 Hi, this is George Gilbert. We're on the ground of the Data Science Summit in Mariette Marquis, San Francisco with Sounder Srinivasan from Robert Bosch, makers of all cool things that power our favorite German cars. And American cars as well. And American cars. There's no car without Bosch. Okay. Sounder, so tell us about the customer journey. Your journey in applying machine learning first within Bosch and then further out and give us a sense for some of the hurdles and that you had to overcome in the maturity in terms of skill set and the technology that had to grow along with it. Sure. Yeah. We have been very fortunate to have a large internal marketplace, if you will, to bootstrap and then firm all our machine learning operations. We are one of the largest manufacturing companies in the world. And as such, we use that as a first starting test bed. There's been a lot of data that is already collected in our manufacturing operations for primarily forensic purposes. Now we have been using that to improve quality and reduced or improve operational efficiency. So give us an example of how the like data that you might have already been collecting that you could repurpose and how you would put in a machine learning to use to improve quality. Sure. So one of the examples I'm going to be talking about at a conference later today is improving what we call test time. So we had a product that goes into your mobile hydraulic devices like your Bobcats, right? So the product that gets made is an assembly line. It takes a certain amount of time. But the testing because Bosch is all about quality, we over test everything. And that was a bottleneck in producing the product, the volume that we needed. So the data that was being collected was about each step of the production process, what happens in each step, what are the components that go in, what are the tolerances, how much time it takes. And then in the testing, what are the various test steps that were happening, how much time each test was taking and whether the test was being passed or not. So we were able to use that data to figure out what are the redundancies in the test. So some tests are covered completely by other tests. Some tests are partially covered and we use machine learning to figure out what is the minimal set of tests that you need that give you the same quality in the final testing result without having this redundancy. So in the end, would the way to measure success have been the reduction in time per unit for testing? Absolutely. So we had about 35 to 45 percent reduction in time. Okay, very impressive. So then if that was one of the first applications, when you felt the tech maturity, you know, evolved and your internal skills evolved to handle greater or more challenging problems, what were some of those? So mostly the challenges were not so much technical, but in kind of bringing together a diverse group of talents that you need in making a successful analytic application. The first is you need a good business owner who understands the value of analytics operations. So somebody who can prioritize different business problems they have and match it with the data that is available and choose the right pilot in his or her area of application. So that's the first challenge. The second is you need deep subject matter experts who can consume the results that we bring in. So one example might be, we might say, this particular analysis shows these two variables are correlated to your quality and then somebody else has to come in and say, oh, I know what this means. This means my inlet valve is open a little too high or my sensor is malfunctioning in this case. So those are the challenges that we had to overcome. Okay, so two follow-up questions on that. Why is the business owner the one who has to prioritize the impact as opposed to someone who's deeper into the data but who might collaborate with a partner on the business side to say, you know, this is not a skin deep problem? Yeah, that's a fair question. And I think it's more one of the success factors that we have seen in many operations. So data scientists is always very interested in harder problems, right? So what are the interesting statistical and computing issues that are solving that can be solved? So they prioritize problems based on those metrics. Difficult, how difficult is it? How interesting it is, as opposed to sometimes you need a very simple technical solution that solves a big business problem and that has a bigger return on investment for the organization. So it would be fair to say it's actually a balance. You like it's it might be a big win business-wise, but it might be extremely difficult technically. So you want to balance the two? Yes, we want to balance the two. So we have an internal filter or selection criteria for these projects that actually looks at both the business ROI as well as technical feasibility. And technical feasibility would consist of data availability and the difficulty of modeling that domain. Yeah, and also difficulty of putting it in operations. Let's go let's go to that. How do you what are the different ways to operationalize the models that you come up with? And so that would be sort of putting it into immediate production. And then on a longer time frame, how do you actually rethink a process? Yeah, both great questions. So on the more immediate what we use is a paradigm of predictive modeling markup language to decouple the data science and the model creation from the actual operational deployment. So that makes it very easy to put it in an IT infrastructure that is already in place. So you minimize the lead time between the tools that you use to develop your solutions from the infrastructure that is already in place in the operations of it. So that's how we quickly bring the solution into production. And in terms of the longer term cycles, what we observe is this then is an engagement or an entry point into engagement with the engineering department where we can have conversations like okay, how do we now design the next generation of the both the product itself, as well as how the product gets made. Yes, exactly. Okay, so now you had told me when we were talking earlier that the next step is to take this beyond the four walls of Bosch. What? What are some of the impediments to doing that technical and non technical? And what are some of the potential benefits? So what we see when we go outside, it's also quite similar when we are inside is people are not looking for technology. They're looking for solutions. So what we need to do is to go from building blocks of their solutions into creating a complete solution that is mature enough that most people can use it out of the box. So that's the I would say the biggest challenge. And how do you identify what the the internally you had a business user who could say, This is my biggest impact. How do you identify the biggest impact business impact problems for external customers or partners? Well, some of it, fortunately, we have a lot of access because we do work with a large customer base. A lot of our customers are B2B customers. And we have had such long good relationship over time that they are very happy to share what are their pain points or what are their business needs in this area. Others you can get through in the standard industry surveys or opinions that you know, the usual marketing channels. And so that's pretty much how we can identify those opportunities and find the similarities between what we do internally within our four walls, and also the external market. Okay, so have you guys started to tackle any of the inter sort of the B2B applications? Or is that very, very early days? I would say it's in the early stages. But our internet of things applications are already out there. There have been few successful pilots in many different areas. And those are the first venues in which we are injecting data science solutions into them. Tell us can you tell us a little bit about what those internet of things apps look like today and what they might look like a little further out. So across all our verticals, right? So whether it's automotive, energy, industrial or consumer goods, these are the we start with sensors that Bosch makes. Bosch is one of the largest sensor men's sensors manufacturing companies in the world. And they are all now IP connected. So they send data into either internally to the client, or externally to a cloud storage, where we can now inject data science applications, either to improve efficiency is one thing. Maintenance is a huge second problem that we have successfully tackled. In one instance, for a customer, what we have done is reduce revisits of both the technician as well as the breakdown times of their end customers. Okay, very interesting to be continued. Yes. So this is George Gilbert again on the ground at Data Science Summit San Francisco at the Marriott Marquis, and we'll be back shortly.