ChatGPT Unleashed: Revolutionizing Web Development from Idea to Deployment

Hey, everyone. Welcome to Chad GPT, Unleashed, Revolutionizing Web Development, From Idea to Deployment. The journey that I went on in this presentation and that I want to take you on as well about my lessons learned is how good is GPT-03-Mini-High at building, testing, and deploying a web application. Right. We all know that it does have some coding capabilities. We all know that it has, you know, reasoning capabilities as well, but what can it actually do and how well does it do it when it comes to building some kind of a web application? And so that's what we're going to figure out today.

So let me tell you a little bit more about myself just before we jump into the presentation. My name is Nikolai, originally born in Uzbekistan. I'm a Russian Ukrainian Jew, migrated to US in 1996. I was raised in like Maryland, Virginia area until I got tired of the snow and the cold, and then I moved to Florida where I've been residing for about six years now, and I absolutely love the blue skies and sunshine. I've been in automated testing, software engineering, DevOps industry since 2008. I love it. It's my passion. It's what I do pretty much every single day since I ever started. Recently, I've been working on my first SaaS startup as a CEO and founder of VolQAI where we are building automated software testing that's actually fun, that actually works, that's actually easy, and makes software development processes extremely better and extremely more reliable using our technologies. And in the picture, I'm a dog dad. You'll see my little baby Mia. She is always here with me as I work from home bossing me around.

So I actually started this journey back in 2022. I don't know if you all remember back in 2022, that's when ChadGPT made its first entry into the world and shocked the world. I think, what was it? Is it number one fastest growing application ever? And they got like millions of users within the first week or something. So it made a huge splash, and so of course I jumped on and tried to see what is this thing? How does it work? And so I think it was about February of 2023, I gave my first presentation at a conference in California about its capabilities. I tried to build a web application with it then. I'm still continuing to build applications with it now and including other technologies. This is how it performed back then in different areas. Ultimately if I was to give you like a TLDR, is that it's good. It was good at technologies that are slower to change, where documentation is better. And it was very poor with rapidly moving technologies, where of course it was trained on data that it couldn't keep up with. So whenever APIs changed, it couldn't keep up with those API changes. And so there were many challenges that broke different kinds of uses of those rapidly moving technologies.

All right, enough talking. Let's actually go for a test drive here and see what we can make happen. So I wanted to start out with a concept and design for the application and the application I wanted to build that I've had a dream about building was a vision board. I have a vision to build a vision board. I've never built a good one. I will continue to try whenever I find free time. And maybe over the years as different LLMs get smarter, my vision board will finally come to life as I continue to do these presentations.

But ultimately, starting with a concept and design. These are the wire frames that the LLM provided me. I did give it a prompt to build me a vision board, and I supplied it a screenshot of my requirements that I had all the way back in twenty twenty two. And so and this is all the information that it told me afterwards. And so you can see here are the different requirements like the log and register page, vision board, the gold editing and responsive nature, which all of which I really like. And so I gave it a little happy face saying that I did a good job.

Here are the user stories. Granted, these user stories came from my original vision and this is exactly what I pasted into the prompt. Actually it's a screenshot, which is super useful. As you can see, it even says here from requirements. And so all of these actually look really good to me. So I gave it a happy face as well because I think that the LLM here did a good job as well. Here are the requirements. There are a lot more, but this is just a screenshot of some of them. Like you use a registration, there's the requirements. It provided some acceptance criteria, which I really like some user user login also with acceptance criteria, even getting to the details of like using JSON Web Tokens or cookie based authentication and then talking about adding goals. And there are others. I think that this is also really good, especially with the acceptance criteria where it kind of gives you a really good idea and really more clarity about what is and is not acceptable. Then it had some design principles that it provided. Simplicity, love that. For me, a clean intuitive interface is always very important. Responsiveness for a vision board. That's extremely important because most of the time I'm actually going to be using it on my mobile phone as I'm walking and walking around, walking my dog and looking at my vision board or maybe I'm going to Whole Foods and looking at my vision board, visualizing my goals coming true.

So mobile mobile responsiveness is extremely important. Accessibility. It's nice to have for us here in U.S. accessibility is less important. Unfortunately, we don't have legal standards that we must meet like you, but it is important regardless. And then there's security and scalability. That part, I think, is important as well.

In terms of error handling. Once I started having the LLM trying to take me through the process of actually setting up the software on my local machine in order to be able to deploy it, it didn't do a good job. I actually spent a good amount of time, probably an hour troubleshooting different kinds of errors. The installation actually went pretty well. But then when we started getting into the installing Tailwind, that's where I started to run into a bunch of issues and it took me some time to troubleshoot these issues. And so I don't think it did a good job. Of course, there are so many dependencies. I think ultimately it ended up being a version of Tailwind that was installed that I needed to downgrade in order for it to work.

The version of Tailwind that I was using was like an alpha version or maybe a beta version that probably wasn't ready. But it was funny because the LLM was the one that recommended me to set up instructions. And so I guess the fact that it couldn't anticipate this or didn't anticipate that, you know, may not bode really well for its reasoning capabilities. But ultimately, we were able to get something up and running, which I thought was actually impressive. Having started this in 2022 to now being 2025 and already being on many versions later since ChatGPT 3.5, I think what we have here is pretty good. You know, it's not perfect. It's not amazing. But it is a solid start for something that is LLM generated. You'll see here the dashboard. You'll see us being able to add a goal. You'll see a nice, clean user interface that's very simple. It works. The date picker is nice. It works as expected. And then it does add the goals to our vision board.

So that's pretty cool. Here's some of the code just to show you what that looks like. And everything here was LLM constructed. I made no updates to any code anywhere. Looking at the code, it looks pretty solid. Clean code. Well-written. Using, you know, good tailwind syntax and so on.

In general, if I was to rate the app creation process from start to finish, like, what did it take to install everything? What did it take to add all the files, paste the code in there, troubleshoot? It was actually a very painful process. When I first started this journey, I created a very simple version of the application. This one I would say was more complicated. It had more pages, more interactions, more capabilities. My very first version, all you could really do was modify the dummy code such as, like, for the goals. I had a dummy JSON file that I could go and modify, and then that would ultimately modify the application. But now this current application, as you saw, you could add, actually, a goal, and it updates the dashboard. So it is a better, stronger version of the app, but it was still a painful process.

I don't know if you all have used other types of models like Anthropx, Cloud, Cloud like SONA 3.7, for example. I'm really used to that one, and that is a fantastic coding experience. Not perfect as well. It does run into issues, but to me, that's a much smoother experience than this version of O3 Mini High, which is actually designed for coding. Literally when you choose it, it says good for logic and coding tasks, but I thought it was significantly worse than an experience with something like Anthropx, for example. But if you do want to see the deployed version of the application, you can go ahead and grab it right there.

So I asked the LLM to code me some end-to-end playwright tests, and this was the outcome here. I wrote five tests. Let's see. Should navigate from home to log in, that test failed. Should navigate from home to register, that test failed. Should add a goal and display it on the dashboard, that one also failed. Should log in and display the dashboard, that one passed. That's a very easy test to pass.

And should view a goal and delete it. Oh, that one is interesting. I'm actually surprised that that one passed, because we don't have delete goals capabilities. So I don't know where that came from and how that test was performed. Maybe we can look at the code later. But overall you see that three of the tests failed, two tests passed. One of them extremely easy. This one I am now questioning the validity of. And it's a poor outcome in my experience. It's a poor outcome from what I think.

Because if you think about it, the LLM had the context of the entire interaction that I had with it. It had me create the entire web application. It knew all of the challenges that I ran into. It knew everything about my project structure. So how did it not create passing automated tests? It's a very disappointing outcome in my opinion. And surprising as well because since the beginning, these LLMs have struggled to write automated tests. And it's funny because years later now, they continue to struggle to write automated tests. And I don't know, to me it's funny because also I guess it's like us humans don't like testing as well. Seems like the LLMs also don't enjoy it.

Here is an execution of the tests. Yes, so it ran through a bunch. But then it got stuck because we do have a 30 second timeout. It got stuck on one of the tests. And as you can see here, we don't have a delete goal capability in our dashboard. So I don't know how it passed that test. That would be an interesting code analysis afterwards. Let's go back and see some of the other tests. I should have ran it slower and one at a time. It did some things, which is cool. But yeah, then it gets stuck.

And then ultimately it fails. And then there is the report that is produced afterwards. Cool.

So that was testing. What about some continuous integration? It also failed here. I tried to deploy to Netlify. And as you'll see here, the build step ultimately ended up failing. I'm not sure why it failed. I didn't actually troubleshoot this because that wasn't the goal of this project was for me to troubleshoot every single error. But back in 2022, the chat GPT struggled with CICD as well. And it looks like it still continues to struggle. I was able to manually deploy the application linking Netlify to my GitHub account. And then Netlify was able to successfully do the deployment. But the LLM, with all of the knowledge and the context that it had, was not able to successfully deploy the app.

So in summary, overall, I would say that chat GPT was more capable this time than before, than back in 2022. So that's really nice. And certainly, there have been a lot of improvements in the LLM over that period. If you saw my previous table before, there was a number of red arrows that were pointing down, which means that it did really terrible. But in this case, it's a few arrows where it did well, and a few arrows where there's still some improvement to do. Like the playwright testing. In general, I see these LLMs have continued to struggle with, which is really interesting and really funny to me. But it is actually a very tough challenge to solve of how do you do testing well? Can you understand the rendered application well enough in order to be able to interact with it in the proper manner? Those are some of the challenges that, for example, we're dealing with at Vault Q. And so I can understand why the LLMs really struggle here as well. GitHub Action CI, still missing a number of details throughout in order to make that work. The working Next.js application, I gave it a yellow arrow because we were able to get an application going. As you saw, the application is not terrible. It's a working application. But it did take me some struggling through many prompts, many errors in order to get it resolved. But I was able to get it resolved. And it was an improvement over chatGPT 3.5, for example.

Installation of tools, majority of them worked. Some of them did not. The easy stuff, right, like text-based information, creative-based information, wireframes, user stories, requirements, easy for the LLM since the beginning, continues to do a good job. And the design in this case, I think it was improved. Previously, the design that chatGPT 3.5 had was pretty ugly and terrible. Now, the current version, I think, was usable. And if I had that kind of an application that looked like that, and I was able to use it, I would not be upset at it. It would be totally usable, and I would continue to use it on a regular basis.

One question that commonly comes up through these kinds of presentations, and I think it will continue to come up for a really long time, is, will AI replace developers? And it's a question that I personally also like to revisit as time goes on, as developments in AI continue to materialize forward, right, and we always have so many developments. I feel like AI keeps moving more and more rapidly, and there keep being more and more innovations throughout the industry. Previously, my answer was I didn't really see how AI was going to be replacing developers. Today, I think it's a mixed answer from what I'm seeing in the industry.

I'll give you an example for myself. This application that I built here, right, it's relatively simple, but it's a solid, maybe MVP, to try out a concept on. And I was able to create it with very little front end development knowledge. I'm no front end expert. I can get around, especially with an LLM, I can follow documentation and be pretty good, but I am no expert. I can't code something like this from scratch by myself without heavy assistance from, like, docs. I was able to create something like this. So that's kind of an enhancement of my skill, right? So now that's one engineer that is more enabled, capable of doing more than I was ever able to do before. So that expands my capabilities, it's not necessarily a replacement. And I think that's one area where AI will really help us to shine is it will make every individual more capable. For example, in software engineering, in automated software testing as well, that's exactly what we are seeing is AI is making everybody more capable to do more with less, right? So maybe instead of supporting one web application, you will now be able to support multiple web applications because you have more time in your day to do the important, complicated tasks.

However, I do think there's another side to this story where there is a replacement happening of engineers. So for example, building a startup, I'm learning a lot about that experience. And one of the most recent interviews that I saw with Y Combinator is that over 90 percent of the code written of the latest Y Combinator startups is generated by AI. And this code is generated by AI, by really smart engineers. These engineers can create their own applications from scratch. And yet they are using AI to generate over 90 percent of the code for their apps. And so what does that mean to me? Of course it makes you more efficient, but also it means that there is less necessity to bring in other engineers to help you with the process because code creation is now easier than ever.

And so maybe people in their current roles are not getting fired because of AI, but certainly new people are not getting hired. And that probably correlates with what we're seeing in the world today, right? We are seeing a lot of tech layoffs. We are seeing it a lot harder to get new positions in the industry. Could that be AI-related? Yeah, maybe. Of course, there are many other factors, but I think it possibly is.

And for example, for me, for myself, building these applications, running a business, I can see how AI has enhanced me so much that I find myself not needing to hire for other roles because AI can handle it for me.

So with all of that said, I really hope that you enjoyed this presentation. Again, I've been Nikolai Advolatkin. You can follow me on LinkedIn right there where I'll post the link to this presentation. You can keep up with me with all of my updates that are going on with my life if you want to keep up with me. That won't be on LinkedIn, that will be on my Instagram, you can find me there. Thank you so much for your time. It's been such a pleasure and I hope you have a beautiful day. Take care.