who are responsible for production, our manufacturing analysts who help aid in troubleshooting problems. If there are breakages, they go analyze why that's happening. Now they have data to tell them what to do about it, and that's the whole journey that we are in, in trying to quantify with the help of data, and be able to connect our systems with data, systems and models that help us analyze what happened and why it happened and what to do before it happens. >> Your team must love this because they're sort of elevating their skills. They're working with rock star data scientists. >> Yes. >> And we've talked about this before. A point that was made here is that it's really important in these projects to have people acting as product owners if you will, subject matter experts, that are on the front line, that do this everyday, not just for the subject matter expertise. I'm sure there's executives that understand it, but when you're done with the model, bringing it to the floor, and talking to their peers about it, there's no better way to drive this cultural change of adopting these things and having one of your peers that you respect talk about it instead of some guy or lady sitting up in the ivory tower saying "thou shalt." >> Now you don't know the outcome yet. It's still early days, but you've got a model built that you've got confidence in, and then you can iterate that model. What's your expectation for the outcome? >> We're hoping that preliminary results help us get up the learning curve of data science and how to leverage data to be able to make decisions. So that's our idea. There are obviously optimal settings that we can use, but it's going to be a trial and error process. And through that, as we collect data, we can understand what settings are optimal and what should we be using in each of the plants. And if the plants decide, hey they have a subjective preference for one profile versus another with the data we are capturing we can measure when they deviated from what we specified. We have a lot of learning coming from the approach that we're taking. You can't control things if you don't measure it first. >> Well, your objectives are to transcend this one project and to do the same thing across. >> And to do the same thing across, yes. >> Essentially pay for it, with a quick return. That's the way to do things these days, right? >> Yes. >> You've got more narrow, small projects that'll give you a quick hit, and then leverage that expertise across the organization to drive more value. >> Yes. >> Love it. What a great story, guys. Thanks so much for coming to theCUBE and sharing. >> Thank you. >> Congratulations. You must be really excited. >> No. It's a fun project. I appreciate it. >> Thanks for having us, Dave. I appreciate it. >> Pleasure, Seth. Always great talking to you, and keep it right there everybody. You're watching theCUBE. We're live from New York City here at the Westin Hotel. cubenyc #cubenyc Check out the ibm.com/winwithai Change the Game: Winning with AI Tonight. We'll be right back after a short break. (minimal upbeat music)