 Thank you everybody for coming in. My name is Piyagudish Suresh Kumar, Vice President of Engineering from Zopper. I just want to have a quick show of hands before I start. Who here have known Zopper before this event today? If you could show your raise your hands and say that you know Zopper before you come in, I see few go up there and few over here. Okay, that's kind of what I expected actually. So, I think Zopper, you just recently opened a center in Bangalore, trying to build an R&D center here and also sales and marketing team that takes care of the southern region. There are 175 strong employees company, most of them are in Noida and there are like, we have like presence in all the cities that we have launched so far. I just wanted to let you guys know that I'm two weeks into this job. I've been previously in the corporate, been working in the corporate world for many years and a couple of startups in the Bay area. Before I moved to Bangalore, relocated for good here and wanted to participate in Zopper to make a change, like a social change, I'll tell you a little bit when I get into what we are doing in Zopper and how that's change is going to come about. And I'm very upbeat about it and looking forward to participate in that. So, the structure of today's talk is going to be talking about the Zopper as a platform, how, what problems we are solving and then talk about some of the interesting areas, couple of interesting areas where it's new and the new products that we're launching there. And also talk about some of the technology stack that we have, followed by a demo of two demos, in fact, showing some of the problems that we are solving in Zopper and how it's getting solved. So that will be taken care by our lead scientist, Vinay Pandey and Siddharth, who is going to come and join me later in the stage. So let's just get right in here. So the message from Zopper is that I think we want to be an e-commerce play, we want to be actually differentiated from the existing players out in the market. And e-commerce, as we know, has actually revolutionized the marketplace already. I think it's created a new business model. And what we are going to be doing here in the e-commerce play is to bring in the hyper-local dimension to it, right? And that changes a lot of equation actually in our mind. And that's the actual, in my opinion, that's an inclusive model where all of the infrastructure are the problems that the e-commerce company is trying to solve, like deliver in a short time and also take care of the price points and discovery and all that. I think, right now, all of those problems are solved by the local firms that are physically present in your neighborhood, right? Those problems are always solved, right? Why don't we tap into that particular expertise that they have already given? So we built a new e-commerce model, which is an inclusive model, where I call it as an Indian model of e-commerce to where I think in India, we have a strong local connection with the neighborhood that we live in, right? So the local community. So this picture kind of depicts that there are stores in your neighborhood where you live and use a shop, use your phone to basically do shopping, kind of know what's going on there. And it is basically offers you convenience because I think in today's world, if you're in Bangalore, how much of a traffic that we need to go to navigate through to go to stores from store A to store B, it's a hassle, right? I think we offer the convenience where you could look at the pricing of a particular product in a different store and also what kind of products that they have carry and 24 by 7, right? I think you don't have to worry about store Rs anymore. I think you can jump into the app and probably take that convenience there. e-commerce have actually offered to everybody, right? And the third point is actually quite interesting is that what I kind of referred to in the initial talk, right? Social impact, if you see that the neighborhood, right? I think the local firms, the local retailers there, they are actually what I call the lifeline of the economy, the local community economy, right? They basically give jobs in the neighborhood and also the real estate that they occupy, right? So there's a lot of social impact that we are trying to bring in as well. When you shop from us, you are actually shopping from one of those stores that in your neighborhood or in your city is going to be servicing your request and that sits coming from them. And I'll talk about a few aspects of the hyper-local. I think the typical shopper in the traditional ways before e-commerce, they go from store-y to store-b, come part of the prices and understand what they offer. And then pick on others because the products that the shoppers buy is kind of like an investment that they make because they're buying a TV or a home appliance, it's not an impulsive purchase. You just wake up in the morning, go online and buy it. I think there's a discovery process, which is a lengthy process. In the hyper-local, the users have been used in a traditional ways to go from shop-y to shop-y, come part of the prices. And we're going to try and bring in all those aspects of this as a hyper-local facet. So as I was also saying that we are not about the impulsive purchase. It's where you just go and buy online and forget about after the transaction is completed. We look at the whole life cycle of that experience. And product discovery, as I mentioned, it depends upon the product. You just like, if you buy clothes, probably just wake up in the morning, buy a couple of t-shirts, that's fine. But I think if you're buying a TV or a home appliance, like washing machines and refrigerators, it's a lengthy process. I think the discovery process is very unique and hyper-local. And these are investments. The purchases we look at as investments, not like transactions. There's a huge difference between investment and a transaction. So these investments need to be protected. And this protection will be going to be not just protection, but also that covers a lot of warranty and services, but also the delivery and installation of these kind of products, essentially. It's not simple. You need technicians come in and do that. So that's the local connection. Basically, you buy it from a local store, you know that that's going to be taken care of, or at least you know that you can go and ask the particular person and say, I want these kind of services, you know, come and deliver, I'm right now available, just call up the guy and say, you know, do you want to come now? And they come and do the installation and set up or whatever. I think that connection is very, very unique in e-commerce, you know, in hyper-local space, which the e-commerce do not offer. Right? And warranty and services is basically, these investments need to be protected. And warranty is basically the warranty given by the manufacturer. But I think we have the extended warranty that covers, you know, beyond that period, like two or three years period, where we're going to introduce new products that are called Sopara sure, that takes care of much of your hassles going through your customer service support. I'll talk about that when I get that slide later in the talk. And then the lifecycle of the product, basically, you know, doesn't end with the services, but also, you know, how do you recycle it? You know, once it gets old, how do you upgrade that to a new version? So you just need to exchange that to a different, you know, new product. So we also come in the local, you know, market, you know, local firms are already doing it, retailers. So we're going to tap into that kind of expertise as well and make sure that, you know, exchanges are also taken care. So you get into a new product cycle all over again. Right? So we look at this hyper-local as a wholesome, you know, into an ecosystem rather than a particular segment of it, just not just delivery or not just, you know, discovery. We look at it from end to end perspective. And product discovery, I want to take, like, two out of that cycle, right? Product discovery is one of the biggest problems that we're trying to solve. And the second aspect of it is, as I mentioned, about the warranty and services. So I'm going to take these big two items and, you know, our elite scientist here would come in and, you know, show some of the problems and how they're solving it. You know, the challenge that we have here is very unique because of the hyper-local, you know, aspect of it. So we have sellers of all different sizes, right, NOS, from their organized, their tech savvy, some of them, you know, some of them may not be, you know, I think we need to build technology for all of them because we are asset-like category. We don't carry any kind of inventory ourselves. It's very easy for e-commerce players to just have inventory, have build a website and sell it online, right? I think we need to build the inventory, right? And its inventory is going to be different from a location to another location, right? Because what's trending here, what's being, you know, bought over here in this area will be entirely different from what's being bought in Mumbai or in Delhi, for example, right? I think, you know, that local connection, that local sort of, you know, knowledge, I think, you know, we need to bring in when we talk about hyper-local. And, you know, our biggest, you know, challenges is to, you know, build technology that works for all of these sellers, where they can come and publish their inventory that can be sold in our marketplace and also the price information, right? So these get fed into our system and we built up the listing based on your location. You know, we kind of come up with the, based on the trending in your area, based on what's going on, you know, what are the sub-buying, what price are the sub-buying. So you get to know all of these information and, you know, we'll probably build you an experience that, you know, e-commerce companies really offer. So behind the scenes, it's entirely different from what others do and what we do. So, you know, I just want to make sure that, you know, product listing is actually one aspect of the discovery. You know, listing is number one and then after you list, their discovery takes place in our app, right? So our app is basically the entry point, you know, today we are only app only. We launched in Android Market. We recently launched in iOS a few weeks ago. So we're going to have Android and iOS apps so you can download and install and, you know, play around with it and, you know, your suggestions or, you know, your purchase. We're looking for your purchases from a local community. So, you know, discovery is something that, you know, I think some of the challenges that you're solving there is, is going to be, there's going to be a demo that Vinay is going to be showing a little later. The next product, the problem, that problem area that you're trying to solve, you know, is the warranty and services area. Like Zopper, Azure is a new product that you're launching, all right, that's been launched. You know, we're going to come up with more extensive coverage for all the products that you buy. And, you know, most of the things that we run into in the past is that, you know, when you want to claim warranty, you typically, most of them won't have the receipts in order, you know, want to even know what modeler make that you purchased or value bought it from or what price you bought it from. All of this information is kind of like, kind of in a clutter or you need to sweep them through. And once you get hold of all this information, you need to call the customer support and I know how much of a pain it is to go through that. And we're going to introduce one-touch claims service where, you know, you would use the app, you know, what all products you purchase, you know, either in our app or in our local community as well. So, we're going to be local store as well. If you know, if you purchase that, we're going to be able to do that one-touch, you know, servicing, no documentation and we're going to give a piece of mine absolutely free, right. And coming to the technology part, these are the stacks that some of the, you know, stacks that we use to solve some of the problems. And this keeps changing, then the list keeps going actually. I could just put some of them here. Some of them are actually, you know, there are some more that we are working on. I think this pretty much covers most part, I guess. That's it for me and I would like, you know, when I point A and so that's to come over here and, you know, before Wiley is coming and setting up here, probably I can ask a couple of questions and we can, you know, give us some goodies if somebody answers some, yeah. So, I thought of taking questions later on, but I think if you have questions, that's fine. I thought, you know, I could run some quiz program. Okay, it's okay, yeah. That's a great question. I think we, what we, you know, I just want to make sure I set the context. I think we are into consumer electronics durables right now, like not all mom-and-pop stores, but all brick-and-mortar stores who sell these goods. So the onboarding process is entirely, you know, entirely different from what typically does. So they do not come on board themselves, right? I think it's an offline process and there is a lot of hand-holding because we go through, you know, tons of checks and balances before we onboard a seller because we want to make sure that, you know, our service level is met, right? You know, I think we want to guarantee that does opera service level and also the products that they buy is, you know, an under-person trusted product, not like any China make or, you know, any other stuff, kind of stuff, right? So I think there's a lot of background, you know, onboarding process that goes on. Before we launch to a new city, there is a lot of, you know, work that goes on. So we are launched in many cities. So one of the questions I want to ask is, can you guys guess how many cities we will be in by the end of the year? Any rough guesses? I know, I've not mentioned that here. Ten by the end of the year? Not even close. 30? Yeah, you're getting there, 40. Somebody said 40, yeah. So, yeah, that's what we're targeting by the end of the year. Okay, any other questions you want to throw in? All right, and we'll keep that for people who ask questions in the end. Yeah, I'll leave the stage to, when I, Pande, when I even, you want to take this from me here? Hi, everyone, I would like to request everyone, anyone who's using like hotspot here, because it's disrupting everything. And because the PyCon India, this SSID should be working. Okay, and if you are using hotspot, it's gonna ruin everyone here. So I'm sorry, yesterday you didn't get internet. So yeah, today hopefully, if you have to use it, then please USB tether or something else. Okay, yeah, I think it's good now. So, yeah, so thanks, Berthesh, for interaction part. So, yeah, I'm Vinay Pande. I'm talking about data science perspective or the technological challenges that actually, you have to face, okay. So basically, what you think is different or challenging in hyper-local market as compared to online e-commerce orders like Flipkart. Can you, any guesses here? Okay, right, so yeah. So basically, we are hyper-local market. So it's like, we are distributed across cities, right? We have lots of cities to cover. And each city has its own set of problems, right? So when we talk about Flipkart, so Flipkart has its own warehouse. So even if you are ordering something from Delhi and its warehouse is in Bangalore, they can actually provide you using their logistics, right? We are not providing logistics. So the first challenge that we have to face is like demand versus supply. So we have to face, we have to check exactly in which area user is searching what and do we have that thing in our inventory, right? So second thing for this, actually we need a very strong catalog. But the catalog population method that we use is like we collect browsers from different vendors for their catalog. We have our own app for vendors as well where they can provide information about their catalog or they can populate their catalog there so that we can load it into our centralized shopper catalog, okay? But many times it happens that they didn't provide complete information. Sometimes the catalog is wrong. Sometimes they just provide partial information, okay? So this is one of the biggest challenge that we faced in last six months. So how we solved it using Python stack and using machine learning that I will show you a demo. And as someone here mentioned here that we have to provide customized solution for each user because each user is unique for us. His requirement is very unique. And if we are not catering that requirement in timely basis and in a proper way, then there may be a chance then he can just go to online market to buy the product, right? So yes, to solve this, so we use smart techniques here. The first thing that the first case study I'm talking about is automatic catalog population, okay? So what was the objective? So many times it happens that we got brochures from vendors but it didn't have complete information. So let's say if you are talking about, if a product is a mobile phone, if a particular seller is selling a mobile phone, then he just gave us a name of that mobile phone or model numbers. All the other attributes that is there that we have to fill it. So there are two options for this. Either we have to populate these using websites, right? So we have to go to different websites like Flipkart, Snapdeal, check there the attributes and then manually fit it into our own system. Or we can automate the system, means basically we can actually crawl the pages, then we can parse it and we can, after parsing we can process the information and solve the process. But this overall pipeline has a few challenges. First challenge is the data that we crawled is not same. It contains lots of noise actually, right? So the first issue that we have faced is how to extract valid information from noise. And the second problem that we faced here was that basically the HTML structure is continuously changing, okay? So for example, when we crawl a particular webpage one month back, its HTML architecture is different. Its nodes are different containing information. But when we crawl is today, either that node is not present there or there are some different nodes which contains the valid information, right? So we actually use ScrapI to scrap all this information and then we have a team which manually checks the nodes, valid nodes, then they actually try to extract valid nodes from HTML and then using XPAC we use to extract fields. But this is very tedious task. So for example, our iterations used to take 15 days just to crawl one, just to complete one cycle, we need 15 days. So if you want to scale and if you actually want to provide good service, then we have to improve, right? So here we try to solve a problem that can we actually automate this process of extracting information or finding valid information from page, okay? So here we have to solve two problems. First, getting the node which actually contains valid information first and then extracting that valid information in a format so that we can populate our catalog, okay? So for example, just I will give you small information. So let's say this is a page from Flipkart, okay? What we are interested is something like this where we have attributes and we want their values, right? But if you actually crawl that page and start processing, what you get is this. It contains valid information, but lot of noise, right? Lots of noise. So how to solve this, okay? So we do have lots of domain specific knowledge with us. Like we actually had a very comprehensive list of brands, comprehensive list of categories, colors and all, okay? So we decided to use that information in our favor. So we first try to use a standard. So basically we pose this problem as a classification problem first, where we actually try to find out whether this particular node is a valid node or not. We have to just classify it, right? So there are lots of standard algorithms like SVM or linear regression and all, right? So we can use that, right? To get the data, but because of noise and because of outliers, the accuracy was not that good. So after lots of, basically after lots of iterations, we decided to use weak learning algorithms. So basically the fund of weak learning algorithm or ensemble is to learn small and weak classifiers. So basically weak classifier is like they use just one rule to classify a particular node, okay? So we actually learn n number of classifiers. Then we combine them using a weighted average to get a final output, okay? So we use our domain specific knowledge for that as well as the past history, okay? So let me just give you one small demo. So after getting information, we learn lots of small classifiers and then we stored that classifier in terms of either a model like CRF model or SVM model or in terms of regular expressions, okay? So we have lots of small, small weak learners that we learned and we use them to get, so let's say we have a HTML page and we need to classify that HTML page into, so basically we need to extract valid nodes from that, right? So after running this process, each and every node is picked from the HTML page. It is passed to all the classifiers and after getting results from all the classifiers, we simply combined the result to check whether that particular node is valid or not. So for example, in this particular HTML that I showed you a few minutes back, it contains 1,250 nodes in overall, which are either recommended URLs, links, some other information, right? But out of that, only 25 nodes are valid or they contains the information which is important to us, right? So after doing that, we actually get class name, a node ID and its tag. So basically after getting this, we can automate the x path process to scrap the data, okay? And using that x path, whatever we get, we simply use it to push it to our attribute extractor, okay? So attribute extractor basically, it also has lots of classifiers, plus we predominantly use CRF, conditional random fields to actually tag different things. So this was the title, okay? And these are the attributes that we extract from here. So after solving this problem using machine learning, the process that took 15 days, it get reduced to almost 30 hours. So that means we can actually do crawling almost after every three or four days, actually, yeah? So yeah, that is the smart way how we solve problems. The other challenge or the other test case that I want to talk about is like, how can we provide a best deal to user? So before going to this, do you have any questions regarding the first demo? Yeah, so as I said, we used to do this manually, right? So because of our crawling algorithm is very good. So we have lots of pages, thousands of pages per category, sorry, per portal. And we have almost, I think, yeah, almost 250 portals in our system. So that means we have sufficient data, okay? So as I said, we used to manually get nodes first, right, X paths. So from that, actually extracting information and processing that information was very easy. So yes, our first phase was manually, right? But after that, it's everything automated. Yeah, so, okay. So basically, as I said, for example, if we are talking about a particular node, okay, so the first classifier that we learn is like, whether that particular node is contained any keyword like copyright, call us, email us, contact us, right? So it's very simple rule-based classifier, which has some certain set of keywords. So if it has that, that means that node is not valid, right? The other feature that you can use is like whether that particular node contains any specific attribute, right? So if you write a classifier for color-based attributes, so you feed all the color names to it as a valid, sorry, as a positive example, and everything else as a negative example, right? So based on that, you'll get, it is also a weak learner, basically. So based on that, you will get one more weightage, right? So after getting all the results, all you have to do is like get the weighted average. So each classifier has its own set of features here. It's not like a common set of features used throughout the process, yeah? This might sound very primitive, but for collecting the attributes instead of going to the next, I mean, the vendor, going to the source didn't work for you, or was there any limitation in that? Like these are the standard products that we are talking about, not non-standard ones. So why don't we, I mean, what was the difficulty in going to the source or the manufacturer to get these attributes? Yeah, so basically, we have typed with lots of brands. So from that, whatever data we get is really good. But if we are here, the problem that I'm talking about is regarding a particular vendor who is providing us just a browser, right? So many times, browsers doesn't contains all the information, all complete information, okay? So there are attributes which are not present there, but as per users, they are important, right? So to populate that attributes, we are using this. My question is, so the different products have different variants according to different colors, the price may vary. So how do you tackling these kind of problems? So there are different swatches like drop-down swatches, containing the size, color swatch, and all different types. So are you taking care of these or they're just like the first price on the page? So basically, we actually divided our categories into two broader part, one is hard categories, and one is soft categories. So in hard categories per se, they are the categories which has lots of things in common and they don't change that frequently. For example, mobiles, right? Or let's say tablets, right? So if you compare mobiles and tablets, they have lots of attributes common to them, right? They have primary camera, they have RAM, they have memory. So we club categories into a different subgroups and we learn classifiers for that particular subgroups only. So whenever we get a particular title, we first categorize it into a specific category and then we use classifiers for that particular classifier on that title only. So yeah. After getting the list of 25 nodes, how are you able to ignore the DOM structure and getting the exact category or exact price out of it? Because the DOM node may vary randomly on each of the product websites. So you meant to say the information that attributes that way extract? Yes, how do you get one-to-one mapping, saying category is short sleeve with a confirmation? So what is the confidence level related to it or what is the logic behind getting that? Okay. So basically after studying lots of portals, we find out that some portals, some standard portals like Flipkart or Snapdeal, they have their own unique way to write information. Okay, so if you, so for example it's, so if you see Flipkart, it follows this convention for all the products irrespective of its category. Okay. So from this, if you actually learn a pattern that if you get two attributes or two keywords there, there may be a chance that this particular keywords is a attribute and this particular keyword is a value. Right? So here we use pattern learning. Okay. Plus actually we have, we build CRF models where we feed information to it in case of if we have a description like this. Okay. So for example, here if you want to check if this particular thing is hand wash or machine wash, right? So for these type of things we simply build a CRF model where we manually tag the data and learn the CRF model. So whenever it gets something like this, it's simply labeled that particular thing as machine wash or hand wash, right? Using some label. Then based on weightage average of both rule base plus CRF model, we can get the final value. Yeah, so as I said, our system actually tries to learn these patterns or this information in each iteration. So whenever we finish our first iteration, we manually check the data. Okay, after manually checking the data we feed up it into a catalog and that information is one again used to run or relearn or models. So basically if it happens that for a particular portal, the pattern got changed. That simply means all the information coming from that portal is not valid, right? So basically instead of say slews, short slews, there may be a chance that slews short slew here and fabric short slew and fabric information here, right? So in that case, maybe we'll based on this rule base system, we'll get slews equals to fabric, right? Which is wrong? Yes, it can be a possibility, but in that case actually we have to think differently, right? But currently for these sites, these models works and if we get wrong data at that time, I think learning patterns from that or getting information using some other technologies will definitely help here. No, see, we are getting data from Flipkart, yes, but we are not selling it. We are simply utilizing the knowledge contains in that particular web page, right? We are not, see basically we are selling products but from vendors. The information that we are crawling from Flipkart is simply attributes or knowledge about that product, right? And even if Flipkart, yeah, so adding to that, actually we have, we are actually planning to tie up with lots of manufacturers, lots of brands directly, right? In that case, even if Flipkart blocks us, we simply get data from any other sources, right? So this is like automation of the manual effort that we are taking, so that we can get information as much fast as possible, right? That's the aspect here. So how do you manage your inventories like reservation? I mean just, suppose a vendor has five mobile phones in stock and I want to buy three online, three web and suppose my friend also visits this retail store and wants three, so how do you manage that? That question actually I think Siddharth or Bhagatish can answer that question. So yeah, basically we are short on time. Can I just cover my second demo first and then we can? Okay, sure, thank you very much. Yeah, okay, the next case study that we are trying to solve here is like how to provide base deal to user, okay? So objective was to basically recommend, so basically if user is searching something, there may be a chance that he has a vague idea about that particular product in his mind, right? So can we actually recommend a good product or a base product which contains that particular attributes plus some other good feature about that, right? So for example, if you are searching a mobile phone which has let's say 13 megapixel camera and 2GB RAM, right? But 8GB memory, so 8GB storage. So can we actually recommend other product which has all these features plus some good features as well, right? So we try to solve that problem using recommendation systems. So we basically combined attribute-based recommendation system plus collaborative filtering, okay? But as we are hyper-local, we have to think about a very important third feature here and that is location. So there may be a chance that what your user is searching and whichever the product is very much similar to that particular product is not present in that location. So how to solve that thing? Then what to recommend to user, right? So here a recommendation system is not only based on user's criteria, it should be based on location as well, okay? So here we build a model where we try to find out a score for each product in our app. We called it as Zopper score. So basically Zopper score is calculated based on the product features. So if a particular product has all the latest features or all the good features, it has more Zopper score, okay? And whenever a particular user is searching something, so we actually study the user's behavior throughout our app, we try to personalize it and then based on user's location, we try to give base product based on Zopper score to him, okay? So basically if a particular user is searching something, in this case a washing machine, okay? So these are the recommended products for a user in Bangalore as compared to user in Delhi. So basically this particular washing machine is 5.5 kg with same brand, okay? But for that particular price range, that particular user can actually get a six kg washing machine, okay, with the same brand and some good features, okay? So here along with 5.5 kg, we are actually recommending him six kg washing machines as well. And that recommendations is actually getting changed based on his location, okay? So yeah, here to actually achieve that, we have to think about two problems. First is personalization and second is data warehousing. So in personalization, based on user sessions and user's past history, we try to build a user persona where we try to find out what are the categories of products user are searching, what are his base categories and products and how we based on that information and by adding it with information of some other users, how can we collaborate the information and give him base results, okay? So for this, we are currently building our own data warehousing system where we try to collect data from all the sources, process it and analyze it. So yeah, that's the next thing that is in pipeline. So any question regarding this demo? No, it's not, it's basically, so basically we actually find out few features for each category or each product, okay? And we assign weight to each feature. So basically it's a linear regression type of things where for each features, we calculate the weight. We dynamically change the weighting based on the new products in the market. And from that, actually we are calculating the Zopper score. So let's say if there is a latest product in the market, it has higher Zopper score as compared to the old product, but as time pass, its Zopper score keeps on decreasing. Hyper market, no, for calculating Zopper score, actually we didn't use hyper market. For calculating Zopper score, we are simply using product features, okay? The features that we actually find out for a particular category. After getting that, while recommending a product, we are using location. So it's two different problem actually, yeah. Hi, how are you measuring the effectiveness of the scores that you are calculating? Effectiveness, okay. So first thing, as I said, the scores that we have calculated, they are based on the latest technologies or latest features in that particular category, right? So if a particular phone is containing everything latest, so for example, let's say in case of one plus one, it has the lowest price, right? It has all the good features that we can call as flagship killer features, right? So after calculating all these features, we actually assigning it a number, right? And all the Zopper score for all the products in that category are relative to the top most, right? So based on that, we are actually calculate the ranking while showing that particular product to the user. So let's say if you are searching a mobile, so whatever mobile you are getting, they are based on a Zopper score, okay? So then we actually check whether a user actually clicking on first or second link or not, right? So based on user's click log, we get to know that even though we are giving more Zopper score to this particular category, sorry, this particular product, user is not clicking on that, right? So that is a feedback loop. So based on that feedback loop, we basically can solve the problem. So yeah, yeah, sure.