0:00 We're usually brought in when the house is on fire, right? 0:02 I don't think it's an automated way how you migrate code bases. 0:05 It's little pieces of of value that I add and that's how I progress from one stage to another. 0:10 But that's intense. The right to these tests beforehand that there's risk mitigation. 0:14 The more coverage you have, more accurate your parity checking is, the less risk you have at the AI is a differentiator. 0:21 It empowers great software architects, software engineers that have a strong software architect system designed capability versus those two. Not to write great software, you still need to be a great engineer. 0:33 That didn't change. If you're interested in system design and modernizing and migrating software, this episode is for you. I'm joined by Dennis Fink, principal consultant over at Xavier, and I loved learning from his approach how he executes on the many ideas and 0:49 projects that he has. So enjoy. 0:55 I looked at your LinkedIn and already in 2024 you landed assignments as AI engineer. Was it all bringing AI to production or was it more also kind of AI in way of working? 1:06 It, it, it was more companies wanting to learn more about AI and, and the very beginnings like models were pretty primitive back then and there was not a lot of knowledge. 1:20 What you see nowadays is that companies are much more self-sufficient. They have a lot of knowledge internally about artificial intelligence, so now the assignments that you currently get are are more complex. 1:35 Gotcha. Yeah. 1:35 But it's still bringing AI and AI component to production. 1:39 Yeah, somewhere within a product, yeah. 1:41 Can you share some of the examples that you've brought to production? For example, knowledge bases are quite popular. So that's, that's the interaction with with the rec system and then the interface around how do you interact with its response times, those type 2:00 of things. But also like a Gentic data research. So you have a big query database, for example, with lots of data in it. 2:10 You want to get insights out of it. 2:12 How do you do that, and how do you do that responsibly? 2:14 Yeah, I know you get a lot of energy out of building things. 2:18 Can you share some of the examples of what you've built recently? Recently I've have a lot of side projects. One of them is Eurosso, which is a genetic platform where you can automate business processes, which is really fun. Another one is a very old game 2:39 that I played about 20 years back that I brought back to life. So I converted it from a binary and a bit of source code which was written in C and C++ to to Rust and then a graphical user face around it so you can play it in the web browser. Nice. 3:04 It's really fun. How? 3:05 How is your CC plus plus or or rust? 3:08 Rusty. Yeah, Rusty, very rusty, yeah. 3:12 I coded C and C++ maybe 30 years ago. 3:16 So that's that's a long time 3025 years ago, more or less. 3:23 And well, then you evil to other other languages, a a bit lower level languages, higher level languages and it it's the knowledge about this programme like it just disappears. 3:39 And that's not an issue with regards to bringing this piece of software back to life. No, like I can still read it. 3:46 And while you're working on such a project like the, the, the patterns are still familiar and you know the principles and you can write pieces of code, you can debug it. 4:01 So it's it's like learning to bike you, you don't forget it. 4:05 Gotcha. Yeah. 4:06 So patterns, principles, those are the knowledge you need to be able to do something like this. Or how do you see that? 4:13 Yeah, absolutely. Yeah. 4:15 Yeah, I think if you like you, you have the term fight coding, right? It is a, a, a bit of a, there's a bit of a smear on the term because it's, it implies being able to create applications without, without programming knowledge. Yes, you're able to do that, but 4:38 you're not able to exercise any control over it. 4:43 So if you, if you fight code something you have serial knowledge, it will just go the direction that the the model thinks is best. And you're not able to assess quality and you are not able to give the proper direction. 4:58 What do you assess in terms of quality, especially when we're talking about a piece of software that's written in a specific language and you want to modernise it? 5:06 Feel like a lot of companies or maybe even people always migrations and modernisations are always a topic. 5:12 But what fundamentals do you really need to have to be able to do this? So you need to be familiar with test driven development. For example, let me let me take the the game accept. All right, so I migrated old, very old game from from binary C++ to Rust. 5:30 But you want to ensure that the mechanics are identical. 5:35 So my approach has been to to write tests, but also to use the code of both systems to assess whether the what it reads and what its outputs, the actions that it does that they perform identical when executed. So if it reads something and it 5:57 parses something in one system, it needs to be identical to the other one, and that's how you prove parity. 6:04 And this is in an automated sense. 6:05 In an automated sense, yeah. So I I created tooling for that to to simulate. For example, in in this game system, you have certain actions that trigger other actions and then wrote a simulator like we, we triggered this action. 6:24 Does it have exactly the same consequences? 6:27 And that's how you how you proceed. 6:28 So that's kind of a brute force way, but that ensures that the end result is is identical even though the code base differs. 6:37 And these actions that you've specified in this way of working, are those actions upfront that you're saying these are the the set of actions we need to validate or are they also generated or randomised or how do they come into place? 6:50 No, they so they they're the the actual code and three points in the original game versus the the new version. 6:58 Culture interesting and looking at a perspective with regards to test driven development. If the old code base didn't have any tests, do you then work to mature that to a certain level before you go modernised? Yeah, absolutely. 7:11 Yeah, Yeah. Interesting. 7:12 Because a lot of organizations that I've looked at, people that I've spoken to when it comes to software that they want to move away from, maybe there's a license component, maybe there's just a knowledge gap there. Maybe they want to move to a 7:24 different platform or a specific language. 7:27 Typically those older code bases, they don't have that level of maturity. They don't have the test framework. That's one why they want to modernize in the 1st place. And sometimes it's even instead of a migration, a complete rebuild and lots of those 7:39 projects in the end fail because it's a huge effort and you need to really have buy in from the business side because in the end it's a very long term play. So moving away when the code base is not really mature, that's I think also part of the 7:55 complexity typically. Absolutely, Yeah. 7:57 It's so writing these tests beforehand that there's risk mitigation and the more the, the more coverage you have, the more accurate your your parity checking is, the the less risk you have at the end. Nice. 8:13 Yeah. So the parity checking that is mainly crucial, I guess. Yeah, you start with a code base, you make sure that maturity wise, test wise, you have a good safety net. Then you modernize and you modernize in an automated way, I'm assuming. 8:26 Does it matter which model or which methodology you use to modernize? I don't think the model matters much as long as it's a competent model like I. 8:39 I like to use open DI myself. Codex models, but you could use cloud. It doesn't. 8:46 It doesn't really matter and methodology does matter. 8:49 You need to have the the instructions in place the parity tooling and I think that begins before you migrate any actual code. Gotcha. 8:58 Yeah. How'd you come into this way of working? Is it just a lot of trial and error experimentation, or? Like I became an enthusiast the moment it, it's, it went from large language models to to agentic. And I realized you could 9:16 automate a lot of tasks with it. That's when I also got very interested in, in, in creating euro cell to, to solve, to automate business processes because I could see the potential of, of agentic. And so I became a enthusiast 9:37 early on and, and started experimenting and, and you see as time progresses, these, these models get more and more more powerful, especially with like the tool calling they can do and the accuracy on that. Yeah, it's interesting because ownership is always a big thing, right? 9:59 I put a lot of effort into my codebase historically with regards to my commit message structure, the way we design systems and thinking about software architecture and applying patterns and keeping things simple and evolvable for the future. And I feel like a lot of those 10:15 are kind of out of the water. And especially when we're talking about a modernization, if I go from codebase X to codebase Y in an automated kind of finger snap way, how much influence do I have on the structure? 10:25 Or should I care about what the code in the end looks like from your perspective? Yeah, I, I don't think, like you say, in an automated way. I don't think it's, it's in an automated way how you migrate code bases. 10:38 So how I typically approach that is how you define like user stories that are really scoped to a specific piece of value. 10:49 And that's also how I instruct the model. 10:53 It's, it's little pieces of, of value that I add. 10:57 And that's how I progress from one stage to another. 10:59 And that's, that's intense. Like for example, the the game server migration overall, including the the web interface and and an editor took about a month of, of work, not full time work, but but still a month of work of, of prompting. 11:16 It's rating how you would do it as a as an engineer as well. 11:21 You, you iterate on on the product and you add little improvements gradually. So I don't think that that differs much. Yeah. 11:29 Are engineers typically good at this because they have to step out of the hands on part and more orchestrate things with regards to, you already mentioned user stories and that's a lot of prompting, a lot of back and forth. 11:40 Typically, that's not really what we historically have spent most of our time on. Most of the time has really been on actually building the thing and being in meetings to discuss about the thing and getting a better understanding. 11:51 But that's one of the execution side of things, yeah. 11:54 Is that changing? I think that is changing and I think you will see that engineers that are like their philosophy is limited by, by execution will love to be able to like zoom out a bit and have a whole overview and being able 12:13 to orchestrate on that level. And you have engineers that are in their in their little bubble that that don't know how to do that. You started creating code very early on. Age wise, you were there very early and typically those are the people that really love the 12:29 hands on part of things as well. But since the hands on thing is evolving, are you still getting the same amount of fulfilment or how do you see that? I think it's accelerated, yeah. 12:39 Like I have, I have a mind that is always exploding with with a million different ideas. I have probably hundreds if not thousands of unfinished projects. 12:50 Oh yeah. Still, yeah, absolutely. 12:52 It's never going to end. Yeah. 12:53 It's, it's never ending and I always have have have different ideas that I don't have time for, for fleshing out. 13:02 And this type of tooling helps me finally finish something you get from from A to C quite rapidly. 13:13 And it allows me to to have that focus to to to finish it. 13:18 For me, I think I'm, I'm the opposite compared to you. 13:21 When I start something, I really want to have the conference confidence that I can do this and actually finish things. 13:27 So if I don't believe that I can finish something, I typically don't start many things because I feel like what's the point? 13:33 If I start and I don't finish, I have to finish or at least see it through to a point where I'm fulfilled. 13:40 What would you advise people that are like me or similar that have trouble starting because starting seems effortless from your side? How do you make starting easy? 13:48 Or does it just come automatically? 13:49 Well it it comes semi automatically so I've HDHD brain. So that means that I start a million things and unlikely to finish all of them. 14:01 But there's also the sense that I like the journey, like I like the the learning journey, the experience of creating and a project doesn't necessarily need to end with the finished products for me. So it's also just just having fun. Yeah. 14:19 Can we go through a scenario? You have an idea and you want to start something. What are some of the first few steps that you typically do with regards to starting? 14:29 First, I've I've also have a business background, so I do some research, but this already exists. 14:35 Is it marketable? Sometimes it it could also be just a fun project for myself, right? 14:41 Yeah, like the game migration. Like the game migration. 14:46 So typically it starts with a bit of research and then formulating a like a high level plan, like what do I want to achieve? What are my goals? 14:57 What is the end state? And then drill down from there into more details, steps. And usually along the way I get a better idea of, of what I want to build. 15:09 And sometimes that changes during the development process as well. Like I get an, a great insight like how this could work better. And then now it's really easy to pivot. So in the past, if you are working on the project and you've invested months 15:26 developing something, if you get a new insight that's requires you to rewrite a large part of your code base, that is very demultivating because you've invested already so much time. 15:38 So you have this, this sunk cost fallacy, which you also don't have now. So it's lowers the the thresholds to to pivot in that sense. 15:49 I've done since I'm prompting a lot, I've installed like speech to text software on my laptop and I just talk a lot more with regards to either brainstorming if I have to do something non code related or if I have to do something software related. 16:02 Still, I feel like prompting now and typing has become my bottleneck because of that interaction. 16:08 I want it faster and faster to get the feedback faster and to get up and running faster. Is that something that you've experienced as well? Partially, yeah, especially also like due due to the speed, like agentic coding is not, it's not instant. It's not like you prompted LLM. 16:25 You get an instant response and and and things take time. 16:33 In terms of prompting, it's it's more the the waiting for a task to be finished, to see, to evaluate. 16:42 Did it do what I want and then re prompting or moving on to the next bit And and it's that that wait, it is a bit of a waste. 16:51 You can mitigate it like doing parallel features, for example, that say that don't touch each other's code base or source files. But there's a, there's a bit of yeah, cutting loss. Is that to say? 17:09 Do you parallelise? So when you're working on one feature, let's talk about the modernisation example of the game, you define your user stories, that's what you mentioned. Do you then see, OK, if we do these in parallel, these are not really going to touch each other 17:22 or how do you approach that, that waiting period? 17:24 If I know for sure there's no overlap, but usually like for example in the game example, I have AI, have a game server, I have a database layer. Yeah, I have an editor and I have a web client and all of those four, they can, they can use some work. So if I paralyse, it's usually 17:47 on the spread between those projects. 17:49 Gotcha. Yeah, So that's more defined boundaries rather than very close to each other feature wise in the same domain. Yeah. 17:57 What do you now do is what do you nowadays do while waiting? 18:00 Because I feel like I am a little bit more waiting nowadays than actually executing. I read a lot based on the output or I go do something else but I am more in a waiting state I feel like. Yeah, yeah, usually. 18:13 So I follow a bit along what it's doing because it's also shares like it's, it's AI thought process. 18:22 And sometimes that leads to new questions or new insights. 18:27 And then I go look it up, make a bit of research, for example, format of specific Sprite library, for example, or thinking, oh, I maybe I want to support X&Y format. 18:42 And then I go look it up if that exists and, and how do you ingest that, for example. So it's it's reading up and if I have nothing else to do, I will get distracted with something else. I'd probably forget what I was doing anyway and then come back half an hour. 18:58 Oh yeah, I was doing that and looking at the output again. 19:02 Gotcha. Yeah, I did a poll recently for the people that are listening. It was what are your favorite topics on the podcast? And system design was the one that won. I think it's like 50%. 19:13 And then I had four options, so head and shoulders above the rest. System design for me is quite interesting. It's not something that I'm very good at because I've always come into projects where something is already live and now we're building on top of that. 19:26 So a lot of artefacts with regards to system design have already been chosen and unless we're bridging a gap with regards to a new feature that we have to do, there's not much thinking in advance. So my experience has not been 19:39 from zero to one that often, but I know throughout your experience you've worked and focused on cloud and also as AI engineer, you've brought a lot of stuff to production, not just in companies but also inside projects. 19:52 Do you think system design is still going to be one of the most core competencies for the people building or how do you view that? I do think it's, it's just going to remain important for sure. It's, it's, you need to have a strong foundation. I'm a I'm a strong believer that 20:11 you will see in the in the in the coming years. 20:16 That AI is a differentiator. It empowers great software architects, software engineers that have a strong software architect system design capability versus those who are not. I believe that the the quality of the of of what you create with with these models is 20:39 exponentially higher in great engineers versus those who are not. Well, first, not to speak of the new generation. I I worry for them. 20:50 Why do you worry for them specifically? 20:52 Because they will not be educating themselves in these principles anymore. Because it's, it's like the Fable, it's, it's take out and you don't have these learning experience anymore. Gotcha. 21:07 What would you advise them then? Because you're saying software architecture and design principles, those are going to be important since you are more orchestrating rather than executing. What would you advise the new generation that's now learning or studying and coming up 21:20 actually? Go go through the big yeah. 21:22 Yeah, do it manually. Is that what you mean or? 21:25 Do it manually, struggle, make mistakes, preferably get from the Internet so you have no resources to look it up and that's that's the way how you truly learn. 21:40 Yeah, that's difficult. I feel like because the the manual approach, it will no longer be what companies expect. 21:47 I feel like if they will hire juniors, they will expect them to operate on a meaner level because you give AI tooling to accelerate people to also accelerate their knowledge, but also their execution power. So then doing the manual route, 22:00 especially on company time, I think that's going to be challenging. It's going to be a hard bargain. 22:04 Yeah, right. Yeah, Yeah, it's. 22:06 Going to be a hard bargain. I think it's, I think great engineers are usually interested in in, in writing software at A at a younger age. And I think that is when you make the investment. And I think the true cost of of the accessibility, the availability of these models is 22:25 that new generations will have lack of understanding of how to build software. And looking at the industry, because I'm curious there how, how you see this a few years back you would be able to do a boot camp or at least this is 22:39 what was out on the market. Boot camp, six weeks, two to three months, education specifically for a core competency, meaning front end or back end. 22:49 And then you would be marketable. 22:51 That was the boot, the boot camp kind of marketing strategy. 22:54 And some people coming out of that boot camp, they really got kind of hooked and they found fulfilment in creating and executing and building. And then they probably landed a job because there was a big demand for software engineers, 23:06 for people that would be able to execute and create. 23:09 I feel like the market is kind of flipped upside down. 23:11 I don't know if it makes sense to go through a boot camp with regards to core competencies or skills with regards to front end and back end anymore kind of cookie cutter model as compared to taking that same time and educating yourself and using 23:25 tooling that is now out there to self educate. 23:28 How do you see that? I think that's a good point. 23:31 You had companies like Kodish, you're right, that's educated developers. I've also personally seen a lot of very good people come out of there. 23:39 Also people that's that drought, that's that were not able to get this fire started in in themselves. 23:49 And I don't know what's what. What is the difference between these people? Like what makes one person gets really enthusiastic and deep into into the matter and and others not? I don't know what what the difference is, but yeah, I don't know where we're going, but. 24:11 That's OK. But I, I do think, yeah, I'm back on track. So I do think that people nowadays will have way more opportunities to educate themselves. There is a lot more information available on the Internet and people can experiment really rapidly. It's really, really rapidly. 24:37 So that might ignite that fire in them. 24:40 Yeah, I feel like one of the things I really underestimated coming into this industry was the amount of collaboration that I do within a team. Not just within the business with regards to what are we building and why is that, But 24:53 also throughout the team. There were a lot of inter dynamics of working together and trusting each other and then letting things be with regards to what other people build and giving feedback and reviews. From my perspective, that is evolving. If I as individual have more 25:07 building power, more execution power, then maybe instead of a team of 6, we can have a team of two or three, and that is really shrinking. So then the focus I had on trying to fix collaboration, I feel like we'll be much more 25:21 focused on the business side of things, trying to understand what are we actually building and why is that? 25:27 I feel like that will be where software engineers are going to grow, not just going deeper on the technical side, but also on the business side, understanding outcomes more. 25:36 You already mentioned when you start a new project, one of the first few things you do is research from the business side and also seeing if it's marketable, which I thought was quite funny. And I also do that. 25:46 And it's because I think I like entrepreneurship. 25:49 I like building something also in terms of outcomes, but I don't, I, I've spoken to many software engineers that are not like that, that also like building for the sake of building. And then when they've built it, that's it basically. That was the fulfilment. 26:02 It wasn't the fulfilment that comes after. 26:03 For me, it's the fulfilment that comes after and those distinct parts, the building for the sake of building or the building for the part that comes after. I see those as very much 2 camps and 1:00 will be happier. I feel like with a gentic 26:16 tooling and the other one building for the sake of building, that is the part that is rapidly evolving. 26:22 So I don't think those people will experience the same fulfilment anymore. Which is, yeah, it's it's new. 26:29 It's never been this crazy in terms of evolution. 26:33 No, no, absolutely. I, I think you're right. 26:35 So that'll be a big differentiator as well. 26:39 But you don't know what happens when people get the headspace to think about these things. So when you have the, the if, if you're more an orchestrator rather than an executor, so your, your roles change. Actually imagine you're stuck in 26:53 the old roles, but you discover a genetic tooling. 26:57 You will get more headspace to also consider other, other things. Like I used to be someone that was deep into my coat as well. That that was my, my whole world at that at that time. Along the way, I, I learned the ability to, to zoom out and look at the bigger picture and I 27:20 found I found enjoyment in that as well. 27:25 But that's something you, you, you learn. 27:27 And I think that learning process is accelerating as well. 27:31 So that might also affect people who are not there now but might be in the future. Yeah, I'm wondering in terms of focus, because I always see myself as software engineer, no matter if I have the responsibility of a person that 27:45 is responsible for product. At my core, I feel like I'm engineer, but I also feel like there is a lot to focus on or be responsible for as software engineers in terms of knowledge. 27:57 And one of the things I didn't really look at was system design, which I think is is going to become more important. 28:02 So I should put my time and effort into that. 28:04 But system design as a whole, it's also how do I make my software actually run or what does it run on and what are the underlying components of that Infrastructure wise, I feel like infrastructure might be a good bet with regards to focus 28:17 because if building is accelerating, then yes, we will have more build pieces of software, but it still has to run somewhere. It still has to scale to a certain component, still has to be in the hands of users. 28:28 And I feel like infrastructure is kind of the the core competency there. I don't know how much innovation has been happening in terms of a gentic solutioning for that front or if people are still very much operating in the same way they did two or three years ago. 28:43 Do you know? I think it's like cloud infrastructure hasn't changed a lot in terms of core principles. 28:53 It's become more accessible so more people can easily deploy cloud solutions. But you're, you're right, it, it's, it is a core component, but it's also also just a small part of the solution, right? So there is a phase where you where you do the the infrastructure underneath your 29:15 replication, but that might be two weeks, a month and then then you have it running in terms of like developing infrastructure scope for example, a genetic models can help with that as well. Yeah, yeah, yeah. 29:31 I did not do any infrastructure as code for a long, long time. 29:34 And then I came into a project where that was part of the responsibility. And then using and leveraging A genetic tooling because it is code, it was quite powerful for me to just illustrate and say, well, these are the things we 29:46 need to change or what is out there in the 1st place. 29:49 It was very good actually at identifying those things infrastructure wise. Yeah. 29:54 Yeah, what what I like about the like infrastructure scout in general, it's a very deterministic role. 30:01 So if I have a template, if I push it to test or production or to acceptance, it will behave identically because it's a deterministic template. It's a declarative with AI coding, that's that's not the case. 30:18 It'll it'll write code and if you prompt the same thing again, it will write different code. So that's, but that, that also has its appeal because it forces you to design the process and the tooling to, to fall into surgical strains. 30:38 So you have predictable output. So for me, the, the allure of, of cloud engineering was to create deterministic roles because I believe that determinism in development processes makes engineers happy, especially DevOps engineers who might otherwise be cold awake at night. 31:02 And for for example for software migration or writing new software it is I I get fulfilment about shaping the constraints around it to produce stable code, high quality code and scalable code. Yeah. 31:21 For people that have then, as an example, a code base that they want to modernise, that they want to move away from, there's a lot of vendors that kind of hook you into a certain way of working with specific components, a specific code. 31:33 And if that is 20 years ago, you might still be building on top of that, which can be quite substantial when we're talking about absolute lines of code. It can be in the millions basically. Also with regards to business 31:45 logic and the complexity that is in there, you might have teams that are kind of maybe not even building features on top of it anymore, but just that are there as insurance to make sure it runs and they maintain it. Now there is an opportunity to modernise. And I feel like that 31:59 modernisation can be accelerated with tooling. 32:03 You already mentioned that we have to make this deterministic or close to deterministic in terms of quality and outcomes. 32:10 It can't just be a finger snap and then you have a new code base and what is the outcome of that can be non deterministic. 32:16 So you have no clue if you can migrate or modernize or shift safely. But do you think this approach can work regardless of the scale of the existing code base? 32:27 Give an infinite tokens, yes. That's the constraint. 32:31 That's the constraint, yeah. Interesting. 32:33 So it gets expensive really, really fast to to give you an impression, the amount of tokens I used for the the game server migration runs in the billions. OK, so that's that's a huge project, but it's also a big code base, right? 32:49 There were more than I would say about half a million of, of, of lines of, of code lines assembly C++ and some some old Flash clients like Flash is also from a from a previous era. 33:11 The Flash client actually ran into the millions lines, lines of code that varies from both. That's that's huge. 33:18 So that that takes a lot of processing power to to process, but you do it in chunks, you do it in little chunks. 33:28 I look at the specific set of functionality I want to migrate this, this piece of of functionality, this slide and then create a plan for it. So I have an executable plan with with check boxes you can take off and I work on that in conjunction with the with the parity tool. 33:47 And then when that piece is done, I archive that the documentation for later reference and also continuously do regression testing. So that's my new change break something that I introduced before. 34:03 Yeah. So the general my way it's it's not a really fast process. OK, So it's not like the snap and you have your new code code base, but I think it's the the the best way of working to ensure that you produce something of value. Yeah. 34:20 Interesting. The finger snaps way, that would scare me, right? Because if all of a sudden there's a new code base and there's no ownership, there's no people that are in between orchestrating the thing, but they still have to own it. That's just as like throwing 34:33 something over the fence basically. 34:34 I don't think that would work. But this way of working, I'm assuming the same principles with regards to a modernization strategy would apply, right? If I have a software piece of software code base I'm modernizing away from, I'm going to look at my modernisation strategy. 34:49 I might want to migrate specific features first or deprecate certain things. I might want to freeze new functionality that comes on top of that, or figure out a way that can build in parallel, although that has risk. 35:01 So you're going to do a lot of it's, I feel like it's a whole risk mitigation exercise. And that still needs to happen even though you have tools that accelerate and you might not have to build hands on yourself as much anymore. 35:14 That would be my same approach. Still figure out what the modernisation strategy should be, then approach it from that way. So. 35:20 Imagine that you're building a product with an with an API, right? And you build functionality on top of that. Or you deprecate certain stuff from that API. You would still have to think about the same principles. I need to deprecate that 35:36 endpoint or I need backwards compatibility and you need to safeguards that. That is actually built in. 35:44 So do the the actual checks, Do the API still do the same? 35:47 If I call the new API with an with old POST parameters, does it? Does it return exactly what I expect? And it it's easy if you have the, the, the codes of the the old API versus the new API to write a tool that automatically checks that like how these are 36:07 the possible input values and you can do assertions, right? 36:12 So if I post this combination of values, I expect this result. 36:18 And if not, there is an error in the API. 36:22 So you can instruct the agents to to actually leverage that. 36:26 Maybe even run simulations for combinations of both parameters and then say does it in each instance return what I expect and if not, revise and and fix the error? 36:43 Yeah, yeah. Interesting. 36:46 I really want to work with this in a production kind of sense and I'm now at a client where I have the opportunity to do so. 36:52 So this is going to be fun. This is the magnitude of things and the teams that are involved in the code base scope. 36:58 I wouldn't know where to start and we're probably going to run experiments, but I'd like your suggestion of the code base is not as mature in the current state. 37:06 You need to work on the maturity side of that first, so looking into the existing flows and creating tests around that, or at least scaffolding around that. 37:16 Absolutely. But also so not not only maturity in the sense of tests and and priority tooling, but also documentation which is often lacking. 37:26 Like the more the more the agent knows about the structure of the old project and the so so the structure and the the the the the layout the better. How much of this do you do manually or how much do you rely on kind of your your agent to 37:45 give you input into structure and formalize that is what I'm still wondering. So when when starting up, there's a lot of manual input because I have a particular idea about how these documents should be structured, what should be in 37:58 there, what is important and what is not important. 38:02 Whereas an agent does a does a wild guess based on inferring from the from the source code. But I have a more high level overview of what's what the scope base is and what all pain points will be. So I make sure that the most 38:20 important points are document as well, and then have an agent revise it once I have the base in place. 38:29 Nice. I want to talk about risks and pick your bane on that front. I've heard a narrative, I don't know if it was in person or online because sometimes those blur together, but that certain agents are not as good when it comes to specific programming languages. 38:43 You already mentioned CC Plus plus from the example of modernizing the the game code base Flash as well. 38:50 Is that something that you've experienced in in practice that for example Rust since it's a newer quote UN quote language, it's not as good as? I found, I found migrations to go to be much better than than Rust. 39:07 Like Rust. It's it would often lead to compile errors because it's Rust. 39:13 Yeah, that's Rust. Yeah, yeah, that's Rust. 39:15 Also, if I do it manually, yeah, yeah. 39:17 Yeah, exactly. So that's that takes a few iterations before it compiles in the 1st place. 39:24 Whereas Go I rarely have compile errors if I if I have an output Go code, I don't think it has a lot of knowledge about Flash in in general and the models that I use. 39:37 Also nobody uses Flash. So yeah, why? 39:39 Why makes sense, makes sense, but it could infer from most of the principles like so my the the Flash client I migrated to HTML JavaScript. OK, it's got the base right in one shot, but particularly implementation details it did not. And then I that that took quite 40:03 a few iterations to to get right. 40:05 Gotcha. And then when we're talking about risk, I feel like hallucination is still a thing. 40:12 It's getting better and better, but how do you build around or with hallucination in mind accommodate for that? 40:21 That's that's the test to me. That's all the scaffolding is basically. Is it just for the non deterministic part? I assume so, right? 40:27 Yeah, nice. The same feature parity sense I could still do regardless of any code base, right? 40:36 Because code base will have features. 40:38 You have a structure around figuring out if the old code base behaves the same as the new code base. 40:45 So I don't see feature parity working differently for different pieces of software. So how do you build on top of that? Because the existing code base for me feature wise has everything. 40:57 It's complete. That's what you're modernizing away from. But the new code base you're going to incrementally build up in terms of features. 41:04 Sometimes it's just a ball of mud and features kind of are very much interwoven. How do you accommodate for that when you're moving to a newer code base? 41:12 So the the advantage of the the approach I described is that you also you, you already have your tests written right in your in your old code base because you're maturing that you already have your parity tooling running properly. 41:25 That also means that in the new code base you will also have tests already written and you will have regression tooling that checks whether a change that you made, but it leads to to a regression bug for example. So that makes it really easy to 41:42 to iterate on it. And if you are going to continue with agentic coding on the new project as well, which I didn't assume is the advantages, you can can use the same pipeline and you can codify that into your documentation that that 42:01 needs to happen every time a build this room for example. 42:04 That newer way of working, if we have an existing code base, you can structure it in a way that agentic tools work quite well. 42:12 There's a narrative online on skills versus agents and how much documentation is too much documentation, how explicit you need to be versus relying on implicit model knowledge. 42:22 I feel like years ago you had to be very explicit, but nowadays there's a lot of implicit model knowledge. 42:30 The way of working that software engineers are going to go about is very much changing. I'm on the train of I feel like spectrum development is quite effective if I put my time and attention in building a spec. And also a spec is quite arbitrary. I've noticed people have 42:44 different interpretations of what it is to create a specification, but if I have a very solid spec that actually aligns with the feature that I'm trying to build, then the building of things is actually quite reliable. 42:55 If I have really good conventions, that is even better. And if I have the right scaffolding and the right kind of safety net in testing and resilience, that's the best, to be honest. 43:06 So I really see that as kind of at least right now how we're going to operate in the future. What is your view on that? 43:13 Yeah, I think that's a devil evolved, right? 43:17 So absolutely. I have a have a fun anecdote. 2 weeks ago I had a session with a colleague of mine who works for CBA and she had an idea for 10 years to create a it's it's for the the real estate market but she's not technical and so she 43:44 asked me and another colleague to to help her out with with AI to create this application for her. 43:51 What I did instead is show her agentic coding. 43:56 How does it work? I set up the the the base for her like this is how you structure your projects. 44:03 Here's how you create a plan and then the first iteration I instructed the the model to create a a plan from her. 44:16 She had a document with functional requirements, which was quite elaborate. And then the model created a implementation plan from those functional requirements and then started implementing them. And within the hour there was an 44:33 like a POC application that she could already showcase to a venture capitalist for for investment, for example. 44:44 But then I gave her the rates. I said, now you're going to do it. She prompted the next bit of functionality. And then I uttered the words, you're now a developer. And she, she lit up. 44:58 She, she realized she was now she had the efficacy to, to, to execute on, on her vision. So I think that's that's that's a great way for her to to to to start. 45:17 And now she understands she's, she's still not technical, but she knows how to structure her project, how to iterate on it, how to, how to create new versions of her application to make her vision come true. I don't know if that answered 45:35 your question, but it's something I had to think about when you said that made. Me think of another question because I know topics like these can be quite controversial for different audiences, but especially the people that have 45:47 gone through education with the software or have years of work experience under the belt, right? 45:53 If you would say she's now a developer, there are many people that would say actually she's not because she doesn't think about XY and Z. But I do agree that this build process, that's the part that gets a lot easier. 46:05 So I'm very much wondering then, when do software engineers or the people that currently are engineers come into play? 46:10 Is it when it reaches a certain level of scale? 46:13 But the people that are driving, they will never realize that, OK, this is when you actually need certain foundational knowledge or expertise. There was a LinkedIn article on someone that made an app and the app was actually, people had to 46:25 upload their pictures and it was kind of a social network app. 46:29 And it was completely orchestrated and built in the way that you exemplified. And there was a security breach and all the images got leaked and stuff like that. 46:38 So that will definitely be concerning. 46:40 I feel like I don't have an answer for that yet. 46:43 No, we we're usually brought in when the house is on fire, right? That's when you need real software. Damage control, Yeah, that's it. 46:49 That's the reality. Yeah, no, So I I took the phrase you're now developer from from Netlifi. 46:55 You're probably familiar with it. 46:56 That's, that's what you see the first time you deploy an application to, to Netlifi. You're now developer. 47:03 And the idea is to, to, to, to empower people to make them enthusiastic so they deep dive more. 47:10 Of course, to, to write great software, you still need to be a great engineer. It didn't change. 47:16 Yeah. Yeah, thanks so much for coming on, Dennis. This was a blast. 47:19 I love picking your brain on modernisations and just your general principles on building nowadays. 47:25 You're more than welcome. Cool. 47:26 We'll round it off here. If you're still listening, let me know in the comments section what you thought of this episode and we'll see you on the next one.
Community discussion