 I'm Zarkmur, I'm CEO and co-founder of Kube26. Before we started Kube26, I used to work at Yahoo with the search team at the California office. And before that, I used to head R&D at an ad startup in New York. We started Kube26 almost like four years back, and we have been particularly working with the Android ecosystem for almost two years. We work with most of the leading Indian OEMs, everyone from Panasonic, Micromax, Carbon. And we do a lot of Android customization. So everything, when you boot up the phone, the lock screen to the launcher, to some of the app that comes on the customization that comes on the phone, there would be some bit of Kube26 involved. We recently raised fun from Tiger Global and Subcard, basically, to customize more of the devices that are out there and venture into different other form factor of devices. These were, I think, two signature products from the past I just wanted to show you guys. One was the look-a-way-to-pause video player integration, where we integrated a feature that if you're watching a video and somebody comes along and you look at them, the video would automatically pause. It was on Micromax Canvas 4. And then multiple products came after that. Bling Play 2 Selfie that was on Panasonic, which is like if your camera is in selfie mode and you just blink your eye, it would take a selfie. So these were some of the enhancements that we did almost like one and a half to two years back. And till now, we are almost on eight million devices that does a lot of these sort of customization. Today, I'm going to talk about app discovery problem, which is something that we do fundamentally at the company, which helps monetize a lot of the devices that are being shipped by the OEM. So as you know that hardware margins are getting thin. So the manufacturers are looking for a way to generate dollars from their device. And we as a company, we sort of come in and help this monetization process by helping user understand their behavior and then download apps from the store, which would make sense to them. Before any discovery process, you need certain ways to collect data. So on the phones that are being shipped out, some of these are, let's say, a music player that we built, a news application. There's a calendar app that's out there. So as users of using these application, we are getting some of the anonymized user data on our platform. This is like a general flow of when the data comes. So we use Kafka as a platform where multiple devices are sending data to our server. Once we collect all these logs, then Hadoop file distribution system comes along that takes this huge amount of bunch of data and pushes to our for batch processing. Spark is a platform which is most probably people who are involved in machine learning or the data science community would know. So on these batch data that we get, we run a lot of distributed processing, which is Spark is one of those platforms, which is better than the HTFS Hadoop in terms of the speed on which it can run. And after, we have run a lot of our algorithms, which is basically profiling users, things like if a user who is on certain kind of, let's say, 3G or this particular RAM or this particular model lives in this network, using this particular kind of app and downloads a certain kind of app. So we map these kinds of activity, and that's what we call profiling, and honestly, the user profile. After that, we basically use Apache high-waist to tag these new users that we feel that, which basically, if a user, a similar user, have downloaded an app, then finding out which would be the other apps that these users use. And then we use Irospy Cache to basically, in real time, push this recommendation out. So this basically goes through the flow of how we are collecting some of the anonymous data on our platform and how we are basically going in the back end and computing a lot of these kinds of recommendation and finally comes up on the platform. This is what we use internally as a collaborative filtering as a machine learning algorithm. There are two types, I mean, user-user, and there's item-item. We use the item-item collaborative filtering, which we call it internally's app app, which basically translates to, say, if Saurav Kumar, who lives in Delhi and uses, has certain properties and like uses, let's say Flipkart app, he has a tendency to use also Snapdeal app or a Mintra app. So this kind of like, you know, and if there is another user who is also like Saurav in terms of the attributes, the way we define him. And if he also has a Flipkart app, then there's a huge amount of tendency that he will download a Mintra or Snapdeal app. So we call this problem statement as an app-app recommendation problem. And generally, like I said, there are two kinds of one is user-user and item-item. So in item-item, basically, what you're saying is, and something just to note, item-item was actually released in 2003. Amazon also uses it on their platform. If you have seen, if you have bought X, you would also like Y. So this is a similar kind of recommendation algorithm that runs on their platform. So this is in general better than just purely saying user profiling and doing user-user because new models are coming in the market. So you're suddenly the way you were profiling these users suddenly changes. So you have to recompute all of them. And just the computing similarity between a Saurav and another guy is extremely computational expensive because we are running into some 1 cross 512 to 1024 attributes to define a user. So the way we do it is actually pretty simple. I mean, there are two stages. One is a model building stage where we look for similarity between pair of apps. So basically saying, if let's just say one pair is Flipkart and Mintra, then basically the profiles of users who have used Flipkart app and then the profiles of user who have used Mintra app, so Flipkart and Mintra. So that would be one of the combination. And you would compute what amount of similarity they have. And based on that, once you have computed that, you come to the recommendation stage. So we're given a user what would be and certain properties that this is the attribute. You sort of compute what would be the next app they would download. So this is what it translates. And when we do that, this is like a store that runs on most of these OEM phones. Here is like some of the apps that we are recommending to the user. We basically built it because a lot of the OEMs that we, when we started working with them, we saw that their phones were usually bloated. So there were a lot of apps that was pre-installed and that used to have almost 90% uninstalled ratios. When we built this, which is like after understanding how a user is using their phone and what kind of apps they are opening or not using or downloading, uninstalling, we sort of dropped the uninstalled rates to almost 30% to 40%, which basically the user are happily downloading the user. And then there's a retention in terms of, from the brand perspective, that the user downloading their app, they tend to use them on a longer scale. So this is just one of the examples that we are using in the mobile domain. We recently, as you know, we've also ventured into Internet of Things and Smart Bulb is one of our products. Similarly here also, we are using like, as people are using their different bulbs, bulbs with different feature, we are trying to understand what kind of apps they are using and based on these apps, we are trying to recommend like different kinds of alert for them that if you are using this particular kind of app, you might be interested in using alert, which is Uber-based alert. So this is what my talk was basically about, how we are using some of the machine learning and the data science to solve the monetization and app discovery problem, which obviously like there are a lot of companies like Google with their Play Store it's solving. What we are coming up is like, one of the Indian tech companies were working with all these OEMs and also fundamentally looking at the same problem. Obviously a lesser scale than Google, but we are also running this on almost eight million user today. So it's a fundamentally extremely big scalable problem that we are solving. So any question that you guys have, I would be happy to answer. Hi, Sandesh here. Just wanted to check, because you are working with a lot of OEMs and a lot of ad platforms are trying to solve this problem on there and as you rightly said, how different is your approach going to be compared to how the other ad platforms are trying to solve this? Or even apps themselves are trying to solve this using the data that they have. So are you saying how are we different from an ad platform that are trying to solve this problem? Yes, in terms of the approach, because ad platforms would have different set of signals, the app in itself, they would have different set of signals they would rely on and how different is your approach going to be because you're directly working with the OEMs. So if you look at ad platform, so either they are integrated on your different webpages or they're inside an app. So the limited understanding of a user that they will have, it will be obviously far too limited than us who exist on their operating system level. So we have a far-reached understanding of user data, not just from what their consumption is on within my app, but also outside what kind of model is running on. If he's on Wi-Fi, if he's on 3G, how much data he's consuming, his internal storage space, which network is on. So when you're on an OS level, so you have a far better data about user and obviously when they're saying if you have more data, then you can recommend better, but then obviously there are trade-offs as well. I mean we understand user fundamentally across like his uses through different apps and if you're just focused on one app, so inside that is if you're inside a news app, so you're consuming different kinds of news content. But so that will give you some understanding of the user, but in general to understand about the spending powers or where is located tier one, tier two. There are a lot of other data that also comes along that helps you discover the different content and services. How do you plan to pass these insights to the app makers if at all you're planning to do that? So insights would be, I mean probably like, I don't know, within your app, if you're trying to solve this recommendation problem, so that was just an insight of like how if you would want to solve this recommendation problem at a scale, then what are the different components that you should be looking at and what are some of the machine learning problems that are there, I mean it's basically algorithms that are there that can help you solve it. So that was just a preview into it and we are extremely like startup and developer friendly, so if there are any questions that you guys think that we can help, so just write us an email on infoatcube26.com and we would be very happy to like work along and find something interesting. Thank you. There's tea break up next and we'll.