Name: [Simu-live] The AI Edge Revolution: Build, Deploy, Dominate
Uploaded: 2026-01-20T06:04:04.708Z
Duration: 44 min 44 s
Description: [Simu-live] The AI Edge Revolution: Build, Deploy, Dominate

Transcript for "[Simu-live] The AI Edge Revolution: Build, Deploy, Dominate": Hello, all, and welcome to today's webinar. My name is Vinay, and I will be helping support today's live session. Before we begin the session, a little bit of housekeeping. On the top right corner of your screen, you will see the live chat option that you can use to interact with your fellow attendees. We have also attached a few documents in the docs section next to the chat. On the extreme right, you will find the q and a box. We encourage you to leave all your questions in that q and a box as we progress through this webinar. We'll make sure to address all your questions towards the end of this presentation. A quick reminder, this session is being recorded, and we'll be sharing the recording with you within forty eight hours. Now please join me in welcoming today's illustrious speaker, Christopher Say, principal solutions architect at CloudFare. Hello, Christopher. Hi, Vinay. Thank you so much for your introduction. Alright. Hi, everybody. Welcome back to our power block session on the Cloudflare Enterprise AI. Today, we are on session number five. Okay. The title is called the AI edge revolution. I'm so excited for that. Why? Let's take a look at the agenda. Okay. So far, we have covered session one to four in the previous four session from I mean, the very fundamental building block of the whole Cloudflare developer platform solution. Alright? So those are really laying the fundamental basic. Today is our first session, session number five, that we already talked about the use cases of our AI. So we have touched on, right, some of the the thing that we are be like, how you use the AI to do your chatting, image processing, voice processing, so and so forth. So let's move on. Okay. Before I touch on today's topic, let's do a quick recap for the last session number four. Okay? In session number four, alright, of course, we have covered things like the durable object, the device data storage, the rest of the session, the fundamental of the workers, so and so forth, that kind of thing. And today, we have already go into the use cases of the AI, as I mentioned. Alright? So on the last session, we touched on the deal durable object. If you remember, okay, the number one thing is that it is something like the full SQLite kind of operation. It can run wherever your data is. Okay? Your object will stick with your application. Okay? So that means, technically, you can achieve near zero latency in any of your so called runtime per se, and your story lives in the same track as the application. So the main portion of that is this is part of the workers' serverless platform. And we touched on one of the so called real time chat room use cases or demonstration. Like, for example, in this demo, if you remember. So we have people coming from I mean, some data from India, Bengaloo sorry, Bengaluru. Okay? The other one from Mumbai and some from Sydney, or you can be from any part of world. Okay? And you can seamlessly create that. User a, user b were put in the same room to chat privately and securely, whereas the rest of the user, although in the same company or maybe they will be a different customer. They can talk via different chat room securely as well. So that's the beauty of the whole thing. And if you remember, we also touched on some other demonstration like the n NBA polling, and we have a very fun game they've created. Okay? It's called the where you can see multiple player as much as the four player coming together and fight against the the common enemy. K? So by demonstrating all these thing, you are able to explore your creativity. Alright? Making use of the technology to become a business generator for whatever, like, application. It can be so called the file sharing collaboration application you want or gaming or video streaming whatsoever. Whatever you think of, Cloudflare developer is there for you. Alright. This week demonstration. So just if you recall if you can remember, I would say that, okay. This power block session is more on demonstration rather than going through the the the deck itself, But we already touched on certain fundamental concept before we go to demo. So we will see that how we create the real time chatbot and also doing a real time text to image processing, which is super, super important. Today, talk about LM, talk about ChatGPT, Gemini Home, and so and so forth. You can create something like that for your own application, be it personal or for corporate corporate usage. Right? You also explore multimodal engine shortly. Okay. So as a full stack application, this is our reference architecture. Right? You see that number one thing, of course, is security is the most important thing. Once you push your application to the public, you are yeah. Yeah. Yeah. You're inviting a lot of malicious user to come in to do whatever they they want they want. So you need to protect yourself against things like DDoS, LASAP server attack, bot management, and, of course, the API issue to protect your API consumption or injection. Right? Then moving down the track, you have your performance, very important. Your performance is bad. Your whole application cannot sell, or it cannot really function and serve the purpose. And the compute is the one that run your runtime, run your core, your application, not to mention about the storage. We have cover extensively for the past two session. How you keep the state? How you keep your database, right, without spending thousands of thousands of dollar, right, using in end of the CSP? So because we charge zero egress cost, if you remember, we kill a stretch of that, that that is really just cost saving. So without touching the rest, of course, we have our Workers AI, which you can touch on today. Vethorization, there will be subsequent session. AI agent, where you use our SDK to build your own agentic AI. Alright? That is the fun part of whole thing. K. So the rest will be, like, your also, how you manage your loss, how you manage and do your analytic. Like, for example, who's consuming your API, who is pumping what kind of a so called prompt, anybody trying to do some prompt injection or or or data poisoning, so and so forth. Right? And okay. This is one of the use cases that are pretty relevant to today's demonstration. It's not the same, but the concept is there. Okay? This is a much more complicated one that you do a serverless image content management. So you have your desktop from user, laptop. It can be mobile. It can be user generated content. It can be images, okay, where you can resize they pump in the images to the whole Cloudflare solution, Cloudflare platform. First, of course, you'll be so called inspected by application security to check whether is there any layer seven attack that this user tried to pump in. It can be embedded the image, or it can be ready some form of the call attack, like cross site scripting, SQL injection, or whatsoever. Alright? So after going through this, you go through a series, okay, of engine, things like content intake, image classification, very important. Without knowing what kind of image you've been pumping, the the whole process is not going to work. Okay? Then go to moderation. If anything that's wrong, you'll be just throw away as garbage, and also move on to storage and go through your resizing and signature validation, and lastly, coming back to all your different user. So this is one of the content management use cases. The next one will be more like you're optimizing your image delivery. K? Just like you're doing any of your CSP, but it's much more effective and cheaper solution. Cheaper because we don't charge egress cost, efficient because the thing is deployed globally in our near 2,000 over pops worldwide. It's not just within the specific region, but once you deploy, it's the whole earth. Alright? Everybody will reach your content at the closest pop to to to the eyeball. So in this use case, someone the user will do an image request, right, through your URL, coming to hitting the Cloudflare CDN, k, that give the user the the the lowest latency and faster response time. And we'll do some transformation rule, of course. Right? If any cache miss, okay, we are fetching from Cloudflare r two, which is very much equivalent to AWS s three. If you remember that in the previous, I if I'm not wrong, the the on the second session. K? So you can take the chance to go back and view the video recording. Then you go through our transformation, which is we'll do resizing of your image. After transforming, you go back to CTN and fetch it back mean, send it back to your user. K? It's pretty similar to just now the the earlier use cases, but this time, right, it's more for optimizing. Let's say, for example, you're creating apps that you require user to upload their photo, okay, and require it to be certain dimensions, certain sizing. So you can make use of this use this methodology to enforce that user comply to your requirement. Alright. After talking so much, let's do a hands on. So as usual, if you are I mean, doing the hands on together with me, first of all, you must have a Cloudflare account, https-.cloudflare.com, and you must have a operating system that run Node. Js that is 20 dot zero and above. So if you're that, please follow me. Okay. Let me share my screen. Just give me a moment. Okay. I'm sharing my screen. You should be seeing my demo lab pretty soon. I'm logging into yep. You can see that. Okay. I'm running Ubuntu and running on Node. Js twenty twenty and above. Okay. So the very first demo I want to show you is okay. I try to get the link so they can follow accordingly. Okay. I'll put this in the chat box. Okay. I paste this in the chat box. Alright. You should see that. You can download and I mean, you can so called download this GitHub. Okay? So once you download the actually, you're downloading the whole template. You will take a look. Okay. There's a templates. If I go inside the templates folder, you have many, many of the demonstration use cases you can try out at your own time. So the one that we're going to touch on is this LM check app template. K? So I'm going to copy it down. Alright. L o m. Check location. A temporary LM check. Okay. I quite copy the whole thing into LM. Check app. Okay. There you go. There is the one. Okay. So go to there. Check apps. Okay. Of course, the first thing, as usual, we will see what's in the package dot JSON file. So, of course, dependency is the one that's been specified upon the encoding of this coding. Alright. The most important thing the thing to take note is on the scripting. Alright? What kind of command we can run like deploy? Deploy is deployed into production. Dev ops is to run your local notebook or PC, k, while you don't have any Internet access at all. Doesn't matter. Right? So the rest is more on that description. So first thing first, if you remember, do a NPM install. To install all the dependency, all the packages, let's take some time to run. Okay. Once done, very simple. NPM, we run on local first instead of running in production. Okay. So it's ready. So this is a URL. 8787 877. Let's take a look. Alright. We have the first very chatbot. Okay. Later, I'll walk you through the coding, but let's see what does he do here first. Okay. First of maybe I say that let's say maybe I want to say, tell me how far is New York from Las Vegas, for example. Right? Okay. Yeah. It's thinking. Okay. The reason why it's think so long is not because it's stuck, but because at the back end, I got multiple account. It doesn't know which account should I use or may I mean, the account in my Cloudflare dashboard. Right? So I specify I'll be using this account. Come back to the yeah. There you go. Alright. So you say the New York last week is approximately approximately about, let's say, four four point five kilometer apart. Okay? So you can say maybe things like maybe I'm doing some mathematics calculation. Alright? It can be anything just like you're using an HRGPT. Okay? So maybe I say that there are 10 levels. I'm doing some physics calculation. Okay? In a building. Okay? Page level is five meters high. If I were to drop a stone from level 10, how soon will it reach the ground? Alright. I believe for those of you who do physique, you know that this are very common question, but also quite tricky. Take note that the stone is $200. If anybody can quit enough, then the GPT I mean, the the chat box. Alright? You're welcome to. Okay. Very fast. Come on, really. Okay. The the the answer came up. The credit is for stone rich. We need to calculate total distance for inspiration. Three point one nine second. Any one of you got a great answer? Right? Feel free to challenge. Okay. So you can make use of this chatbot just like any of the GPT and I mean, to to customize this to your own function your your requirement. You can use that for your private or your own corporate usage. There's no open to public as well. Okay? All depending on your requirement. Now let's take a look at the coding. Let me stop this one. Go to the source folder. Okay. The coding is dictated by the index dot t s. Alright? Index dot t s. Take a look. Okay. The very first thing is that we need to find out what model are we using, alright, at the back end for the chat to to develop the chatbot. So, yep, there you see. The model ID is NAMA version three. Okay? This portion. And the most important thing is a system prompt. Any of this chatbot, right, the model require you to specify how you want to respond to be in. So in this case, I will say that, okay. I'll tell the the the model say, yeah. Helpful, friendly, assistant. Please provide concise and accurate response. You can say that you can change to, let's say, I need academic style of respond. Okay? Or I need it to be more more fun, more cheering, whatsoever. You can specify here. So it's all depending on your requirement as well. K? The rest of this, like, default would be something like your default constructor if are doing a c programming, k, or Java. Alright? The most important thing will be the function, how you handle the chat request. And you will notice that okay. First of all, message system, await AI run. Okay. This thing is super important. Measurement token one zero two four, what it does is that I want to specify, saying the respond for a model cannot exceed one zero two four token. So if you're asking what is token means, okay, one token is almost equivalent to one WORD, WORD. WORD. Alright? But it's not exactly the same because one token is actually equal to about 2.7 WORD. So one word is actually equal to about 1.3 token per se, okay, if you do some mathematics. Hopefully, I'm not confusing confusing you. Okay. But this is super important because you don't want it, the the bot I mean, the the model return you 10 pages long of answer. Number one thing, it cost you money, of course, and second, it'll be very well the high latency because of chunk of data they fetched. Right? And, normally, user doesn't want that. There's only one quick answer. K? Like, just now what we did on the mathematics or checking on the distant. So this is something that you need to take note of. K? The rest is just super I mean, the the the respond. So this is the back end. So where is the front end? Front end is at this let me go down to the public. Okay. This is where you specify your front end that you see that g GUI. So I won't go through detail because it's very straightforward. Okay. First of you got CSS. Okay? Then the rest will be, like, type your message, the message box accordingly, and then you call the AI template. Alright? So where to call the AI template is specified in this CheckJS. What is important? Because down here, see, the box I mean, when I click the submit, right, I don't know where to go. So let's take a look at JS. So one thing that you need to take note of is if I go down, okay. Maybe I search a bit faster. K? API. Yes. Alright. This URI, also called the route path API check, is very important because you need to tidy with your back end. So okay. Keep keep in mind that on my JavaScript, I say that I'm gonna send my post my request to API check. So if I go back to my back end script, okay, if you check the respond, let me search. That'd be faster. API. Yep. You see? You you you make sure that the post is trying to post I mean, assessing the API chat route. Alright? If anything that doesn't match, you just say no file, and the thing will not work. So that's where the common mistake is when developer try to deliver that. It doesn't work. It it throw error four zero four. It's really because it doesn't match. Okay? So these are the things you need to really take a look. Okay. Now if I bring you to this model, if I let me give the link and copy it down in the chat box. Okay. It's in your chat box. Okay. The link is showing what kind of model can you consume from our platform. Alright? We got 50 over of them. If I let me enlarge a little bit. Depending on what you want to do. Okay? Let's say if I have to do, like, text to speech. Okay? These are the model. K? Things like Aura, one two. Okay? Mellox. If I wanna do, like, let's say, text to image. Okay? You have your Fras two, Phoenix, Fras one, Dreamshaper, Stable Fusion, Preosys from ByteDance. K? So you got many of these thing depending on your usage as well as, let's say, you want to do image to text. Okay? You have lava. You you form gen two. So it all depend what you want. So just now we're using the number 3.3 or if if I do memory on 2.1. Okay? So if, let's say, I want to do now is trying to do the image. Okay? I mean, text to image. I want to say allow you to type in text and generate the image just like, you know, do dot a I. What you can do is that you can select things like text to image. Okay? So you can choose things like flux two, phonics, Lucid origin. I have already pre deployed that. Okay. I still got pre coded that using Lucid origin. Let's take a look. This is under, actually, you know, the AI. So what I can do is that I pre deploy that to save time. Right? Now it's December. Everybody's in the fastest season mode, so I don't wanna take so much of your time. So I try to quickenote the whole thing. Okay. I got this thing called the LM image app. Let's go to there. Okay. Let's go to take a look at the source. The TypeScript. Okay. So let's take a look what model are we using here. Okay. Is okay. SCF, that means it's hosted by CF. I mean I mean, CrowdFleur. Alright? It's hosted by CrowdFleur, and you are assessing or utilizing for another model called Lucid Origin. K. So you just have to change this model ID, but there are some differences. Once you change the model, there are some differences is that you come to here. Let's say, Lucid Origin. If I open another tab compare to, let's say, just now I'm doing test enrich and if I'm not wrong. Okay. Let's say I pick any one of that. K. Number three. Okay? Take a look. So for example, in number three, of course, as a developer, you know you need to look in the documentation. What are the expected parameter as a prompt, the Lora, respond format, you need to follow exactly. So, like, right now, I'm doing a text to image. Okay? So I must make sure that my code, alright, supply all this input accordingly. So once you know get a hang of this thing, k, some of you are already a developer. You know what I'm talking about. So once you just follow this very detailed laid out requirement, You can just consume from any of the model easily. You can have things like, for example, you can create your own application for automatic speech recognition, translation to to different languages. Okay? You can make use of things from the Hugging Face, from the Google, from Gemini, so and so forth. So it's really unlimited to what you can create from here. So let's take a look at this. Let's say if I run this code. So just give you a feel how it works. NPM run dev. Okay. Now it's listening on port eight seven eight seven. Let's fire up there. Yeah. You'll find that I'm running locally. I'm not running production. Okay. Okay. That's the one. For example, let's say, I say that Spider Man rabbit here. I tried that for for for the whole thing. So, okay, just let's take a look. What does it generate? So it's generating the image. Okay. There you go. Okay. Previously, when I generated that, I see a spider web with two years, but now only one year just like any of the Hollow Loop, some Hollow Loop complete of one year, they kinda thing. So you can do anything. Let's say, I don't know, maybe Cloudflare. Wait. Liar. I'm not too sure what will you generate. Let's try. Let's take a look here. Wow. Okay. I don't see a CrowdStrike logo here. Okay. Probably they cannot I mean, the another model cannot interpret. What does it mean? But, nevertheless, allow you all to create application for your end user to creatively consume this text to image function. K? If you got any question, please post that post that in the q and a. We'll answer that towards the end. Alright. So far, we have explored the text to text I mean, also text to image. Let's take a look at the other application, which is very fun. Okay. I try to copy the link. Okay. I paste this at the chat box. Alright. Okay. So you see there what I can do as usual. You just git clone the link. Just follow accordingly to the step. Okay. This one is a little bit tricky because it's something need to change inside. Okay. Let me show you. Once I done that, it's under the whatever iFly workers. Just oh, whatever. So sorry. AI. Okay. So first and first, NPM install. Very important. Don't forget the step. This will install all the dependency. Okay. The code that you downloaded from this link is a little bit how should I say? I will say it's very old. Okay? But there's something you need to do. Okay? First of all, take a look. You need to install a dependency, which is called a base 64. Okay? I have put the command in the chat box. Alright? So you just follow accordingly. Without that, you'll see a lot error. So I I don't let use face error rate. I mean, I just install you install for you first. Else, you can try without installing and run. You you see a whole chunk of error. Okay? But after doing that, you will not. Okay? So that is not the end yet. Okay? So let's take a look at the source. So you find that there's an index dot t s x. Okay? Okay. You will find that later, I'll show you the demonstration on this portion, but just a highlight. You you will use three model, okay, to achieve what this app is trying to do. Number one thing, he's trying to use it byte dance. No. Not this one yet. Okay. Let's take a look at yep. Okay. First of you use this UFORM gen two model. This is the multimodal. That means you can take input as an image or a text format. Okay? After that, you generate a description. Based on let's say, I I I put image. You describe the image. For example, you're wearing a hat. You're wearing a hat. Later, you see in the demo. So this is what this model is trying to do. After you after you have done the model, right, I mean, describe the the input picture in this example. So you use a. Okay? The second model. Okay. Now you find that this is at h f. That means use using hugging face. It is this is a pretty old code. It's obstructed. Alright? So we are gonna change this. So we are changing to use the Cloudflare hosted model. The link will be this one in the chat box. Okay. That's the one if you refer to the chat box. And I'm going to change that because this this model has been oscillated. Right? By the way, it's not hosted by us because it's at hugging face. So you need to remove that. Paste. Okay. Now it's hosted with CraftLay CF. Okay? It's another model that does the similar thing, and there's a third model. So what this model does is that it based on the generate generated description of the image combined with some other hobby that you could input in, you create another description and finally pass to a third model. Right? I know they sound complicated here, but just bear with me. Just want to show you the flow first because without correcting this, the the the apps cannot run. K? You you hit error. So app post. Where's the thing? Yep. This one. Okay. Previously, using byte, but this model has been obsoleted. Okay. So we are using another model, which I'm again going to paste on the chat bot. Okay. There you go. K. Just follow and change that. That's the fun of the whole thing, right, as a developer or when you're experimenting with all the model. You got to play around, change around, and see how does it behave. Okay. Now you have touched on three of these. Okay? So it's time to save, okay, and do a build first. Okay. To build successful. Okay. If just now you didn't install that depend dependency on the basis for, you see a whole chunk of error here. Now I don't see that because it corrected. Okay. Let's do a run. Run dev because I want to run locally. Sorry. Okay. Is this now a port eight seven eight eight? Okay. Let's try to go to here. Local host eight seventy eight. Okay. What it does is that it try to take your image with some description and change you into a logo Lego kind of so called picture or kitten fly, sesame street fly. Alright? Or change it to a superhero. Let's say I try I try the LEGO LEGO e LEGO iFry. Okay. If you pronounce that correctly. Okay. Maybe my hobby is called, let's say, basketball. Okay? Then you see that take a photo of yourself or you can upload for me. Probably, I'll just take your photo. So this is my photo. Okay. I'm looking at my camera now because just I was looking away into my screen. Let's take your photo. Okay. So I will Barash, you catch your photo. Okay. I got a photo. Okay. Then you move on. You find that the first one, they ask us to input the hobby. Second one, you take a photo. But after getting a photo, remember just I say they're using this multimodal model to describe what you have seen in that model. So you find that, hey. How can I hang here again? No. It's not your hang because I need to select the account. I got many, many more account. If you only got single account, then it should be safe. Alright? For me, I need to select. Okay? Once that let's go back to here. Give me some time. You need to so called analyze the photo. Yep. There you go. You see? You see that this is man with glasses and the brat shirt. It's the main subject of the image. He's looking directly at the camera. He says, style, thank you. He has a slight smile. He's like, yeah. I'm a very serious looking person. So add a friendly, approachable aura to the image. Okay. That is the description based on this multimodal multimodal model, okay, that describe the photo. Okay? Then the next one will combine this description as my hobby with this thing, which is based on the MISTRA. You suggest, okay, some description to modify that based on what I described here plus photo and try to transform me into Lego looking person. There you go. K? So the stable fusion, okay, is the one that is the last border that's remember, just we modified. Okay. I don't know whether this is very low like me. Okay. But in spare, okay, try the red hair, playing basketball. Yeah. Of course, you can try it yourself. Okay. Take your own photo and try. Of course, I can also do things like if I wanna go back I mean, for sale type, you can also try to convert yourself into a superhero or sexismistic figure or to be a kitten. K? Try that out. Let me know. K? So this to demonstrate to you that how easy it is that you can manipulate around using different model or craft this to your own way, your liking, your requirement. Hopefully, that's a fun part of the coding. Now we go to the last demonstration, which is very important. I pass you the link. This link do not need you to do any coding because it's in a public website readily for any user to try that out. I will bring a window Okay. Hold on. Here. Okay. And I go to this. Okay. This is hosted by CrowdStrike. It's ready made to populate. You can go to link, follow me, right, anytime from here. So let me click all things first. I'm gonna show something very nice here. So first of all, to have a microphone. I want to do a speech to text to demonstrate something else, and I will need to write something. No. I want to do image. Let me try to delete the work first. Okay. Image, audio, and I will use numbers m p. We should instruct to do that. Okay? So later, the image will go to here. The audio, which is my speech, will go to here. Okay? A 3.28. It'll be okay. So sorry. I need to use my whisper that is doing a speech to text and then text to text. Okay. So as you drag around, as you deploy this, what happened behind is that, automatically, you know how to okay. The the script would know how to divert converting, transforming your your media, be it voice or text or image from one part to another part, from one model to another model. So right now, I need to choose a image. I'm going to try something here. I say maybe financial report. Okay. I just draw from any images. I'm not too sure whether we should k. Maybe I just say this one. Okay. I just save image to my desktop. I will say that this is financial report o one. Alright? Save. Come into here. I say I choose my image. Financial report o one. Alright? I put that int into here. Then I will say that okay. Let me do recording here. Please explain what this image is doing. Okay? Record it and then run. Give you a moment. I'm waiting for the output. Okay. My pronunciation is not that good. It thought that I'm talking other language. Okay. Let me try again. Alright. Maybe I'll say that analyze the image. Something not right. I don't know what happened to my microphone. It's not converting correctly. It was it was doing well just now. Check for any error. I see how okay. Now it's correct. Okay. If you write Python script, right, you know that Whisper is one of the Python library. Okay? So you converted my speech into text called check for any error, And then as as the image being inputted in this normal model. Okay? Then now you see what you can see. The table provider seems to be a financial statement, which is true. Okay? That's brilliant. It's a balance sheet. Okay? Then as you go out, doing the structure of contain columns or issue to check for any error, which is my request in my speech, one would typically look for inconsistency without special information or what? You could provide more information being described. Right? So okay. Rightfully, if I provide enough detail, maybe the image is not that clear. Okay? It's able to demonstrate things. I already find out the error. Okay. So if I have already captured the image previously in one of the financial report, which is better, You can see my image here. Okay. It's the same thing. It can really tell me that these are the so called error they detected. Alright? So for the same so called methodology, you can drag and drop. You can make use of things like any let's say, you want to do a webcam. You can drag. Okay. Then you can see my camera, which I'm not gonna do it now. Then maybe I don't know. You can this is for the imaging producing. Okay. You can drag and drop anything you want. Okay. Let me create this. So there are some ready made or things like maybe image arena. Let's load this. Okay? So maybe I'll say that this will be sunflower, for example. Okay? Then let's try. You go through a series of this thing. Okay? Sunflower. Okay. You find that, remember, this is my prompt called sunflower. Then this is what I instruct the model. If you remember just like in the first demonstration, I got this. So because it's the prompt to tell the model what do I expect this to be. Okay? So my instruction say enhance the above prompt, which is this guy. Can feed this in image generation, that kind of thing. So by doing that, it's able to call many other model, number one, phonics one point zero, Lucid Origin, case stable execution, FRAS one, FRAS two. So it demonstrated that we can run multimodal in a single application. Although now it's dropped in a very visual way, you can do the encoding, whatever, alright, depending on your how you want to provide the visualization, data visualization for your user, how you want your UX UI to be in. K? So, hopefully, by doing this, this really excite you. K? Everything is done on a Cloudflare platform easily. K? Simultaneously, you can just use the same base code and then just modify your model, tweak it a little bit based on the input requirement, and you get to see what you want to see. Now if you ask me, what's the difference? If I have to run-in Gemini, run-in JetGPT, why do I need CrowdFlare? The number one thing is that if you are only using a single model, by of course, by all means. Right? But the cost will be higher. You need to pay the subscription. Let's say you use a friend ChatGPT. You pay, I don't know, maybe one month, $20.30 dollar of defending your enterprise will be higher cost, so so and so forth. Or maybe application require multiple model. How that means you need to subscribe. Pay a fixed sum of money to different provider. So in terms of handling, finance, management is going to be messy, chaotic, and cost saving is not that easy to achieve. So what we try to offer to use that, everything go to Cloudflare. And behind Cloudflare is all the model that is there's a readily for you to ingest, for you to make use of. But the question is, how much the cost? Okay? So let's give you the guide on that. Let's take a look. Model pricing. Go to this link. Owner. Okay. I tried to put in the chat box. Okay. You go to this link. Let me show you here. Alright. This will list out all the pricing. You do not need to subscribe to any of the so called LM model. It all go through us. So we work by what we call the neurons. Alright? So as a free account or pay account, the first 1,000 neuron is free for you, k, which is going to $0.0.011 cents. How does it work? Per day, you got 10,000 neurons that you can use free at no charge. Anything above that, then you start to charge that then you start to be chargeable. Like, let's see. For free 1, 10,000. Pay one also got 10,000. Okay? But for the pay one the pay account, every 1,000 neuron is 0.011. The question is, what are neurons? You know? Okay. You can you can you can you can liken that neurons are something the the same as the token, and a token is current to one word. So you can think that a neuron is equivalent to one single word of input or output. I mean, the the problem let's say the problem is a sum for our test two words. So two neurons. Okay? But by right, technically, one neuron equals to one token equals to 0.7 word. So depending how you do a mathematics. Okay? So one word equals to 1.3 neurons or 1.3 token, if you think of that. But for the state of an easy mental mathematics on computation, just think that it's one to one. Easy for you to compute because the difference is not that much. K? So if I'm consuming from things like the normal three dot two, there will be 0.02¢ per 1,000,000 input token. Okay? So that means 1,000,000 worth is only 0.02¢. That's super, super cheap. Okay? So in top in neurons will be 2450¢. They they are equivalent. K? There's a output. That means a return for the model. You're charging as a 2.201 per million up output token. If you got if if if your application is running into a millions of so called input per day, that means your business is doing very well. Then this sum of money is I mean, it it means nothing to you. Right? Okay? So the rest is the same thing, like, depend on model. Let's say you're on the top about, let's say, image model pricing, there'll be different chart. It's even cheaper. If per 5.512 time one five zero two tile feature is only zero point zero. I I don't know how to describe that. $1.02 $3.04 $5.06. Wow. It's as though it's cheap as though it's it doesn't cost anything. It's 4.8 neurons, okay, per five one two five one five two image. K? Like, if you want to consume for another, there will be six three six neurons. Audio, model, other model, pricing, all price out here. You can do your mathematics. Okay? So by looking at this, what does it mean? One thing using Craftware developer platform to do your AI processing, That means it's a so called single pane of glass. You can write your base code and consume for all model with certain tweak of the input parameters just like you're doing API call. Okay? You do not need to subscribe to any of the LM vendor. Everything go to us, and we we we charge you very, very cheap, okay, in terms of the consumption. So I'm done with the demo. If you got any question, please put in the q and a. And I just wanna encourage you to try that out. Even if you are not on the pay model, you try on the free model. The one k. If I do remember wrongly, 10,000 neurons per day. That's enough for you try out with your basic testing, basic consumption. Any more than that, you just pay a little little I mean, just a raise a small sum of money, okay, for the for the rest of the consumption. Right? Okay. I'm done.