Scale Like a CEO

Balancing AI Ambitions and Realities with Johann Beukes | Scale Like a CEO

• Justin Reinert • Season 1 • Episode 9

Share episode

In this episode of Scale Like a CEO, host Justin welcomes Johann Beukes—a fractional CTO with over 15 years of experience in AI and data science. Johann shares his journey from software development to solving complex AI challenges for both startups and Fortune 500 companies. He discusses the current hype and misconceptions around AI, the importance of aligning AI projects with business objectives, and the unique approach of his company, Quest AI Solutions, in delivering practical, effective AI solutions. Johann also reveals the challenges of building a talented team and navigating the noisy landscape of AI expertise. Tune in to learn how Johann's practical insights and strategic thinking can help your business leverage AI for real-world impact.

Speaker 1:

People are either conservative because they're afraid to fail, or they're over-optimistic and think that AI can solve all problems. Finding that balance is going to be the biggest challenge.

Speaker 2:

Welcome to Scale Like a CEO. Today, we're joined by Johan Bux, a fractional CTO and partner in multiple technology ventures, who brings over 15 years of experience in data science and AI implementation. As businesses navigate the complex landscape of artificial intelligence, Johan's practical approach and real-world insights help leaders cut through the hype and deliver actual value. From working with startups to Fortune 500 companies, he's built a reputation for solving complex challenges and building effective teams that drive results. Join us as we explore the realities of AI adoption for solving complex challenges and building effective teams that drive results. Join us as we explore the realities of AI adoption and learn from Johan's unique perspective on scaling technology solutions.

Speaker 3:

Johan, thank you so much for joining me on Scale Like a CEO To get us started. If you wouldn't mind, just give us a 90-second intro to yourself and your business.

Speaker 1:

Johan Bierkus, I'm a fractional CTO for many different companies. I'm also a partner in two companies. One, I'm the CTO, where we do agentic type solutions, ai solutions, data science. In the other company, I'm a partner and my role there is to be the CTO for startups. The one is more startup focused, the other is more enterprise, fortune 500 focused, but both require AI and technology. So that's the commonality between the two.

Speaker 3:

So you know, if you look at the industries that you serve, what's one of these problems that you see and how are you solving it?

Speaker 1:

I think the biggest challenge right now is just the hype cycle that we're in. I've been in data science and big data and machine learning for probably about 15 years and prior to that, you know, did software development and I've seen the hype cycles. When I went to school, I did my master's in data science. One of the things they did is walk us through from the 60s all the way. You know how we got to today. And when I talk to CTOs, a lot of these guys are very well established in what they're doing, but they're afraid to take a step because they see all of the hype, but they don't want to be the ones that go to their bosses and say you know what? This failed, we tried it and we couldn't succeed. So people are either conservative because they're afraid to fail, or they're over optimistic and think that AI can solve all problems. Finding that balance is going to be the biggest challenge. The second biggest challenge is that there's a lack of understanding about AI projects and a lot of times people will compare it with software development projects. If you came to me, justin, and you said, johan, can you build me a website with a shopping cart and Stripe integration and all of that. It's pretty predictable what you're going to get. If I've done this before, I can also give you the estimated timeline and predictable deliverable. However, if you came to me and said, johan, I want you to build me a model that can predict with 99.999% accuracy, that's not the same.

Speaker 1:

Ai is probabilistic in nature and you have to always tie the business case to the solution. We have processes in place, but it boils down to instead of you telling me I need this kind of accuracy, we always focus on what's the business problem. So, for instance, if you say I need this kind of accuracy, we always focus on what's the business problem. So, for instance, if you say I need to shave off two hours out of an eight-hour workday, that is achievable. I might even achieve that with a model that is 50% accurate. It's always interesting when I start engaging with clients how they see AI solve their business problems and then having to educate and change their mindset of how the process helps them to get it into production.

Speaker 3:

Yeah, I mean there is a long fight around the possibility of AI, but getting it to practical solutions and making it work like we want it, so you can be challenging. I'm curious what makes you unique in the way that you work with your clients?

Speaker 1:

A lot of times, and I'll talk about Quest AI for a second. That's the AI solutions agency. The first question we get is how are you able to compete with the big consulting firms or the big agencies out there? It's a pretty easy answer, because we're not trying to be the biggest, we're trying to be the best, and for that to be true, we need a few people that can actually do the work, because not everybody can. So I'll give you a practical example. A couple of years ago and just this is an example that shows you the difference between how we operate and what makes us unique versus other agencies A client came to us and it's I'm in South Florida.

Speaker 1:

There's a lot of cruise lines down here, so I'm not going to name which one, but one of the big ones came to us and said they just got a quote from one of the big consulting firms. They need to have a pilot on one of their ships actually one of the biggest ships where they're counting people. And they need to count people because right now, the way they do it in the buffet area is they, every 15 minutes, go take a napkin with the knives and forks and count how many there were before and how many is gone, and that's how they keep track of how many people are in an area, which is terrible, it's unique, but that's not a good one. So they're like we want to use computer vision and we engage with this consulting company that's pretty well known and they gave us a solution. That's about $2 million. It requires us to implement new camera systems on the ship and it also requires us to install some hardware that can connect to the cloud, because they need cloud access to run some of the models, and it was just mind-blowing that anybody would propose that. Especially for a ship, the bandwidth, the internet access, is not the same as what we have at home, so you can't build that kind of infrastructure on a ship. We looked at the problem and found that they're not trying to identify people, they're just trying to count people, and that's a very big distinction. If I needed to identify and say you know, every photo that I see, I need to compare it with Justin to identify our faces, that is much more computation or needs a lot more computation than if all I wanted to do is say, count how many people is in this picture. It's two different problems and the way that we approached. It also required us not to install any more hardware.

Speaker 1:

For instance, if we use this, what we did is we ended up using the security cameras. The reason that the consulting firm didn't want to use it is because it's fish lens and I didn't feel familiar with like there's fish eye security cameras. The reason that the consulting firm didn't want to use it is because it's fish lens and I didn't feel familiar with like there's fish eye security cameras. So if you create a model, it means you have to create a custom computer vision model to recognize people. There's no off the shelf open source version, so we did that and we had about 6,000 images that we labeled and what that means is we basically take an image, give it to a human. A human will draw this rectangle area of interest around a person. We train an algorithm to recognize whenever you see that little square and it looks like this that is a person. That is how you train these models. The alternative is somebody else did that and you can, if it's open source, use that, and that's what this company wanted to do. But there weren't any models available at that time. There was open source with a fisheye lens human detector, so we built that Three months later about $250,000 in and it was much more than we usually charge.

Speaker 1:

But we had to get hardware. I'm going to say Lambda Labs is the company we used. We got one server, used the existing camera, so the ship doesn't have to be dry, docked to retrofit cameras, and we had to have people actually go on a cruise, which the employees were thrilled with. About three months later we had a model working, counting people and actually counting heads, not just people, and when I say heads, bold hats, all kinds of types of heads it would count. And with the fish lens model there were actually three different models one in the middle, one between the middle and the edges and one at the edge. So we had three data sets and three different models to count people and it was pretty accurate. So that's a long story to tell you.

Speaker 1:

The way we look at problems is very unique. Ai doesn't always solve the problem. Sometimes just a good if statement is good enough, sometimes a state machine, something that in computer science classes they teach kids about state machines, and it's not AI. When I went to school we had a professor who was teaching advanced statistics course and he's like he was bragging because he's been doing that for 20 years at the time and he was saying everybody's starting to say what I'm doing is machine learning and AI, but it's statistics. It's been around since the 60s.

Speaker 1:

That's part of the process is looking at what's the tools out there, be practical about it, select the right tool. And then the last thing I would say that makes us very distinct is all of our projects. Unless we've really done a lot of it, we approach with eight to 10 weeks off a prototype. The reason I do that is a lot of agencies will go in and say we're going to build the web app, the mobile app, integrate with your SAP system all as part of the proof of concept, but they haven't tested if AI is the right solution. So we always focus eight to 10 weeks on what is the core problem. Let's see if you have the right data enough data to solve your business problem. Only when we prove that out we'll find people that can build a website or integration or other things that cost money. Those things are predictable. The piece that's not predictable is the AI.

Speaker 3:

Yeah, that's so great. I love the cruise ship and you got to see your team on the cruise to do the work. I'm curious to hear a little bit more about your team and you know how you've gone about building a team within your business.

Speaker 1:

This is a very interesting business. I started Quest, the solutions company, with a friend of mine that I've used for 15 years as a recruiter. His background is recruiting and we started talking about the challenge of finding good people that's smart, that will stick to a job, and, unfortunately, the people that's really good at their job in AI and machine learning get bored very quickly. They don't want to do the same thing for the next five years. They just they want to move on. What's the next problem? You know those are the kids that solve puzzles and take things apart. You know it's just curious minds and that's totally fine. Those are the kind of people that thrive. They enjoy the type of work.

Speaker 1:

So we looked at our business model and we're like most companies. Most of our clients do not need full-time data scientists for every single business function. They might need one or two, and then when there's a big project, maybe a couple more people experts in machine learning or data science or building out engineering pipelines up and running, somebody that's got technical experience, like a DevOps or somebody that has done technical work. We can educate them, upskill or cross-skill them to maintain it. So what happens is we look at the business model and like all right, what if we work on projects? We focus on projects, we rebuild it out for a client and have that very first project as a reference implementation. Their team can monitor it. They work with us during that process. They also get you know the best practices. That goes along with that, because that's what we do for a living. We constantly keep up with what's out there. When we step away, they can either own it or continue. We'll help hire and place people or a lot of times this has happened too We'll bring in partners Capgemini or KPMG or all these big consulting firms that are good at throwing bodies at a problem, and sometimes that's what you need to scale it, but that initial work to actually solve it. It doesn't work if you just throw bodies at it. You need specialists. So when we work through that with our clients, we typically have some of the best people in the industry work on those projects and they stay on that until they solved it and then we move them to the next projects.

Speaker 1:

We're pretty picky in the type of projects that we take on, and it's two reasons. One is we want to solve the difficult problems that nobody else is really attacking, and then we also have the kind of team that thrive in solving difficult problems. So if I start giving them things that they know how to do or they think is something that is not really that difficult, they're going to get bored. So that is a challenge as somebody that needs to manage a company and client expectations and all of these different aspects. You have to keep the client and the employees and everybody happy but fun work. I mean, that's really what it boils down to is everybody that's on the team enjoys what they're doing. They don't mind working on it every day and they get that thrill of finishing it, getting to a point where, yes, we did something nobody else has done before.

Speaker 3:

It's great. It sounds like you've got a good formula for how you move people into the business and different projects. I'm curious what are some of the challenges that you faced as you've scaled the team?

Speaker 1:

The challenge has become where a lot of the interviews and I've changed how I do interviews these days A lot of our team members roll from one project to interviews, and I've changed how I do interviews. These days A lot of our team members roll from one project to the next. I've worked with some of these folks for years, but every now and then we don't have enough people we have to recruit. It's becoming more difficult with language models, you know chat, gpt and all these things. I've had people that's looking at me while they're doing the interview and I can tell they're searching for answers. So we've totally changed how we interview. I don't give them code challenges anymore. We walk through problems we look at.

Speaker 1:

I want to say there's probably four aspects that we really focus on. One would be critical thinking skills. I have a degree in cognitive psychology. I have a degree in data science. It's a really good I want to say combination to analyze people as well. I'm not a psychologist not pretending to be but I understand how people process information, think, store all those things.

Speaker 1:

So part of that interview process is to really see how they reason. What if they get stuck? Do they panic, do they look for a solution. We'll challenge them, give them problems that we don't know if there's an answer, just to see if they get frustrated and how they respond, because that is what's going to happen in the job, right? We also look at how well do they communicate, how clearly do they explain stuff? Are they able to articulate something to us? What if we ask them to write it down? These days, we almost I want to say all of our clients in the last month at least require developers or engineers that uses AI to help them code. Now, if you just use AI and you ask it to do your work and you don't know how to plan up front, how to communicate the product requirements, it's going to go crazy and it's going to cause more trouble than it's going to be worth. So that's the second thing is we look at how well they document, communicate and explain. We also look at how much they keep up with the industry.

Speaker 1:

Do they know that OpenAI has all these different models and how do they know which ones are good for what? One of the things and I'm giving away some of the interview questions here, but I would give them a question that requires and you can probably solve it with OpenAI's ChatGPT where it's doing some kind of prediction for you. Right, these language models? They're not prediction engines, they're not recommendation engines. What they do is they use tools that they outsource that specific problem to and they get back the answer.

Speaker 1:

If you created a solution for a client, let's say the client is asking us we have this list of work orders and we need you to categorize them and list them in a level of importance and look at the historical backlog to see which ones have a higher probability of getting fixed quickly, let's say those are the parameters. You can feed that into a language model and it's going to give you an answer. However, it's not going to be always the best model. What you want to do is you want to go through and see where the language model interprets the text that the person is typing or whatever input mechanism they use, and then have a tool that's got a traditional data science model. The one that I talked about before is from the 60s maybe, but it's a statistical sound model that gets called by the language model and it returns the answer and the language model gives you the answer. Architecting that understanding why that's important. It's just one of those things that very quickly in the interview. We go down those paths and we realize if somebody is an engineer that entered into AI, they'll probably trip up on that question. If somebody is a data scientist, economist or even a bioinformatica student, those folks typically have a really good background in using machine learning, using AI, to solve their specific problems and they typically don't get tripped up. So that whole process of interviewing, finding people has definitely changed for us in how we filter through and get those people. Now, that's on the technical side.

Speaker 1:

Your question was also what other challenges do we have? And on the business side, marketing and lead generation. I think that's one of the first areas that we. It's so much noise and it just feels like it's a waste of time to go there. I have agents. I have AI agents that curate information for me. If there's topics that we're interested in as a company and we want to write about, we have an agent that does research for us and sends us an email with the topics, summaries and a link.

Speaker 1:

If we're looking for candidates that's very hard to find and I'll give you an example in a minute We'll scan and see if there's any candidates looking for work that matches the criteria that we have. It's just using the tools you know to kind of find where can we automate the process you know and then, once that is done, have a human in the loop. So those would be the two challenges, because I think the biggest reason there's such a amount of noise is just the hype. I mean, everybody's talking about AI or they're an expert at AI somehow, so it's just getting through. That really is tough. So just a quick sidebar on the difficulty of finding positions right.

Speaker 1:

So we had a client come to us a couple of years ago and they wanted to create the next version of a shopping cart experience. They were looking at using AR and VR, augmented reality and virtual reality. And if you have an iPhone, there's a software kit on iPhones, the modern iPhones that we, as a developer, we can use to scan a room and create a 3D map of that room. Now if I was a retailer and I wanted to sell furniture, I can scan this room and put my furniture in the room. I can see it, not just like on Amazon, on a flat 2D picture. I can actually see and almost like in a game and look at it. So we needed to develop something similar to that. But if you think about the technical challenges. If you're looking around the room and you remove an object, now there's a hole the texture, the map, like color of the chair or the material that's gone so you have to fill that in. That's a really good use case for generative AI. Think about if you have to fill in the gap on a carpet or a tile or paint on the walls.

Speaker 1:

We needed somebody that could do that, but we also needed somebody that understands 3D environments, ar and VR. My partner and some of the other partners in the company because they have a background in recruiting hunted and found a person that worked at Meta. We offered them the project and he came on board. So that is some of the cases where it's really rewarding. I understand the technology side of it really well. That's my role, but then I have a team behind me. They can find people. You have to have a really good network to get in touch with the right people like that, yeah, it sounds can find people. You have to have a really good network to get in touch with the right people like that.

Speaker 3:

Yeah, it sounds like you're doing a really great job at cutting through the hype and doing some great work. I love that, johan, thank you so much for joining me on the podcast today. If people want to reach out to you, how can they get in touch?

Speaker 1:

The easiest is on LinkedIn. I will get the messages, but then also Johan Beekhuis. There's not a lot of us in the world. It's a very common name in South Africa, where I'm from, but johanbeekhuiscom is the landing page my companies that I have. There's too many, so it's just easier to go to one place and you'll find me, and LinkedIn, of course.