Let's dive in. Okay, so the first demo is a state of the art figma to HTML generation agent. It was created for a company that creates newsletters for some of the biggest brands in the world. So basically what this agent does is it creates HTML files directly from figma design mockups. All you have to do is provide it with the necessary assets, the figma mockup, and then it proceeds to work. So as you can see here, it uses one special function that is extremely complex and uses a lot of other, API calls to OpenAI under the hood and then it generates this newsletter that looks almost exactly as the target design mockup. This process is obviously extremely time consuming and it used to take a lot longer, but you might be wondering, you know, what's the reason for using an the agent here since the agent just uses one function and then provides us with the HTML file So the answer is that because often times after generating an initial HTML you have to iterate with the developer and adjust it based either on your or your client's feedback. And the same can be done here instead of you just getting that final HTML file, which would happen with the AI automation And then you would have to adjust it manually. Here you can say to the agent the specific adjustments that need to be done. As you can see here, Nick enters some specific instructions for which sections to adjust and how. And then the agent proceeds with aligning these sections better according to the feedback. And as you can see, the generated HTML closely resembles the design mockup, which is amazing. A key takeaway from this project is that you must always ensure that the agents are flexible enough with the provided tools so they can adjust according to feedback. If you are new to this channel all of our agents are developed using my own agentic framework called Agency Swarm on GITHUB If you wanna learn it make sure to check out my previous videos later Okay, so the next project is a unit test generating agency developed for an I.T. consulting firm called “Mobiik” that specializes in software development data. AI and cloud services. The problem was that the process for generating unit tests was relatively simple, but at the same time extremely time consuming. And it couldn't be automated with a basic automation because obviously each report is different. That requires different instructions and again, it requires iteration based on feedback. So here we decided to integrate this agent directly into the same Azure DevOps environment, which their developers use daily. As you can see here, we have a board with different statuses, and this is exactly how this company used to perform this process manually. You know, once you have an item in specific column, it means that certain tasks need to be completed. So here's the first issue. You know, developing tests for .MP3 files just as an example. And on the right, we can we can also see the Docker containers to monitor this process. So let's now drop this task into the Create Technical Report column. Okay. As you can see now in the Docker container, the task has been received and the processing has started. The system then searches and analyzes the code base based on the status of the item, just like a normal developer would. And soon as you can see, a comment will appear here with the revised user story and the generated technical report. So the technical report includes introduction and code analysis with code snippets, user story analysis and the conclusion. So next we create the test plan by dropping it into the next column. And again, the benefit here is that you can also adjust those generated stories and generated test plans before proceeding to generating the unit test. This allows you to control your agent much better. And finally, we drop this item into the unit test column. So the system will now create unit tests. It will create a new branch of the Azure DevOps environment, and then it will create the pull request with the generated unit test for review. So you can see these unit tests being created here and then the new branch is made and now you can review the commits in the repository showing the new files created by the unit test agent. In summary, this simplifies the whole unit's generation workflow and seamlessly integrates into the client's business because we're using the exact same systems that this client's developers use daily. So the key takeaway is that you have to integrate your agents into the exact same systems that your clients use daily and the cost for all of those user stories. test plans and unit tests However, keep in mind that is a very basic example. It's not a real production repo. And in a real production repo typically there is a lot more code and a lot more complex unit tests that need to be generated. All right. The next project is a very common data analytics use case that we developed for an online payment processing company. The problem here is that they have a lot of different tables and data sets. And I mean, like, you know, not 10 or 20 tables, but 100 of tables and extracting insights from those tables is not only time consuming, but an extremely challenging task. So all we have to do to use this agent is simply to send it what query for what we are looking for. It can include, you know, dates, specific metrics and maybe even aggregations. So let's try another query and think I can show it to you here. But as you can see, the agent can perform those queries for any metrics that you want. Yeah, and the costs were only a few cents depending on the amount of context that the agent needs to retrieve in order to perform the query. As I said in my previous video, those data sharing and data analytics use cases are typically my favorite to get started with because it's first of all, extremely easy to implement it. And secondly, it's extremely easy to build upon. So once the agent has access to all of that data and their data sets, we can not only allow it to query that data and provide us with the insights, but also take certain actions on our behalf. So this would be the next step in this project. It would be making sure that the agent can actually perform certain actions, like for example, detect fraud and then block certain user accounts. Okay. The next agent that I wanted to show you is a Tare Sheet Agent This agent was developed for a marketing agency called ESM that helps brands to run successful marketing campaigns and visualize the results. Tare Sheets in marketing are essentially a special PowerPoint presentations where you can see how different ads will look like on different websites. So to test this agent, we first have to upload the ad creatives and then insert into the message all of the websites where we want to preview them. The agent will then proceed with scraping the website and converting it into an image using Flash API. I always do recommend using third party APIs whenever you can because even though we could implement this tool with selenium using a third party API is easier because first of all, it simplifies the deployment and second of all it significantly simplifies the management of this agent. So whenever you can create a tool using the third party API, always opt out for that. Okay. So then as you can see, the agent replaces those ads on the web pages that we just scraped and finally provides us with a downloadable link to the PowerPoint. Let's see what it looks like. Awesome. So as you can see here, we got the first New York Times website with the ad Creative. Clearly replaced on top. And we also have Forbes but since like Forbes didn't have any ads. And finally we have TechCrunch also with those ad creatives correctly placed instead of the current ads on this website. The way this ad replacement tool works is actually quite interesting because it uses another custom AI model under the hood. So as you can see here, we're calling YOLO that is deployed on the same infrastructure for the client. And then this tool performs object detection and then replaces it with the ad creatives provided by the client. A key takeaway from this project is that you should always use premade solutions for the tools if possible. So you know, if there is a pre-made ad detection model like the one we use here or an API available that performs the exact same task, always just wrap those API’s or external code bases into a tool so that you can save on development efforts. And the cost for this agent are quite low because again, it runs a custom model on its own infrastructure and doesn't even use tokens for that. Okay. And the final agent that I wanted to show you is an AMD sheet agent that is made for a business that owns multiple e-commerce brands like this one. And essentially they needed an agent that would fill out essentially special product sheets that are then inserted into marketplaces like Zalando in order to post the product. So let's try to upload some product information required to fill out this sheet, like, for example, packaging materials and other product info in .CSV files. Okay, So now I'm just going to copy the product name and then hit send. The agent then asks me which fields do I want to fill out. So I'm just going to say, “all” so then the agent is gonna check this product from one of the sheets and then proceed with filling out the AMD sheet So finally the agent finished processing all the fields in the .CSV files, and then it tells us that there were some errors. So the user can again provide some feedback and correct them. However, the real benefit of using AI agents here is that after we have this fully filled out AMD sheet here with all the necessary product information, as you can see right here, in all of those fields, basically the client then sends this AMD sheet to the retailer and the retailer then checks it and sends another sheet back which the client also has to fill out. So basically what happens next is, you know, the client can take this new sheet and then select another agent that is specialized for a processing this next flexible AMD sheet. So then the client can re-upload the file and again start the processing for the next sheet with another agent. In conclusion, I wanted to provide you There wasn't a single agent that didn't demand some coding experience. Now, keep in mind that there are a lot more caveats that you have to consider when deploying such agents in production, which is why I recommend watching my previous video now, where I discussed a lot more of those insights and answered the top 12 common questions about where and how to deploy AI agents in businesses.
