Catalin Popa: Product Recognition in Retail/E-Commerce with Azure AI

This talk discusses the application of artificial intelligence in retail and e-commerce, focusing on a live demonstration of a product identification application built using Microsoft Azure AI services. The speaker, Catalin Popa, highlights the simplicity of integrating AI into applications, even for those who are new to coding. He showcases a demo app that allows users to upload images of products, which are then identified and linked to an e-commerce site for purchasing. Throughout the presentation, he emphasizes various AI use cases within the industry, such as visual product recognition and inventory management, while also mentioning the importance of image quality and training data for effective AI model performance. The session concludes with insights on how AI can be used to enhance user experience in retail.

Thank you for that and thank you for joining this session in which for the next 40 45 minutes I guess we're going to talk a bit about artificial intelligence in Asia and yeah we're going to see a bit how artificial intelligence can be applied I guess into retail and e-commerce. So I'm going to tell you from the start this is going to be a live demo. What

you're going to see in here is literally things that we were we are going to build together, right? Oh, well, I'm going to build it. You're going to watch it. Sorry for that, right? But it's going to be a live demo. We're going to try to build a simple solution that kind of demonstrates how easy it is to use Asure AI, right? The components from the Microsoft

cloud in terms of artificial intelligence and how those can be applied into your application. It's going to be a really simple application. And I'm pretty sure that everybody who's watching here live or online are better developers than me. So if you see at some point things that don't make sense from a code perspective, keep in mind I'm a noob from a developer perspective. But enough about that.

Let's go straight into it. U just a few things about me because you're probably wondering why I'm here and why I'm talking to you guys. Uh, my name is Catalin Popa and all of those badgets and all of those titles that you see in there are basically in there because I've been working with Microsoft Technologies for the past 12 years. That's my whole career. I'm going to

be honest with you, right? Uh, since from my very first job up until now, I've been working with Asia and I've been implementing Asia solutions. So, I'm a consultant. uh at this point I'm implementing solutions usually for major enterprises and for governments and some of them in the past years uh because talking about artificial intelligence here in here some of them include a AI components as well

more and more companies and more and more governments as well try to implement these kind of solutions as things go along right the technology evolve and with that evolution as well that's how how things are now right it involves uh artificial intelligence pencils. So, just to show you a quick idea of what we're going to build and how it's going to look like. The simple solution that

we're going to that we're going to present today, it basically looks like this. We're going to have a very simple application which will be that product identifier application that you see in there. In there some user is going to go it's going to upload an application and then once that application is being uploaded then Asure does kind of its magic in which it will try to identify

what product that is and if that product is identified then it will basically make a redirect towards an e-commerce website. I have a small e-commerce website. It's one of our my customers actually right locally in Romania. That's where I'm from. And yeah, it's just going to redirect to that product. What you're seeing in here to be honest is basically kind of a let's say enterprise version so

to speak of this Google lens cap capability that you have in google.com. Probably most of you already used it. But if you guys go to Google and if you upload a file in here, right? Then basically Google is going to tell you, hey, I found this little toy that I just search that you just searched. I I found it on Amazon, on Printmix, on wuria.com and whatever.

Right now, imagine it's not as general as that, right? Imagine it's just for your business. It's just for your application. And the demo that we're going to build, I'm going to show you the end product so that you guys understand how we're going to get there. Kind of looks like this. It's a demo app that I built in here. If you want to search directly for your

product, you can literally go upload your file. You want to search for the product directly. You're going to get a message at some point. It found the Mutella toy. That's how it's called, right? And then afterwards, you will be redirected to the product page at some point. Here it is. It co costs around €9. That's kind of the conversion in here. And you can actually buy it

from the customer website, right? So this is what we're going to build. Very simple. I'm going to explain you why I want to show you these why I want to show you guys this example because this is just a mere example of product recognition when it comes to visual product recognition and visual searching. But in terms of artificial intelligence, all of these components have certain use cases

that can be applied in different in different scenarios. This is just one of them, probably the simplest one that I can show you and it has an visual impact. But we're going to talk about other use cases as well of artificial intelligence specifically in this industry in e-commerce and retail in general. Now, enough with the PowerPoints and with the slides. As you can see, I'm not a

PowerPoint fan. Never was. Sorry for that. So, let's get straight into it. And I'm going to open my Asia portal. Some of you might be familiar with Asia. Some of you might not be. Bear with me. I'm not going to explain all of the things. I'm going to just go really through really quickly through all of it because it's not about Asia. What I'm trying to demonstrate

here is literally how easy it is to use these technologies regardless if you're an AI expert or an Asia expert. It's literally just kind of plugand play so to speak on certain features that you can actually that you can actually use from here. So without further ado, let's actually go and create the only resource that I'm I will need in this demo. And I'm going to use

one of my subscriptions in here. I already have a resource group which I called product recognition demo. And in here I am going to create an AI resource. Now I'm going to filter all of the Asia resources that you have from an AI perspective. And you can see here a bunch of them to be honest that are presented by Microsoft. The one I'm going that I'm going

to use is the first one. This Asia AI services, which is which is the general component when it comes to building AI components in in in Asia, in Microsoft Asia. Now, you're probably wondering, okay, but why that and not some other components that we have in there, right? The reason is because if you want to use AI as a general matter, then probably that's the the thing

that you want to that you want to uh that you want to use in your solutions. If you have something really specific that you want to do, for example, if you want at some point to I don't know do translations, then you probably will use a language service. If you want to do speech to text or text to speech, you're probably going to use speech, right? If

you want to do cog cognitive vision, right, as we do computer vision, sorry, or custom vision as we do today, you're probably going to use those components, right? But this one literally it's incorporating all of them. So if you don't know what to use or if you just want one one resource to manage, just keep in mind this is the general one that you want to use.

So let's build one of these. Shouldn't take too long. I'll select the closest region, I guess, or the most common one, I guess, for Europeans. Let's put West Europe. And let's call it, I don't know, pets demo. Something like that. I guess I'm going to use the pricing tire, the only one that I have available at the moment, which is standard. I will acknowledge that I read

all of those terms that I haven't actually read, right? And let's go quickly through it. From a networking perspective, I'm just going to leave it open for the demo's sake. Identity as well. Tags, these are minor details when it comes to building resources in Asia. So, I wouldn't really focus on them for a while. So, let's create it. Right. One of what I want to show you

in here is that with just a few simple clicks or a few simple commands or with infrastructure as code if you're more wellversed in Asia let's put like that you can easily create a such resource and now when it comes to that resource let me refresh it until ah here it was sorry for that when it comes to that resource now you're not going to see too

much too things in here that you can actually configure from an Asia perspective. You're just going to see this resource, some information about it, but the most important part. You're going to see some keys and some endpoints. And yeah, for you guys who are developers, you probably already guessed it. The reason why you see these keys and these these endpoints in here is because you can do

APIs towards these this service to have some predefined models that you can already use or some semi-predefined models that you can already use so that you can literally embed them into your application right as a developer. So we're going to try to do just that, right? If you remember our diagram which we showed earlier, it will literally be just like that. We're going to build our semi

model if you want to call it like that right uh in which we will train some things to identify some products. Then based on that we're then going to do an API request toward that that model so that we can get results on new images that we're going to send to be recognized to be identified. Now the simplest way on how you can do that to train

your own model from that perspective is by using one of the other portals that Microsoft has which is customvision.ai AI and in here if you log in with the same account as mine which is here my live account you can see in here I'm logged in in the same manner right here you can build different projects right now of course going it's going to ask you for

the project name so let's call it pets demo as well it's going to ask you for a description let's put it object detection certain products on a e-commerce website doesn't really matter right now. It's going to identify already the resources that you have in Asia. So what you see at the moment those are two resources that I already have built. The one that I just built with

you guys a few minutes ago and one which I had in here for a couple of days because oh well I have to test my demo until I present it to you. So that's the resource that I just that I used a couple of days ago. But let's use this new one that I just created. Right now afterwards you'll see that you can choose between two project

types classification and object detection. And while some of them do different things, at the end of the day, right, we will want to use object detection because object detection will basically lets us find content or specify something inside an image. You can have multiple objects inside an image, for that you'll use object detection. If you just want to tag an image in a certain way and put

a label on that whole image, then probably what you want to do is classification. So for our sake, what we're going to do today is object detection because maybe at some point in our pictures that we're going to send, we're going to have multiple products. That might be a case as well, right? So you you would want to detect all of those products in that picture. Now,

in terms of domains, when to selecting one of these in here, these are kind of predefined models that you can that you can use, which Microsoft puts at your disposal. It's kind of specific for what you're trying to do and you can find documentation and information about how those uh are specific for your case in the official Microsoft pages, right? But you can see that if you

don't really know what you want to detect, probably you're going to want to use a general domain. If you want to detect specific things like logos or products on shelves or so on, right? Probably you're going to want to use one of those domains as well. Now I'm telling you all of these details because as I mentioned earlier right at the start of the session it's important

to understand the outcome that you're trying to do right because in terms of AI you've probably already seen it for from from some of the other sessions that were uh into this conference before and even today these are building blocks it's not literally something off the shelf that you will use and that's it usually these kind of hyperscalers right when it comes to Microsoft Amazon Amazon, Google,

you name it. It doesn't really matter, right? They put at your disposal these kind of resources, AI resources with some predefined models, but they're not going to give you a finite product. It might be in some other cases. But when it comes to building a cloud solutions, these are building blocks, Legos if you want, in which you have to take that component, figure out in your architecture

how it might help you and then put that resource to use so that you can actually build or do whatever you're trying to do. Right? So let's do the same thing as well. Let's actually create a project. Takes a while. Here it is. And in here it will be very intuitive. Right now as a disclaimer everything that I'm doing here with clicks you can also do it

programmatically. Right? So if at a certain point you want to script all of these things that you see in here or if you want to I know just have another application doing it. You can do that via PowerShell commands. You can do that via infrastructures code. You can do that via API requests into your application. What I'm doing here is visual because oh well, let's try to

explain it as clear as possible if if we're here to explain all of these things. Right now, I'm going to add here a lot of images. So, I'm going to show you I have a specific folder in here in which I have a lot of pictures with a lot of things that usually an e-commerce website for pets would usually sell, right? They're m mo mostly four products.

this wonderful toy that you see in here, right? Some treats for dogs, I guess. Then you have other treats for I guess it's cats or even dogs again. I honestly don't know, right? And some cream, some balm that you use for your pets as well, right? So for product, let's try to see if we can train our artificial intelligence to recognize these once uh once we want

to interrogate it. So, we'll do just that, right? I have uh what's the number in here? 17 total without these folders. I have 68 pictures which I will use for my training. And I'm going to tell you in just a minute why we use so many pictures, but you probably guessed it. So, let's open all of these. Let's see how fast the internet connection is is in

here for 68 file upload. Yeah, pretty quick. Okay. And now once they're uploaded, you actually have to do let's say the work, right? You have to go through all of these pictures and at some point you have to select the area that represents a product or an object that you want to identify, right? I'm not going to do that for the moment, but you want to select

that and then you'll have to put a tag on it or or by default the technology actually kind of detects objects in there. So if you see if I just hover with my mouse on the picture, it's already kind of predefining me an area in which it basically says, hey, I'm kind of identifying something in here. Do you want to select it and add a root mode

tag for it? And I'm going to do just that. Let's call it Smith and Bert and Balm or whatever. Right now, once I put a tag on each image, you're going to see that I'm going to find it in the tagged section, of course, right? And I'm only have at one at the moment. So presumably if I want to train my artificial intelligence, my model to be

as accurate as possible, I will have to do these kind of tags, these kind of labels for all of my pictures or at least for most of them. Let's go like that. Right? I'm not going to do that. It's probably going to take us 30 minutes to do it. But bear with me until I tag at least two of each. Let's see. Right? I'm going to call

this Nusper snack. That's the name, I guess. And the last one. Which was it? Which was it? This one. This Volute treats. Right. See, uh, this one, it's kind of fishy. It doesn't show me the entire product, for example. So, I'll have to do this manually myself. Sorry for that. I'm going to do it like this. Here is, right? So, let's call it wolf. put. All right.

So, once I did that, I should presumably have four tags in here with one image in each uh tag. Let's do one more of each, right? Bear with me a second. You can see the pictures, right, from a let's say uh from a diversity perspective uh are of all sorts. Some of the pictures are incomplete. Some of the pictures might be fuzzy. Some of the pictures might

be in different angles. The more variation you have, the better it will be because that will basically help your model to understand all different kinds of scenarios when a user is taking a picture. I know in a let's say utopic world, we would want users that will only give us exquisite pictures, but realistically that will probably not be the case. So the more you attend for these

kind of scenarios, the more the more probability it will be that you will also recognize odd pictures that users might send at a certain point. Right? So let's do one more. Bear with me a second. I know this is a bit boring from a demo perspective, but I'll promise you it has a reason. Burton. Oh, wait. Let's do a Vol split also. Right. And then one more

which was the other one, the Windows SP snack. Right. Sorry, not this. This one. Great. So now I have two of each. I still have a lot of them untagged. If I try to train my model right here, I usually have two options. It's not going to work. I'm going to tell in just a why it's not going to work for the moment, right? Uh but you

usually have quick training or advanced training. Now, quick training, it's basically a thing that Microsoft puts at your disposal and says, "Hey, I'm going to use the minimal resources that I need from a computing perspective and time perspective so that I can train a model that would be at least decent." Right? If you want to do more advanced training and if you really want to spend time

and have a budget for training, you can actually specify in here as many hours as you want for that model to be trained. which basically means that more computational power would be put in place would be put at the disposal of the model of the service to actually train this this solution. I'm going to try quick training. It's going to fail, right? The reason it's going to

fail is because even though this kind of training model, it's a click- toclick kind of situation in which it's really intuitive, you can't train on minimal data, right? As any train that's any artificial intelligence, right? At the end of AI, as any model, you have to have a minimal set of data to actually do that. So, at this point, you probably already seen it. If I want

to train a project in terms of object detection, I will need to make sure that I have at least 15 images for every tag. So, that's why I have 68 images in total. It's because I have to kind of respect this threshold of having 15 per product. Well, the good thing is that we don't have to do that manually ourselves today, right? I already have a project

which has all of those tags. You can see here that I have 30 33 pictures of the Mutella toy, 23 of Smith and Burden, 24 of the Nusper snack thingy, right? The Volut as well, 23. The reason why I have so many is because in some pictures it might be that I have a picture with both products, right? So some of them might coincide. You can put

multiple labels. You can put multiple tags on a picture. It doesn't have to be a specific product. It can contain uh multiple products at once. Right now once you train it, once you train the model, it probably it's going to take about 5 10 minutes if you did a quick the quick uh training. It also depends on your data. How many pictures do you have? and so

on so forth. But once you have that, you're going to receive some info about your training, right? The precision, recall, m map, right? Uh components that you usually have for a model in which as you can see for my model, if a tag will be predicted by my model, I am 91.7% likely that that prediction will be right. Then in terms of recall, right, if it recognizes

attack specifically, the percentage will be 100%. Right? And in here, of course, the overall performance of object detection if it happens, it will be 100%. Now, the reason why this happens and I'm having such good numbers is because, oh well, my example is really simple, right? I only have four products, four tags. Most of the pictures are kind of close up and clear as you probably saw.

But keep in mind that once you train this right, you probably want to have better and better results with each iteration. Now you can improve these results either by literally feeding more data into the model of course or or putting more compute time into your uh into your training. It's or both of course at at some point. Now once you have these right because these look so

nice and well you're probably wondering okay I have a model I trained my model how do I use it now you have this predictions section in here in which you either can see your endpoints from an API perspective right and you have usually two options in which you can send an image towards this model you either can send a URL if your image is somewhere on the

internet or you're saving it somewhere into some storage account or whatever Or if you want to literally send the file directly towards the API, you can use this endpoint in here in which of course you have to have specific headers for authentication and you have to have specific content specific body so that it will actually understand what you're trying to do right there are some security measures

as well into that API request but at the end of the day it's an API request in which you send an image to be predicted right to be detected. Now, I'm not going to do that via endpoint just now. I'm going to do a quick test in here. And I have these pictures for testing in which I'm going to upload this other picture of Motelato toy because

I want to show you guys what the result usually is, right? So, you can see that it kind of detected two Mutella toys in here. some of them with uh some of some of them with a specific pro probability right the reason why it does that is because if you guys remember our uh our percentages right I have a 91% chance that it will probably be predicted

correctly or I will predict something in an image now now that doesn't mean that I will I will have my model which will be spoton when it takes that the more images you put in there the more data data you put the more specific it will be but in this case although I two products that have been detected, right? You can see that one of them is

actually predicted with a 99.9% probability. And that's one of the indicators that probably already know, right? You should take in consideration when you're analyzing the information that you're getting back from your models. You probably will look for something that it's over 70% or 80% or 90% because at some point some things might be fuzzy. A picture like this, right? Might generate you two responses on the same

toy and the same product most probably, right? Or it might be at a certain point that you have multiple products and you want to identify the one from the middle of the of the image. Different situations, right? But keep in mind that always always your response will be like this. You will get a in an API version you will get a JSON in which in that JSON

it will specify the name of the tag that has been identified and a probability right in which the model basically says to you I'm looking at this I'm identifying this in this region and I'm 99.9% sure that's a mutella toy based on the training that I just had right so yeah that's kind of how it looks like from that perspective Let's also jump in a bit into

this part right here. If you guys remember, we had this small application and this is where the dev part starts. I guess I won't bore you too much with code, but yeah, it's kind of a very simple application as I said, right? You can see that I have some headers in here which is literally which are literally the headers that you see in the application as well,

right? find out if an item is available basically this row right here and then I have of course a part of my application which is letting me upload a file an image then based on that image I'm actually doing a JavaScript in which I'm doing some processing of that image now that processing is using a task that I have in here in home controller right this where

is it sorry this save file tag task that I have in here. And in this save file task, you'll probably see that API request that I was bugging with bugging you guys about uh earlier, right? You will see that I'm adding the prediction key, right? Don't do this. This is unsafe, of course, but as I said, I'm a noob when it comes to uh developing. I'm pretty

sure from a security perspective, you guys will be way more attentive than me. Uh I'm pointing towards the UR URI as well, right? So this URI that you see in here, this West Europe API cognitive.microsoft.com, it's basically Sorry for that. Here it is. It's basically that endpoint that we were talking about in here. So if I go to predictions and I want to show you guys the

end point. So let me delete this. All right. It has the same It has the same uh structure as well, right? And you can see that I'm building here an API request as I'm being instructed into this into this small let's say disclaimer right here, right? I'm adding my content, right? The file object and the file name, right? Then I'm setting up the content type and then

I'm waiting for a response which will be the of course the response of the API request. Now the response of the API request then I'm literally going to parse it into a J token and then I'm going to look for that prediction score. If you guys remember that percentage that we were talking about this is the prediction score right in which I'm looking for a prediction score

which is over 90%. Right? If it's over 90% in my specific case in this demo right then I want to return that prediction from the uh from the model. If it's under 90%, I'm just going to say, "Yeah, I don't know if I have that product on stock, so try again with another picture or I don't know, I just don't have it." And that's it. Right? So

that's kind of how it looks like, right? I also have, of course, some fail safe in case this get goes wrong. I have some other error messages in here and so on so forth. But function is executed right here. Sorry for that. A bit further up. Right. In which if it's a success then yeah I will alert my application that I found the item with that data

and it will be redirected towards the product page. If that doesn't happen then the item it's not available. I just I'm just showing an error message right in that point in which I'm saying oh well I don't really find it now in terms of mapping right and because I do that redirect as well you've probably seen it but yeah this is kind of simple I guess uh

if the product will be a Mutella toy or that snack product I'm literally just pointing to this website right here pipedia.com in which literally this e-commerce website is selling these kind of products What I'm showing you here are real products. I'm not joking. This is actually a real website which sells e-commerce uh things for for pets in Romania. Right now, that's kind of look that's how it

kind of looks like. Let's do a quick demo again to show different use cases, I guess. Right. I'm going to use a different picture for testing. Let's use one that shows me that ball, right? If I'm searching for the product, it's going to find me the balm. It's going to redirect me, right? Takes a while, but here it is. Ah, doesn't have the product anymore. So, sorry

for that. I didn't did I didn't do that research anymore. Probably that product is not available and and anymore on the website. Let's do split then. Let's see if this fuzzy gets recognized. It does. And here it is. Right. So different cases. Let's do also one which is not containing a product. Let's see how how that goes. Right. So I have here a picture which I took

in my hotel I guess a while ago which has I don't know some food sort of. Let's try to search for this. Probably won't find anything. There it is. It's not available. Just try again. Right. If I do a wool, a simple empty wall probably it will be the same thing, right? So a very simple a very simple example of open AI right of AI services sorry

that uh that can be used and can be implemented in your application. Now, I'm showing you because when it comes to sorry, when it comes to use cases in e-commerce, that's just one of them, right? Uh you're probably wondering, okay, how that visual search capability makes a difference from a product perspective. Well, for some companies, for some businesses, it does. I've seen companies which were specifically looking

for these kind of implementations because they had users that were complaining that the search bar wasn't doing its job. Or you might have users, I've been one of them as well in which I'm seeing a product in a country, I just take a picture of it and then when I want to search it in my own country, I just want to input that picture, see if that

uh uh if that website actually has it and that's it. Right? some other things when it comes to use cases uh in e-commerce and in retail because I I was mentioning to you guys those are building blocks right this is just one of the scenarios that you can do with AI when it comes to when it comes to Asia components right but keep in mind that you

can do all sorts of things all of these things that you see on the screen right now these are real life uh scenarios that I've seen implemented in customers that I work with or they had the intention to implement at a certain point price tag detection right for example if you're in a retail uh shop in the retail store at some point and you want to make

sure that a price tag actually actually exists on your products or of course if you want to I know scan certain things and you don't want to do the infrared solution in which it uses the code bar right you can actually use it for that as well right automated checkout and payment as well uh I know you guys here in vill I've seen it with my own

eyes you have those automated stores in which you just go in front of the uh cashier and basically everything is scanned. It just tells you you have €25 to pay and so on so forth. You can do the same thing with AI as well, right? Instead of the infrared solution, you can put all of your products probably I know in a specific area then some camera can

actually take a picture or a couple of pictures live and it can basically tell you I've identified this. This is how much you have to pay. Right? That has been implemented as well. Uh you can also do inventory management. I've seen a lot of things and this is actually a pretty common thing in in e-commerce and in retail actually sorry in which if you want to find

out how many products you have on a shelf of a specific kind I know of a specific uh uh a specific product you can literally take a bunch of pictures and then artificial intelligence for you will say okay you have 44 products uh based on this shelf that you can you can put it in inventory certain things that you can do right as I said uh there's

There's no limit in here. And probably you've heard this from from every person that has ever talked about artificial intelligence in here. It's all about having a building block, right? It's up to you as a business. It's up to you as a customer, as a user of these components to figure out how you want to place them and how you want to use them. So that's about

it. Uh we still have 10 minutes, I think. Do we have questions here? Sorry. Okay, great. What is a good number of images for each tag to avoid overfitting in the model? Well, it depends out of my experience probably yeah somewhere between 40 and 50 I would say if you want to do this specific thing that I'm doing right here right because yeah at a certain point

just imagine this imagine you have a product right uh there is a finite a limited number of I don't know uh angles or limited number of I don't know situations in which you can picture that and you can take a picture of that product, right? You can have 5Z pictures, you can have pictures from a specific angle or whatever, from top, from bottom, you name it. But

at some point, they're kind of being redundant, right? You'll probably just repeat the same scenarios. Out of my experience, probably 30, 50, something like that, right? But I've seen really good working models with just 20 pictures. It really depends. Uh it's a trial and error. I'm that's how models work. Just try it out. train it, see the results of that model. If the numbers, those percentages that

you've probably seen, right, in terms of prediction and so on so forth are agreeable to you, then that's kind of it. Second one, what augmentation techniques are you constantly using while working with image recognition? Well, this kind of depends, right? Uh it's up to you to use your own technologies in here in terms of techniques. I'm not using any of it at the moment. So what you

saw here it's literally brute nothing at all right but if you want to use augmentation techniques to do image recognition for that yeah I guess you can use it as well it doesn't doesn't really matter so you can go from that spectra you can go from not using anything to using all kinds of techniques doesn't really I mean of course it makes a difference don't get me

wrong what I'm trying to say is that you don't necessarily have to use them to actually do what I've been showing in here what minimum percentage of prediction accuracy is used for a production environment. Well, I guess it depends on your customers, right? Uh I'm joking here, but yeah, I'm I would recommend if it were after me at least 70%, and I'm telling you this because you've

probably seen in my uh in my app in here that I had a 90% limit, right? Well, I'm doing that because it's a simple it's a very simple scenario, right? I had four products in which my pictures were kind of up close and clear and that was it. I've seen cases and I'll give you an industry case at some point, right? In which there was a manufacturing

company in which they wanted at some point into their production line to differentiate the things that they were manufacturing, building, whatever, right? So they had uh a bunch of parts, mechanical parts that were building for cars and at some point they want to identify if one of those mechanical parts was intrusive in the line of production that was in there. Right now as you can imagine the

scenario that I'm trying to explain to you guys, even if you do pictures with high-end cameras and if you even if you monitor all of those production lines very closely, the quality in there won't be great. Right? So the lower the quality that you have on pictures probably you can ease up a bit on the prediction uh percentage right on the prediction accuracy. I'm saying that because

at a certain point 70% might be enough 65% might be enough but it really depends on what you're monitoring what you're analyzing from an image perspective. If you have things that are really specific with just one product and they're closeup like I had go up with the with the with a percentage. If the pictures become more and more let's say uh complicated in which you have multiple

things inside there and you have multiple things that can be detected probably you can go a bit lower but not too lower because otherwise you might have false positives well. Which LLM model is it using? So those models that you see in there those are black boxes from Microsoft. You don't really have details about the models. I mean you can probably of course uh read a bit

on the Microsoft documentation for it but these are readytouse models that Microsoft puts at your disposal. They're a bit customizable. I've probably seen it, right? Because I can go, I can put my own pictures and so on so forth. But what's inside, what technology is inside there, how the model is constructed and so on so forth. Yeah, Microsoft doesn't really give you those details. However, however, don't

get me wrong, it's not like you're being stranded to not use your own models. If you want to build your own models in Microsoft, in Asia, you have Asia Machine Learning and other components that you can build your own models yourself. These are readytouse models because Microsoft and whatever I'm talking here about Microsoft but again it can be Amazon it can be Google it can be any

cloudscaler doesn't really matter these are kind of offering customers right uh readyto use components for people which are not that wellversed from an AI perspective so that they can implement in their solutions this is just a visual one but imagine if you want to I know extract data from images a lot of companies nowadays banks or I know insurance companies and so on so forth do a

lot of document processing in which they extract data from passports or from uh I don't know identity cards or you name it right so you don't really know what's behind it you just know that you want to use it right at some point what happens in the back end I know it's something that some of you geeks right here and me also don't want to don't want

to admit but yeah it's a blackbox model it's in there Microsoft gives you very limited about image recognition has implicit text recognition to use it for proposing similar pro products. Yes, you have text recognition as well. Now when it comes to text recognition depends where you're going with with this. If you're talking about text recognition in terms of I want to identify if it's handwriting or yeah

computerbased writing so to speak that simple you have that model as well. It was one of those actual models. You remember when we went to Asia and I showed you guys all of those different components that you can build in Asia. Well, some of them are specific for a specific things, right? You have one which is called text analytics at some point in which you can input

a text into that model and you can get all sorts of stuff. You can get the sentiment of that text. You can know if it's positive or negative, right? if can analyze if you have at some point I know the what what are the main subjects that are being talked about when into that text and so on so forth. So there are different things that you can

do of course with uh with AI this included. That's it. That's all. Okay, then I'm right on time, I guess. Thank you for that. Thank you for watching. Thank you for joining. And if you have any questions, just yeah, hit me up on LinkedIn, I guess. I will be more than happy to respond to your questions if there are any. If you want to talk about artificial

intelligence, I I will put your I will put my knowledge at your disposal if I can answer to your questions, of course. So, thank you again for joining and enjoy.