 Yeah, hi. I'm Shashank from Razor Bay. We capture billions of events each day and we have a lot of problems to solve using data and AI. So one of the problems I want to discuss here is that we detect and prevent fraud in digital and e-commerce transactions. So as everybody knows where is the money, there is the fraud. So as a payment gateway and as a payment service provider, we have to prevent the losses of our merchants. So we have different products and different problems to solve. So one is that fraud prevention in e-commerce transactions. So how we does that and what we does that. So we captures 200 plus parameters from the e-commerce websites and applications using plugins, HDKs and REST APIs. So once we have that, we move that into Kafka and then we have Apache link. So where we generate 300 to 400 features out of that. So features like user behavior feature, how many transactions you have done in past or past week, past month, these type of features. And location feature, location is five star location or four star location. Same with the IP and the phone and user journey, how many events you have clicked on that. And end of the transactions within 200, 300 milliseconds we predict whether this is like fraudulent transaction or genuine transaction. So main thing is like how you run your AI models on the scale in real time and how you train that in that period only. Because you have to do incremental learning and because whenever fraud lends are fraud lends users are doing to trying to play with the system or try to do some frauds, they will do that within some small time frame, maybe in one minute, five minutes or half an hour. So your model should be up to date to that. So we use incremental learning algorithms there and that feature generation four in a feature generation and prediction of model. We does that in 200 milliseconds. So for that, for that, currently we are deploying our model in fling application itself. And once we predict that model, predict the class, whether it is like fraud lends or genuine with the probability score, we send that to our customer using postpacks and real time dashboards where they can decide whether to process that transaction or not. So this is the use case we augment lot of third party data also on that like on city we does that. What's the, it's tier three city or tier two city or tier one city, what's the population of the city, what's the type to achieve that, what's the pain capability of the user. Is there any anomalies in the transaction or merchant transaction, these things we does this. Okay, questions, questions for him real time fraud detection big topic anomaly detection a big topic. Yeah, we've got one here one now just the flash talk speakers are helping each other out. Go for it. Thanks. That's a great topic that you spoke about. So whatever we do fraud detection or anomaly detection, I think one of the problems in security field is whether you're doing the detection at a transaction level or at a user level. And how are you actually achieving that because some of the projects or resources that I did spoke about when they when you're doing the detection at the user level, they recommend that you use different model for different users. So how did you go about solving the problem, especially at scale, but you have a lot of users can just explain briefly on that. So actually we solved that problem at transaction level not user level user level we take it as a you can say some features. In our model. So as you as you are saying, yeah, this is the problem because in transaction time you you only get some milliseconds to predict that. And you have to take in account all the user history what you have done in the past and what other users are doing in similar like like if you have to do clustering at these things. So in our case, so so fun example in e-commerce frauds what happens so e-commerce frauds is not generally some geography specific or user specific it change. So if you are in North reason, then there are different type of fraud. If you are in South reason, there are different type of fraud. If you are selling some fashion product, they're different type of fraud. If you are selling some electronics products, they're different type of fraud. So we create some micro models, some address models, some user models, some geography models, and by enabling that then we generate the results. Okay, great. One last question from Hussain. Yeah. I think the follow up of this question, like I just wanted to know like how complex is the model and what are the scale problems? Yeah, so currently we are using two type of models. One is classification model as we have the label data. Because we know that what happened to that transaction in past or what happened to that order in past. So if we have that like one year or two years of that transactions and with the label, so we use classification models on that. And we have an LSTM models also. So in LSTM we generally draw something like it, when user opens the website, when he has done first event, when he added something into the cart. So these type of journey, we train LSTM models with that, with features and with time spent on each event. So then after four or five events, we can predict whether this will be like fraudulent transaction or genuine transaction by using user behavior. The user behavior model and classification models. Okay. Well, usual applause for Shashank from Razor Pay.