What It Looks Like to Give AI 80% of Your Project with Dan Cardinal, CTO at Beghou Consulting

(Dan Cardinal at 00:00:00) In my opinion, you can't do any production AI development without those types of controls in place at any kind of scale. I went in skeptical, came away really impressed by the rigor of thought they have put around to the operational aspects of it.

(Joel Beasley at 00:00:11) I see a lot of AIs not delivering results, and I contrast that with people I have interviews with.

(Dan Cardinal at 00:00:16) It got us very quickly to this 80-ish percent point.

(Intro Narrator at 00:00:19) We were able to finish and get it all the way into a production pipeline now. Today, we're talking to Dan Cardinal, CTO at Beghou, about how he's using tools like Blitzy to achieve 80% autonomous project completion at scale. You're listening to Joel Beasley, Modern CTO.

(Dan Cardinal at 00:00:40) Our company is a tech-enabled services firm, consulting firm. We've had proprietary technology as part of our services for really 30 years, and in the past 12 years, have really developed a number of sort of product-y solutions that we use to deliver our services or that our clients use directly for their operations. And my job is all about making sure those technology offerings are not only supporting those services, but sort of amplifying them as we grow, because we are growing quickly and really rely on technology to help us scale in a sane way in this business environment.

(Joel Beasley at 00:01:12) Is that the main sale platform? Is that the one you use?

(Dan Cardinal at 00:01:15) Yeah, it is. We rebranded in the second half of last year to ARC, Beghou ARC. So it's kind of a fancier name.

(Dan Cardinal at 00:01:21) Its original name was the Data Collection Framework before any clients used it. That was 13 years ago. So, anyways, it's currently called Beghou ARC, and that is the platform that we use Blitzy for. That's an application-building platform that really combines the capabilities of custom software, but without the fragility and difficult-to-maintain parts that custom software has. So we went through this process of looking at what our clients do and what the kind of software they wanted, and we built a platform that we at Beghou could configure to deliver these software applications to them more quickly, to meet their usually pretty custom needs.

(Joel Beasley at 00:02:01) And what's the scale of this? Are we talking two engineers, 50 engineers? What's your current org size?

(Dan Cardinal at 00:02:08) Yeah, we're relatively small. So our engineering team is around 35 to 40 people, so not tiny. The different code bases each, though, have active engineering teams of maybe five or six people on them. So relatively small teams. I really like that small groups of smart developers working on things, which is one of the reasons we chose to try to invest in AI, because we have a lot more work that we want to do. So we were excited to look at AI as an option to get more throughput out of the folks we have.

(Joel Beasley at 00:02:38) You're in the healthcare pharma space, right?

(Dan Cardinal at 00:02:41) Yep, we are.

(Joel Beasley at 00:02:42) In that space as a whole, are they pretty receptive to AI?

(Dan Cardinal at 00:02:48) It depends. A lot of their work involves, you know, in the most sensitive case, patient-level data and things like that that you just absolutely can't mess around with how that data is handled. And so that industry, like many industries that handle sensitive data, have been very cautious about AI adoption. And so as a service provider to that industry, we have to be very aware of that philosophical agreement that we have with our clients, as well as tons of contractual agreement that came out right away after AI started to take off where you can't put our data in AI, you can't use AI for our project work at all without telling us. And so that, with the data, it's been the biggest challenge.

(Dan Cardinal at 00:03:25) With the code development, that's less challenging because clients are still fine as long as we're developing our code in a compliant way that we were before as well, and that applies to AI as well. So we already—the nice part for us that helps us adopt AI is we already had a number of controls in place to help ensure quality, to help ensure policies were followed, and for human-in-the-loop code review. It was sort of required to maintain the level of compliance that our clients expected. And so at least we weren't starting from scratch because, in my opinion, you can't do any production AI development without those types of controls in place at any kind of scale, or else you'll just rapidly, you know, at the worst end of the spectrum, have some massive security problem. The much more likely end of the spectrum is you'll just have a big unmaintainable mess before you know it.

(Joel Beasley at 00:04:12) And I was really impressed this year with this company, Blitzy, that I had on the show. When I first met their founder, I was like, yeah, I don't think you got it. I don't think—he showed it to me. I was blown away. I was like, no way, this is so cool. And so that's how we got connected, right? You're one of the Blitzy customers. Is that right?

(Dan Cardinal at 00:04:29) We are, yes. Yes. Yeah.

(Joel Beasley at 00:04:31) And so you use them for this .NET project?

(Dan Cardinal at 00:04:34) We did, yes. Yes. And similar introduction. It went really well. So the introduction was similar. We got referred by someone else who had had some success already with Blitzy. I went in skeptical, right? Came away really impressed by the technology and really impressed by the rigor of thought they had put around to the operational aspects of it. Right? Once I came out of that call with them, I'm like, oh, I see better now how this could work for larger development tasks than I did before that call, and we're willing to give it a try. And I also saw the value in what they had done was going to take us a long time to do if we wanted to try to do it ourselves, and who knows if we'd get it right.

(Joel Beasley at 00:05:09) How did you work with your peers on this? If you're like, hey, we have this—it was a proprietary code base. Is that true? Okay. So we have this proprietary IP, this code base. How did that decision go for you and your executive team?

(Dan Cardinal at 00:05:21) Sure. There was a discussion about it, particularly with our governance, risk, and compliance head, to make sure that this was an appropriate thing to do. There's some aspects of doing this with a software-as-a-service provider like Blitzy that check actually a lot of the boxes, because it really literally only has connection to the code, right? And the code is sensitive and it's our IP, and we want to protect that, but at least it doesn't have access to our other systems and things like that when you get into the advanced security threats of using AI, which are problems we have to solve when we try to use it internally. So in some ways, it was an easier discussion than some of the internal discussions we've had on how we want to use AI for our software engineering.

(Joel Beasley at 00:06:01) Help fill in this gray area for me. Is working with Blitzy, is it you go into the with a specific project and you engage for that, or do you get an annual license? How does it work?

(Dan Cardinal at 00:06:12) Yeah. For us, it might vary for certain clients of theirs. For us, first of all, they were very thoughtful through the sales process of pre-identifying some projects and types of work that would work well. And then the licensing model is really around how much code we ask it to ingest and then how much code it produces for us, with a nice little counter on how we're tracking on that overtime, which is helpful.

(Joel Beasley at 00:06:33) Give me an example of this first project that you've done.

(Dan Cardinal at 00:06:38) Sure. So we did pick something that we thought would maximize our chances for success. For that, it was a migration of a code base from an older version of .NET to the latest version of .NET Framework. So it stood out to us because it's low-strategy. It's pretty clear what needs to be done. It's high-effort, and it's low-interest work for our developers to work on. It's just very tedious, right? And so it sort of checked a bunch of boxes that made it, in our minds, perfect for AI, because it minimized the chance it could misunderstand what it needed to do, and really get a big unit of work done that we were having trouble getting to because it sort of fell into that potential technical debt category. So I was excited. We tried this project with them, and it went great. It got us very quickly to this 80-ish percent point, and then—

(Intro Narrator at 00:07:24) We were able to finish and get it

(Dan Cardinal at 00:07:26) all the way into a production pipeline now.

(Joel Beasley at 00:07:28) What does it look like, that last 20%? Do they give you a checklist of things that it couldn't figure out or questions or to-dos, or what is the last 20% look like?

(Dan Cardinal at 00:07:37) Yeah, it's mostly the judgment calls, right? So it took care of tons of the mechanical work that is required for something like this, and then it came up with a very nice list of things that it thought we should apply our judgment to before doing anything, and we did. And then there were some nuance to things that it didn't catch or sort of made the wrong judgment call, but that was pretty minimal. It was really more about leaving those decision points to us, and it did provide us with that as part of its documented output of where we want to look at that.

(Joel Beasley at 00:08:07) That's pretty cool because these are the same conversations we'd be having anyway. I've done so many software projects in my life. You know what frustrates me is when people look at this new technology and they scoff at it unless it's perfect. And it doesn't need to be perfect. What it needs to be is valuable.

(Dan Cardinal at 00:08:25) Yeah, absolutely. It's all about value, and that cuts both ways with AI. You get the skeptics. Oh, it didn't work, so we're not going to use it. And then you get folks who use it way differently than would actually be adding value for them just because it's sort of fascinating and fun and all of that. But you're absolutely right. Once you break it down into the types of work it can do to add real value, this conversation's a lot easier about where you want to use it and invest in it.

(Joel Beasley at 00:08:50) Would you recommend Blitzy to other people?

(Dan Cardinal at 00:08:53) I would, yes. It went—I was optimistic it would go well. It went better than I expected, and I would definitely recommend it for someone looking to get into AI software development. Even if you're looking at other options, as are we in a whole technical landscape level of how we want to evolve our software development lifecycle, it's been super helpful to partner with a pro firm or whoever has a bunch of experience there and leverage their platform, which, as long as they're well positioned to stay ahead of the operationalization of AI software development, they'll continue to add value for us as well.

(Joel Beasley at 00:09:26) I saw that one of your projects was generation of documentation from the code base. And my background, software development, I've just built a new application. I hadn't written code in a while, and I built an application these past two weeks. And so I'm in it right now for after not having been in it for a while. And to me, it almost seems like the moment we got tools that could generate documentation is the moment we don't really need documentation anymore. Have you found that any truth in that?

(Dan Cardinal at 00:09:55) I would say yes. And there was an interesting phenomenon with that, which was kind of a—we hadn't really thought about generating documentation when we kicked this off with Blitzy. There's an interesting compounding effect of that, though, which is that once you have all of that documentation, if you have another AI use case that has anything to do with that first one, you've got a plethora of context to give to the next use case. So in our case, this configurable framework that I talked about that we built last year before even engagement with Blitzy, we built an assistant for our engineers to use to more quickly—an AI-driven assistant to help us configure that tool. Its ability to configure things totally revolves around its understanding of what the rules are, and the documentation we had then was pretty sparse. And so it works pretty well, but it has gaps in its knowledge. Now we've taken the Blitzy documentation and fed it into that other AI tool, and it's much better and is, I think, going to get much better as we continue to sort of refine it. So it's a yes-and for me. It's like, yes, you can sort of interact directly with the code. You don't need as much documentation, and that documentation is actually really valuable for other AI use cases as well or valuable for whatever RAG-based or similar chat conversational assistant you want for people to interact with instead of the documentation directly. Our services—our knowledge, right? People come to us because we know how to do things in a way that we have a lot of deep expertise. And we're so excited at the success we've had, and our plans for taking that knowledge and really kind of embedding it into the technology, because AI makes so much more of that possible now, to bring really, really cool new offerings to our clients. So it's a really fun time, but it makes us want to develop much faster than we had before.

(Joel Beasley at 00:11:36) Let's wrap this up on some leadership advice and insight. So I like to always ask guests for leadership advice, and I'm going to give you some constraints. Okay? It has to be something that someone else told you that you implemented. It worked very well, and you've kept it with you for a while now.

(Dan Cardinal at 00:11:57) Yeah. Great question. I think for me, it has been—someone, I've had variations of this told to me over the years, but really pointedly and relatively recently in the last five to 10 years, let's say—you're in the seat for a reason. Just make a choice and lead, right? So I'm—it's not in my nature to be overly confident, and I really would like to look at every sort of part of a decision, every aspect of a decision before I make one. And that's good. That's a good trait. But the thing that really stuck with me is, if you are leading people, you are there for a reason most likely, and don't let—don't have a lack of self—avoid self-confidence issues and self-doubt issues around that and just make decisions and go. And that was super impactful for me. I went from being a very well-liked, passive, productive, but kind of passive leader, to someone who's sort of affecting change, in my opinion, at least. That's the newfound confidence coming through. So that—that'll sound simple, but that finally clicked and was transformative for me. I'll quickly add that I've been through lots of different leadership trainings and coaching and read leadership books and things over the years. And some advice I got that also stuck with me: a lot of people don't find one of those things solves all their problems. In these different leadership models that you can listen for and kind of make your own. And there have been some things that have come out of that that have really been transformative for me as well as a leader.

(Joel Beasley at 00:13:41) The—when you made that shift, the first piece of advice you gave, when you made that shift about decision-making, what was the actual situation that you were experiencing that caused you to say, okay, I have to make this change?

(Dan Cardinal at 00:13:58) As, especially now at this level of senior leadership, it became clear that we needed to decide on some things. We can—even—AI is actually a rife ground for this too, right? Because they're all risky decisions right now with AI because it's so new. But even before that, there are decisions that we need to make as a company that I—that no one else is going to make. You're never going to just be like, oh, this is the perfect answer to this. You're just going to have to make a decision and go. And I'd say, well, who am I to decide? Are you kidding? You're our CTO.

(Dan Cardinal at 00:14:31) Oh, right. And, you know, other people are just people too, and everybody has this angst and you just have to move. So there have been lots of cases. The AI is a legit one. Like, deciding to go with Blitsy is, you know, there were nice parts about it that made it less risky to me, but it's a big, you know, spend a bunch of money on this brand new way of doing things, and there's lots of ways it could go wrong.

(Dan Cardinal at 00:14:53) But nevertheless, someone has to decide and you just have to sort of follow your gut at some point.

(Joel Beasley at 00:14:58) You followed your gut and it worked out.

(Dan Cardinal at 00:14:59) And it worked out. Yeah.

(Joel Beasley at 00:15:01) Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you'd like to hear discussed on the podcast, either add me on LinkedIn or send me an email [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.