222 | Scrape, Analyze, Generate: Build Scalable Content Systems That Works with Nadia Privalikhina

0:00Hello, and welcome to another live episode of the Leveraging AI podcast, the podcast that shares practical ethical ways to leverage AI, to improve efficiency, grow your business, grow your business, grow your business, grow, and advance your career. This is Isar, Maitis, your host. And if you are listening to this podcast, it means you are consuming content that somebody else generates, in this particular case, me. But generating content in general is a great way to build an audience and create relationships. And if you nurture these relationships and you continuously provide them value, it will over time lead to business.

0:31I've been a big believer in that for many years when I worked for large companies and now when I'm doing my thing right now on my own. But either way, creating valuable content that really helps a specific audience that is your target audience is a great way to grow your business. However, knowing exactly what to post and how you should post it is not that easy. Many organizations and many people have been running very, very fast on the content treadmill, but going nowhere because they're not necessarily sharing content that helps their target audience in a meaningful way. but to do that properly it's actually a lot of work right you got to do research and figure out what's actually connecting with these people you got to do research and figure out what's trending right now and what they might be interested in you got to then analyze that content and figure out which components they might be interested in and only then you need to start generating the content and generating the images and all it's a lot of work and so doing that manually is something that many companies and definitely individuals are not doing or at least not doing effectively However, AI is a great platform that knows how to do all these things.

1:37It knows how to do research. It knows how to analyze data. It knows how to create content. So if you can take each and every one of these components with the right prompts and then combine them together into a single automated process, you can create an incredibly effective content machine, which is exactly what we're going to show you how to do today. Now, our guest today, Nadia Privalikina, is an expert in AI automation. She's been building AI automations for multiple clients in the past few years. Now, her background is being a software engineer, which means she definitely understands the technical side and how to build things that flow effectively with minimal interference and issues.

2:16But she also spent a while running an e-commerce business, which means she definitely understands what motivates people to buy and how to connect with them in order to drive them to take actions, which this combination makes her the perfect person to show us exactly how to do this process. Now, when I saw her post about this on LinkedIn, I'm like, oh my God, this is absolutely brilliant. I need to bring her on the show. So here we are, and I'm personally very excited for this episode. I know it's pure gold. Again, you will see how to do all the things we've just talked about in an automated way, leveraging AI together with automation capabilities.

2:51And so, Nadia, welcome to Leveraging AI. In the next few years, AI technology will change our world dramatically. Whether you are a business executive trying to catapult your business forward, or just somebody who refuses to be left behind and want to advance your career, this is the show for you. I'm your host, Isar Matis, a serial entrepreneur and an AI enthusiast. You'll hear invaluable practical tips from innovative business leaders, AI practitioners, and some of the brightest AI minds in our world today on how you can leverage AI in ethical ways to advance your career and grow your business.

3:36Thank you so much, Yusuf, for such a nice introduction. Today, we are going to talk about how to create YouTube thumbnails specifically. That is the use case that we are covering, but I believe the same concept can be applied to different industries and different types of content. The automation that Nadia showed is how to take what's working for other people on LinkedIn, combine it with the stuff that you want to include in your thumbnails, and then create thumbnails on the fly for your content. The same thing can be brought into anything that you want to create by researching, analyzing, and then creating based on the same exact parameters.

4:13You'll see it. It feels like magic once you'll see the results. For all of you who are joining us live, whether you are joining us live on LinkedIn or whether you are joining us live on the Zoom call, then feel free to ask any questions. First of all, introduce yourself right now in the chat. Say where you're from. Say who you work for. Send your LinkedIn link if you are on the Zoom chat so other people can connect with you. And tell us in two words what is your biggest interest in this particular use case. and if you're not here with us live, the question is, why aren't you here with us live?

4:48We do this every single Thursday, noon Eastern time. We have amazing people like Nadia that are going to share exactly how to do different use cases and then you can ask questions, which is the last thing that I'm going to say and then I'm going to give it back to Nadia. Feel free to ask questions in the chat on LinkedIn or in the Zoom chat and I will bring it up to Nadia and we're going to answer your questions. With that, Nadia, the stage is yours. Walk us through your magic. Yeah, thank you. So I think that I will share my screen because it is the best way to showcase what we are going to talk about.

5:18For those of you who don't see the screen, by the way, because you're listening to the podcast after the fact, then we're going to explain everything that's on the screen so you can follow us as long, despite the fact that you're driving or walking your dog or on the treadmill or whatever it is that you're doing. However, if you have the opportunity to also watch the YouTube, then that's another great benefit because you can see afterwards what we've done. Okay. And right now I'm sharing the screen with my LinkedIn post that attracted that attention. And what happened is that I am part of Liam's only private group where we learn how to use AI and how to sell AI services to businesses.

5:55But actually, Liam runs a big channel with more than half a million subscribers. And one of the problems that he faced is that they need to constantly create thumbnails. They do research, they create those different variations, and then post those variations. So there was a hackathon, and I created the system. I will quickly run through it right now. But here you can see the first examples of what types of thumbnails the system produced. And yeah, I haven't mentioned that I won the hackathon. I took the first place with those images. And since then, a lot of people reached out saying that those are amazing because if you don't know how Liam looks, he looks exactly like those images that I'm sharing right now to those who are listening.

6:40You have to believe me. And that was pre-Nano Banana, right? So now this is probably even easier. I would still say that character consistency can be a problem. So I tried after Nano Banana release to do the same with it, but the character consistency is still a problem. So I will tell you how I approached this problem and how I solved it. So again, those of you not seeing, there are multiple thumbnails with a very clear style. So they have these kind of like beige background or kind of like off-white background with graphics that all look the same as far as like sketches and images of the person in multiple directions and angles.

7:22But his face is in most of these thumbnails. And they look very consistent from a style perspective. And they look highly professional. And yet no person has created them, which is the whole magic here. And maybe even share a few examples a little bit closer. So here are those finished thumbnails. Amazing. Okay. Let me exit this screen. And now I will tell you how I approach this problem and how you can do this. Awesome. Just a quick question. How long was the hackathon? Like how much time did you have to work on this? Well, in theory, we had almost one month. But because I have my clients and I believe everyone else faced the same problem.

8:07So we finished it in around three or four days. Oh, wow. Okay, cool. Okay. And now I will share. So to prepare this and how would you create a thumbnail or any other content, you would first want to understand what works and what works for others. And this is what I also did. So now you can see on the screen. I believe this will give it a moment and it will display the thumbnails. I scraped all the thumbnails from multiple creators on YouTube and created my database of Vinyin thumbnails. And later, initially, my idea was to analyze those thumbnails because you have models that can transform visuals into text, for example.

8:50And then my idea was to train to fine tune a model. But actually, it didn't work, even though I use that fine tuning for my own LinkedIn content. Interesting. But in this case with thumbnails, for some reason it didn't work. So what I ended up doing is I created, I scraped multiple channels. Obviously there is still some human work. I had to pick those channels on my own. And then the system could, let me show you what is going behind the scenes of that scraping part. So now I'm sharing my NITN workflow. It is super easy and quick. and someone mentioned that it's probably not legal to scrape YouTube, but actually YouTube gives you quite a lot of data points, including thumbnails that you can get absolutely legally with some limitations.

9:38So first I set the channels that I want to monitor. And then for each of those channels, I get all of the videos from the past year for them. And you can see... When you're saying get the video, you're getting the thumbnail and the description? Let me show you what we have here. YouTube provides a lot of information about those videos. So you can see we have title, we have channel, name, it is somewhere, description, and then thumbnails. Yeah, thumbnails are here. And this node in NA10, that's what it does? It just grabs information from specific channels? Yes. So it takes a channel ID and I say that I want to get all the data from past year in this example.

10:22And then I get all of that video. So it found, I believe I entered around 10 channels and almost every channel posted more than 100 videos. So I got all of those 100 videos for each channel, which makes it up to 1 ,000 items in this case. And later, obviously, I want to filter them. And yeah, there is another data point I will not be sharing today because it will expose my API key. But there is another option to get more stats about that videos. views count and comments and whatever you want. I would say there is still a lot of information that can be... And that's from what tool did you use that?

11:01Is that like an appify? No, it is also YouTube's official API. Oh, so it's YouTube's own resources. Okay, cool. Yes, yeah. I just... And it doesn't give you an option to list all of those. So in this case, I used... So to explain to people how these work, so all these tools work the same, whether using NA10 or Make or Zapier or any other, they, in all their tools, they don't expose every single function of the API that exists in the tools they connect to in this particular case, YouTube, which means you can connect to the simple node and it will give you several functions that you can do in the node, but you can use an HTTP call, which is very, very easy and you don't have to know how to code.

11:42You just have to get the documentation and drop it into your favorite large language model and say, I want to connect to that with an HTTP call and I want to find that information. They will give you the code and you can drop it here on NA10 and then you can call the rest of the functions that did not exist and was not pre-built into the tool itself. Awesome explanation. Thank you so much for that. And after we get that information, in my case, I wanted to filter out short form videos, but for some people, it may be the opposite case. So they would want to see what's trending for short form videos.

12:13And I had to use a little bit of code for that, but I believe that it's also possible to do without code because I have technical background for me, it was easier to use code in this case. And then I just filter out short videos that are shorter than two minutes. And after that, I get also, because we scraped all the videos from the past year, we can get data about what is the average view for this specific channel and what is their average likes rate. And this will help us to identify what are the outliers for this channel. Oh, cool. I see where this is going. Yeah, it's not that simple. I mean, it is quite simple, but you need to have your own strategy.

12:56This is where the human part comes in. And so, yeah, we get the average views per hour for this channel and average likes per video. And then this information helps us to identify which videos are good and which videos are not that good. And after that, I use that information and save it to my database. And again, to explain to people, when you say database, people are like, oh my god, I don't know any databases. It just saves information to Airtable. And those of you who haven't used Airtable, Airtable is like Excel with a better and easier to use user interface, right? So it allows you to do more filtering and control more the look and feel and add images.

13:33Like you see in the example here, you see all the thumbnails from all the different channels that this process has brought, but it's saved into Airtable, which is a very user-friendly tool. you don't need to know how to run databases or set them up or so on you literally go to airtable create a new space in airtable define what you want to have in it is it images and what's going to be the columns and what fields do you want in it and then you just can feed that from whatever source in this particular case from any tent absolutely i have non-technical clients who are quite uh who are navigating airtable very easily yeah and so i save those uh thumbnails and then I have one more workflow that actually analyzes those thumbnails.

14:14So I use an agent in this case, but why I use agent? Maybe it is not so specific and you can use, let's say, your OpenAI can analyze what is on that image. But in my case, I decided that I will do this with agents. It's just a small limitation of anything that I found or what I wanted to use that I had to pass the image and use an agent for that. So my agent analyzes those thumbnails and saves them to, again, to an air table. Can you explain what exactly it analyzes? Like, what is it looking for? Okay, let me see. So the prompt says that you are a world-class YouTube strategist with attention to detail.

14:53You live and breathe YouTube content marketing strategy packaging and are skilled at creating winning packaging for high-performing YouTube videos. You always output value JSON. So this is one trick that we just needed for a structured output from this tool, but it is not mandatory, I would say, in this specific case. And then I give an additional extended prompt where I say that this agent will be given a YouTube video that performed well. And, okay, I also feed it with a transcript, a title, and thumbnail. And the mission of this agent is to deconstruct why this thumbnail works for the video and identify the thumbnail archetype.

15:33And this way... Before you close this, again, to people. So what is JSON and how does it work? So JSON is a very simple coding language that basically describes the data as if it is a table. So it tells you which fields they are, what values they can have. And the benefit of using JSON is that more or less every tool out there knows how to use it. And so if you export and or import that, then you get consistent results because it's going to come in a very well-defined format. And again, you don't need to know what it is. You can go to whichever tool and say, I need a JSON that will give me these parameters and it will write the code for you and it can paste it in here.

16:12The benefit of that, once the data gets into an automation tool like NATN or MAKE or whatever, you can identify each and every one of the fields. You can map them to the columns in the table in Airtable and so on. So it just gives you a structured output versus just free text. Exactly. And maybe I will... Okay, let's move on. And so this system could look at all the thumbnails and transform them into text representation. And then I save that information again to Airtable. And so by the end of that second workflow, what I have is a lot of structured data. I have titles of highly performing videos, channel names, summary, and then thumbnail description.

16:56And for a thumbnail description, it says what is in the background, what is the text on the thumbnail, and a few other things. What is on the foreground? Those are the main components. So you're basically deconstructing the thumbnail into its various components. And because these are all the top performing videos from the specific channels, then that hints that these are solid descriptions and or thumbnails or a combination of the two, which gives you hints to what you need to do in order to create successful ones yourself. Yes. Now we are just preparing the data for the future automation. Awesome.

17:38And this is the data preparation step. What we do now, Airtable allows you to export that data. I don't remember where it is, but you can actually export all the table into a CSV file. And what we do next, and you can do it with any types of content. I did the same with LinkedIn content. So you export it. And then in my case, I went to Google AI Studio, or you can go to Gemini, but Google AI Studio is a similar version. So it is just Gemini for developers, where you can safely test your ideas and experiment with that model and how it works. So what I did is I went to this Google AI Studio and I exported and I gave it that CSV file with description of best performing videos alongside with the thumbnails because we have already transformed the visual thumbnails into text.

18:26And so Gemini can work now with that text. And why Gemini? Because it has a huge context window. It has more than one million token context window, which means that you can basically feed it with, in my case, there was around 100 videos, but you can potentially give it 2 ,000 videos for analysis. And so after that, after I give it this CC file, I instructed that you now see top-performing thumbnails for videos, their description, explanation, and so on. And your mission now is to generate a prompt for an LL. This is also a secret trick that I use. I always ask Gemini to create prompts for me because it knows how to create prompts better than me.

19:08And also I give it a link to the prompting guide for Gemini. Google publishes those prompting guides. They are officially available. So I gave this prompting guide along with the CSV file. And now what Gemini did on this step is that it came up with a good prompt. It now created a structural prompt with a persona. So again, it says that UI world class YouTube strategies, and it is not my prompt already here. This is what Gemini created. And it created the task. So task is to analyze and provide video title and summary and generate, in this case, in the context. I mentioned that I want to generate based on the title of the video and transcription of the video, I want to generate those thumbnails.

19:55And so all that information came into the prompt. And later I made a few modifications to inject the branding guidelines that I wanted it to have. But the structure is the same. I will just iterate upon the first version and ask to add a few more components to it. I believe this is the second one. I ask it to deconstruct and reverse engineer how I can get a clear description of a thumbnail based on title and transcript. And I used one of those prompts later on in Insights and Automation. So the prompt, again, just to understand what it does, it takes all the inputs from the CSV file that was created based on all the data that we collected previously.

20:37And now it gets an input for a new video. I ask it now to create prompt for basically itself and say that it will be given, yes, title and transcription. of the video. And it needs to then define what the exact... Yeah. Okay. What the exact thumbnail needs to be based on the best practices that it learned from all the previous material, combining it with the new need for the new video. Something like that. Yes. Okay. Is there a section for like brand guidelines and something like that? Because again, we saw everything was perfectly branded. Yes. later I injected it manually. So you added manually a segment about the branding.

21:23Exactly. Yes. And after this step, I have a prompt that I can use. But let's go back to the this is now on the screen, you see the final automation that I created, it starts with a form. So we need to somehow instruct the system that hey, this is our title. And hey, this is our, what is our video about. And so it starts with a form where we feel that information. and later it can summarize the transcript if it's needed. And then I use that exact prompt from Gemini. It has this persona about world class, YouTube thumbnail designer and task and instructions and so on. And then it comes up with a few concepts.

22:04We don't want to have only one thumbnail because most of the channels test multiple thumbnails and we want this system to create multiple concepts. So it comes up with a few concepts And it is based on concept archetypes and so on. And it outputs visual descriptions of what should be in the background, what should be the primary subject. Is it a person or is it just an object which fits this video? And what are the elements? If there are some arrows or additional, I don't know, a pile of money or something to evoke some emotions. And then text elements. And this is all what we need to, almost all what we need to generate and visualize those concepts.

22:46But what else do we need? If it is a personal brand, then we obviously want to have a person on those thumbnails. And if you don't know, then OpenAI has this image model, which allows you to reference images. Now we have NanoBanana before, we didn't have it, but still OpenAI has that possibility to reference not only one image, but multiple images, and it also can be done via API, which means that it can be automated. Yeah. Now on the screen, you see one of the examples that OpenAI provides. There are four objects, and then if you give it a prompt to compile those objects into one image, then it combines all of them and puts them into one image.

23:32And we use a similar approach with thumbnails. Okay. So what I did here is I used a few thumbnail examples from the channel. Okay, let me open it. So here is one example and there are a few more. So now I use those three examples. And then there are just a few supporting blocks. But later I ask an open AI model. So I reference those images, multiple three images. And then I say that again to that it is an expert YouTube art director. and a master of photorealistic concept executions using Generative AI. Its sole mission is to create a single masterpiece-level YouTube thumbnail based on the detailed concept provided below.

24:17You must perfectly fuse the visual style guide and the perfect and specific concept into one cohesive hyper-realistic image, and so on. And here I also give it additional references. So I mentioned that the images that I provided are the reference of a creator of this channel. And I also give additional styling guidelines right here in the prompt. So we give styling guidelines into multiple places because the first place, it affects the layout of the image. And in the second place, when we actually generate that image, then styling guidelines affect how it looks and what colors it uses. And so let's actually view.

24:58So this is one image, another input image, and then we give it a concept. And finally, I will not show it right here, but let's go here. Okay, this is... There are a few examples. So again, what you're not seeing is that basically what this step does in the automation is it takes basically everything we've prepared so far, but mostly it has a prompt on exactly what to include in several different variations of the thumbnail. So it takes these, as you call it, archetypes of different what works, basically, So it's going to generate several different versions to use. It's using images from previous thumbnails to get a visual reference of the brand guidelines, as well as the look of the person.

25:46And then it combines all of that into outputs. So every time you run this, you're going to get X, I don't know how many, three, four different options for thumbnails that are already aligned with the brand, that already has the person, the text, the background, the supporting graphics, whatever needs to be in there already in the image. Yeah, exactly. So we first distilled what is working into a prompt, then that prompt created a few variations, a few text representations of variations, and then we converted that text into an actual image together with a few reference images that we use them only to reference base of a creator that should be on the thumbnail.

26:26But if you are watching it, you see that the final images, they are not perfect. So even though we give ChatGPT the model behind ChatGPT reference images, it doesn't preserve that character so well. And that is why we need additional steps. And anyway, this image is not in the proper format and so on. So we want to do the face swap, and face swap was one of the hardest things to figure out. But what we found is that there is this flux context model. You probably know about it before. Again, before Nano Banana, flux context was one of the best models for visual editing. And with it you can modify some text on an existing image.

27:11You can transform that image into a different style. And one good thing about it is that you can actually create a fine-tuned version of this flux context. And what it allows to do, let's say we have an image with not a perfect face. Let's say we have this image and we want to use, we want to have an image with an accurate face. So we would train a model that has an input image with someone else's face and the final image would be the face that we want to see. And in this way we can later on run that fine-tuned model on the images that our chat GPT model produced. I hope it makes sense. Yeah, so I'll read.