WEBVTT 00:00.018 --> 00:08.662 I've actually gone through some sort of an existential crisis, to be honest, where you kind of think what the hell is going on and where is this all headed, sort of thing. 00:09.142 --> 00:10.995 I think we are headed 00:11.591 --> 00:14.890 people need to start thinking product more. 00:14.890 --> 00:15.213 Yeah. 00:15.254 --> 00:17.736 product engineer thing is really kind of happening. 00:17.736 --> 00:18.250 Yeah. 00:18.433 --> 00:25.827 Guru prasad Venkatesh, Guru to his friends, joined the ThoughtWorks Bangalore office in 2006, where he continued to work for 12 years. 00:25.827 --> 00:29.961 Guru and I got to work together in Cincinnati, Ohio for Kroger, the big grocery chain. 00:29.961 --> 00:37.445 After ThoughtWorks and a brief stint at Cloudera, Guru went on to co-found Beskar Technologies, where he serves today as chief architect. 00:37.505 --> 00:44.725 In addition to their custom software development services for customers, Beskar is also building Baid an agentic AI dev tool, and 00:44.725 --> 00:47.771 Guru is our guest today on the Engineering with AI podcast. 01:04.809 --> 01:06.226 Guru, thanks so much for joining today. 01:06.226 --> 01:07.553 I really appreciate it. 01:08.149 --> 01:10.245 Hi Kyle, yeah thanks for having me here. 01:10.245 --> 01:14.213 It was really nice working with you as well at Kroger in Cincinnati. 01:14.400 --> 01:15.680 For sure, those were the good old days. 01:15.680 --> 01:20.180 I remember you endlessly running JMeter to fix performance problems that the team and I had made. 01:20.300 --> 01:21.721 So thank you. 01:21.721 --> 01:23.359 So let's kind of start with you. 01:23.364 --> 01:24.019 Yeah, 01:24.458 --> 01:28.299 Beskar there are two other partners that we started Beskar with. 01:28.299 --> 01:31.259 And one of them, Jake, is an ex-thought worker. 01:31.319 --> 01:34.999 In fact, he was the one who referred me to ThoughtWorks back in 2006. 01:36.319 --> 01:38.399 And interestingly, we are school buddies. 01:38.399 --> 01:45.219 I've known him from my third grade, third class, class three to class eight, after which he moved out our town. 01:45.219 --> 01:47.999 His dad got transferred to Gujarat. 01:48.361 --> 01:50.093 And we kept in touch after that. 01:50.093 --> 01:52.115 And he came back to Bangalore. 01:52.115 --> 01:56.179 And he had started this other company called Mavenhive once he left ThoughtWorks. 01:56.179 --> 01:57.601 And that got acquired. 01:57.601 --> 02:00.804 And then he wanted to leave that new company. 02:00.804 --> 02:06.934 And then I had left ThoughtWorks and joined CloudEra because I was very interested in data engineering and data science and that domain. 02:06.934 --> 02:10.344 But then after 02:10.344 --> 02:16.584 I think one and a half years at Cloudera I decided to move on and at the same point he was leaving his new company. 02:16.844 --> 02:25.364 So we both got together and decided that, yeah, I mean, he always was, tried to hire me into Mavenhive. 02:25.464 --> 02:28.864 And then at some point we decided, yeah, it's a good time to start. 02:28.864 --> 02:34.405 So him, Nitin is the other partner and me, we started Baskar. 02:34.405 --> 02:36.665 the core... 02:36.746 --> 02:49.686 philosophy, mean the ideas, the way we work, it's very ThoughtWorks influenced because a lot of, mean the folks that we worked with spent many years at ThoughtWorks. 02:49.686 --> 02:51.766 Yeah, so it's pretty, it's quite interesting. 02:51.766 --> 02:54.126 We've not grown a lot. 02:54.126 --> 03:00.666 We're just like 14, 15 strong after five years. 03:01.126 --> 03:03.046 Yeah, so that's the best. 03:03.185 --> 03:06.489 When you say ThoughtWorks like some of our audience will know what that means. 03:06.489 --> 03:14.055 For the rest, I think very extreme programming oriented, so test-driven development and continuous integration, continuous delivery. 03:14.055 --> 03:15.977 the other thing that comes to mind for me is, 03:15.977 --> 03:18.230 very different hierarchical or no hierarch. 03:18.230 --> 03:24.847 I remember telling everybody I'm sure I didn't have a boss for a lot of years, like where I didn't know who my boss was at least, and pretty sure I didn't have one. 03:24.847 --> 03:27.350 So I wonder like how about those parts? 03:27.350 --> 03:28.741 did that also come through? 03:28.741 --> 03:29.872 Yeah. 03:30.198 --> 03:30.668 Yeah, yeah. 03:30.668 --> 03:41.376 I mean, I think that's one of the things that I had a hard time dealing with after I left ThoughtWorks because it's so transparent and you don't actually report to anybody or ask 03:41.376 --> 03:42.247 anybody for... 03:42.247 --> 03:46.009 I mean, you share your leave plans beforehand, but apart from that, there's nothing. 03:46.009 --> 03:49.612 Yeah, mean, so Beskar is, yeah, it's similar. 03:49.612 --> 03:52.714 We three are partners, but we are the same. 03:52.714 --> 03:56.397 We kind of work on projects just like everybody else. 03:56.397 --> 03:58.198 We don different hats. 03:58.198 --> 04:08.385 in pursuing leads, like I might play the tech architect role when somebody is asking for an RFP for a new project, things like that. 04:08.385 --> 04:10.706 Jake might be involved initially. 04:10.706 --> 04:12.868 It actually totally depends. 04:12.868 --> 04:16.710 But it's pretty fluid and we've kind of grown organically as well. 04:16.710 --> 04:25.768 And the ways of working that's very thought works influenced in the sense that, like you said, it's very extreme programming. 04:25.768 --> 04:34.310 very visible short sprints like we try and show our work as often as possible yeah all that good stuff 04:34.310 --> 04:42.088 Well, I like to start with how you personally are using these tools and being the chief architect, like maybe you're coding every day still. 04:42.088 --> 04:46.750 You tell me, but I can tell you probably got some use of these tools. 04:46.750 --> 04:48.892 So what do you like? 04:48.892 --> 04:50.373 Maybe a little bit about like what 04:50.373 --> 04:57.600 languages you're working in, but more importantly, what Gen.AI tools are you using and what's working for you and what's not working for you? 04:57.600 --> 04:59.413 What do you like, personally? 04:59.436 --> 05:09.950 Yeah, so I've kind of dabbled with LLM's Gen.AI from GPT 3.5 days, like 2022 initially when it was released. 05:10.395 --> 05:12.684 at that point, we working with 05:12.684 --> 05:24.481 this company that was trying to solve interesting problems in the recruitment space, trying to improve an interviewer's process and make interviewing better And this LLM GPT 05:24.481 --> 05:34.425 3.5 fits so perfectly there because they were trying to summarize interview notes and do things like that that actually help in the interviewing process. 05:34.425 --> 05:37.356 And this is even when lang chain was not a thing. 05:37.356 --> 05:39.047 So we had to develop our own. 05:39.257 --> 05:42.530 AI client that now people call agents. 05:42.530 --> 05:48.476 So essentially that had to build that system and use a prompt and keep sending it to the LLM to give back answers. 05:48.476 --> 05:51.639 We had to interpret answers, build those chains and things like that. 05:51.639 --> 05:58.747 So yeah, mean, at that point, it was very interesting, but I didn't have a lot of confidence because there were so many hallucinations and... 05:58.841 --> 06:04.815 you ask it for a yes or no answer and it kind of always, that doesn't give you that consistently, right? 06:04.815 --> 06:06.176 But then it kept improving. 06:06.176 --> 06:07.837 That's what I've seen over the years. 06:07.837 --> 06:17.965 And one of the things that we used is the function spec, wherein you can say that I want the answer in this format and you can share a function spec with it. 06:17.965 --> 06:20.928 So that improved the quality of answers and things like that. 06:20.928 --> 06:23.191 personally, I've used ChatGPT right now. 06:23.191 --> 06:26.933 I use ChatGPT for a lot of the personal stuff. 06:27.011 --> 06:30.865 like managing health or basically asking it for diet. 06:30.865 --> 06:36.548 I go to the gym so it kind of plans my meal plans or workout what I should be doing and things like that. 06:36.548 --> 06:39.829 In fact, we were discussing some product ideas around that. 06:39.829 --> 06:42.351 But yeah, I use Claude code heavily. 06:42.351 --> 06:46.492 And Claude code is one of the best tools, think, out there. 06:46.492 --> 06:48.993 Played around with WindSurf a little bit. 06:50.459 --> 06:50.981 Yeah. 06:50.981 --> 06:52.741 with GitHub Copilot. 06:52.741 --> 06:54.721 I thought it was really, really good. 06:54.841 --> 06:59.701 Magical, basically the way it kind of generated your code based on context. 07:00.901 --> 07:09.501 But at that point, my thought process was that it is good for eliminating the need to write or type plumbing code. 07:10.801 --> 07:15.001 So based on the code in a file, can just auto-generate stuff. 07:15.171 --> 07:16.991 But it kept getting better and better. 07:16.991 --> 07:24.411 with Claude Code, in Beskar, I've been the most skeptical in using any of these tools. 07:24.851 --> 07:30.471 And I've always been played the devil's advocate when people have conversations around this saying, about this? 07:30.671 --> 07:33.071 It made this mistake and things like that. 07:33.071 --> 07:36.011 And I still am quite skeptical. 07:36.262 --> 07:38.662 But yeah, Claude Code I think, is very powerful. 07:38.662 --> 07:40.902 And that's what we use on a day-to-day basis. 07:41.332 --> 07:41.625 Yeah. 07:41.625 --> 07:51.945 And also one other really powerful tool is Kiro I don't know if you've kind of heard of it, it is primarily a spec-driven development tool. 07:52.965 --> 07:57.585 It is like, yeah, yeah, yeah. 07:57.585 --> 08:02.649 So it's basically based on Visual Studio Code and... 08:02.649 --> 08:05.069 they've introduced this workflow. 08:05.149 --> 08:12.649 So they are kind of having an opinion on how you should go about using AI agent in code coding assistance. 08:12.909 --> 08:20.369 So it starts with writing a specification and then that gets converted to your user stories and then tasks. 08:20.369 --> 08:27.349 And they have this sort of a markdown template where you can go and click on a task and say, want to execute this. 08:27.549 --> 08:31.149 So I found that to be really useful. 08:31.974 --> 08:36.803 And I think that's the way we are advocating the usage of Claude Code as well right now. 08:36.803 --> 08:41.277 So yeah, think Claude Code is the primary coding tool we're using. 08:41.277 --> 08:41.878 That's interesting. 08:41.878 --> 08:46.862 The first mention I'd heard of Kiro, we had Tim Cochran on the show recently. 08:46.862 --> 08:51.406 I think he's the latest article right now, though Chris Ford's will drop on Sunday. 08:51.406 --> 08:54.849 yeah, and he uses it because he's at Amazon, right? 08:54.849 --> 08:56.111 And they're the ones that make it. 08:56.111 --> 08:57.632 And I think there's a rule. 08:57.632 --> 08:58.673 I think they have to use it. 08:58.673 --> 08:59.863 I don't think they have an option. 08:59.863 --> 09:03.834 And so, but you're actually almost telling me more about how it works than he did. 09:03.834 --> 09:05.625 He just sort of assumed some knowledge. 09:05.625 --> 09:07.225 Anyhow, you're making me want to go try it. 09:07.225 --> 09:07.906 That's cool. 09:07.906 --> 09:16.330 do you use it somehow in conjunction with Claude Code or you, or it's just sort of like, nah, you, you like the spec driven in part, but it's, it's not as good at code. 09:16.330 --> 09:17.000 So you don't use it. 09:17.000 --> 09:19.871 Like how, help me tease out your thoughts there. 09:20.062 --> 09:25.506 Yeah, mostly it was an experiment with Kiro Just to was curious on that way of working. 09:25.506 --> 09:30.319 Because it sounded really interesting and in line with the way we do things usually. 09:30.319 --> 09:35.132 Like we don't kind of give a general prompt and say, hey, build this whole entire system. 09:35.132 --> 09:41.206 We kind of break it down and then do it a uh thin slice and in iterations. 09:41.206 --> 09:41.734 So. 09:41.734 --> 09:47.381 Yeah, I I've played around with Kiro but haven't used it heavily, to be honest. 09:47.442 --> 09:54.752 And one of the distinguished engineers in Amazon who I follow, he was kind of advocating Kiro. 09:54.752 --> 09:56.314 And that's how I got to know about it. 09:56.314 --> 09:59.517 And I was interested in learning about it. 09:59.517 --> 10:00.718 Cool, cool. 10:00.718 --> 10:03.260 I like how you talked about originally as plumbing code. 10:03.260 --> 10:09.824 I think one of my first ever interactions, it was actually ChatGPT that made me think, wow, this thing can write code. 10:09.824 --> 10:14.248 I was working on an aviation project and I wanted to deinterlace. 10:14.501 --> 10:15.902 radar signals. 10:15.902 --> 10:20.685 I had, it looked like I had an airplane that was jumping all over the screen and it didn't make any sense. 10:20.685 --> 10:24.207 But what was really happening is it was a time signal from 10 a.m. 10:24.207 --> 10:25.909 and another signal from 11 a.m. 10:25.909 --> 10:30.092 that had been like merged as though it was a time zone problem, right? 10:30.092 --> 10:31.343 And so the 10 a.m. 10:31.343 --> 10:32.283 flight and the 11 a.m. 10:32.283 --> 10:37.058 flight were like being merged like a zipper together and that's why it didn't make any sense. 10:37.058 --> 10:40.944 So I went and asked Chat GPT like is there math that can help me do this? 10:40.944 --> 10:44.307 And it was like, here's the history of radar and how this has worked. 10:44.307 --> 10:49.962 I'm like, oh, and then at the end, here is like 25 lines of Python and they worked. 10:49.962 --> 10:56.217 And that was like one of my first introductions to like, yeah, this thing, this is more than just, you know, vibe coding. 10:56.217 --> 10:57.449 There's something really here. 10:57.449 --> 11:07.375 Yeah, but like you say, the plumbing code, taking the drudgery away, I think that's an early like win, I think, that gets a lot of people interested. 11:07.744 --> 11:15.596 Yeah, but it sounded like you're doing more than just plumbing code now, but you also said your approach is normally to break things down into smaller pieces. 11:15.596 --> 11:21.211 Can you talk more about what sorts of guardrails you do use when you're approaching things? 11:21.211 --> 11:26.084 Yeah, so I don't write a lot of code on a daily basis now. 11:26.084 --> 11:32.970 I'm mostly involved with client, maintaining client relationships and architecture and things like that. 11:32.970 --> 11:40.596 So I use Claude in the plan mode mostly just initially to come up with ADRs. 11:40.596 --> 11:46.410 Basically, you use it as a bouncing board just to figure out what the architecture options are. 11:46.454 --> 11:49.995 trade-offs of you going down one path, uh things like that. 11:49.995 --> 11:51.367 And it's very good. 11:51.367 --> 11:57.552 It's very powerful because it's yeah, it's like talking to another senior developer and bouncing off ideas, right? 11:57.552 --> 12:00.233 So uh that's very good. 12:00.233 --> 12:12.800 But yeah, when I have done development, I start out with plan mode and try and, like, if I had to do a big refactoring, right, where I wanna migrate. 12:12.800 --> 12:18.494 Like I want to improve the quality of code in certain parts of the code base. 12:18.494 --> 12:24.038 I try and break it down and basically ask Claude Code in the plan mode to do it. 12:24.318 --> 12:31.823 And then it comes up with a plan and then I kind of review it and then go and see what can be improved there. 12:31.824 --> 12:33.525 And then do it step by step. 12:33.525 --> 12:37.962 think the idea is to avoid making big bang changes as much as possible. 12:37.962 --> 12:38.849 Yeah. 12:38.862 --> 12:46.167 keep things a little more controlled because it can actually go berserk sometimes and it also sometimes complicates things. 12:46.167 --> 12:54.871 Like you have to point out that the problem is already solved in a script somewhere and you can just reuse that rather than write another new script. 12:55.692 --> 12:56.872 Things like that. 12:57.233 --> 13:03.587 I have fallen in this trap where you kind of give it access to everything like AWS CloudWatch logs. 13:03.587 --> 13:05.468 go and look at your CI failures and everything. 13:05.468 --> 13:09.242 And then you keep saying, yes, yes, yes to our whatever it does and then become lazy. 13:09.242 --> 13:17.168 It actually does a decent job most of the times, but it can get into a loop and the solution usually is very simple. 13:17.168 --> 13:20.350 Like it would have been using the wrong binary on your machine. 13:20.350 --> 13:24.592 Like that's built for Intel rather than for, and then it'll crack its head. 13:25.233 --> 13:27.955 And then I'm like, yeah. 13:29.582 --> 13:31.223 It just keeps working hard. 13:31.223 --> 13:33.895 There's a lot of trial in the way it does things sometimes. 13:33.895 --> 13:38.777 And I think when you step in at that point, you have to nudge it. 13:40.208 --> 13:45.571 whatever we have thought of and ideas that we come up with, we see that Claude or Anthropic. 13:45.571 --> 13:48.172 They implement that in a week or so. 13:48.172 --> 13:52.583 So every week, they come up with these features that are kind of really... 13:52.583 --> 14:02.668 very powerful, like for example, the skills feature, or you can have your own agents, or the tools you can use, the memory tool especially, things like that. 14:02.583 --> 14:15.423 recently we've started to do trainings with companies where we kind of train on how to use coding agents like Claude code effectively. 14:15.674 --> 14:26.770 One of the main things there is to do this breakdown of tasks into small tasks and then use the plan mode, use skills, use memory as much as possible and try and be efficient 14:26.770 --> 14:34.886 about the usage of context basically the idea is that you have this sort of structure in the way you approach. 14:34.886 --> 14:44.591 you're utilizing Claude Code and have certain things in place, like a skill set, basically workflows for solving various problems, very, very different hats, like for example, a QA 14:44.591 --> 14:49.282 hat or a Cloud Architect hat or a security SecOps hat. 14:49.442 --> 14:56.244 And then look at your statement, like basically problem statement in these different angles, right? 14:56.244 --> 14:58.345 Yeah, so some of that. 14:58.410 --> 15:04.989 like saying, Claude, act as a quality assurance person in that kind of style where you're giving it like an agent's description. 15:06.011 --> 15:08.577 Yeah, got it, got it, got it, for sure. 15:08.577 --> 15:18.557 that this particular folder has some changes that are related to authentication or RBAC or things like that. 15:18.557 --> 15:21.217 And we want to be doubly sure that we're not breaking anything. 15:21.477 --> 15:31.597 So essentially, you ask Claude to execute a story that is maybe related to introducing some sort of token-based authentication mechanism. 15:31.671 --> 15:40.708 And then you use one of these agents, sub-subagents, that can use a particular skill set to verify that change and ensure that nothing is broken. 15:40.708 --> 15:47.092 I mean, of course, that can also be done with tests, but just use the non-deterministic nature of things. 15:47.092 --> 15:49.694 You ask it to verify as well. 15:49.722 --> 15:53.067 Let's start thinking about also the teams that you serve, right? 15:53.067 --> 16:00.696 So not just your own work, but teams at Beskar, maybe customer teams that you can see when you're working on custom software development. 16:00.696 --> 16:03.511 What kinds of changes are you seeing in? 16:03.511 --> 16:11.795 And maybe if you want to think about this in terms of what a mature team is doing versus what teams that are new are doing, whatever stories you want to share are interesting. 16:11.795 --> 16:15.237 But how are you seeing this impact teams in a big way? 16:15.237 --> 16:19.950 Yeah, I think in general, right now the expectations are high. 16:19.950 --> 16:26.923 essentially, people, customers expect you to be very productive, given that you use these powerful tools. 16:27.184 --> 16:28.344 That's one change. 16:28.344 --> 16:35.568 yeah, I mean, just in terms of the usage of tools themselves, customers have come to us asking for help. 16:35.810 --> 16:39.243 in figuring out how to use these tools effectively, like I was saying. 16:39.243 --> 16:48.731 So essentially, how to avoid just vibe coding and just build something up very quickly that quickly becomes unmaintainable. 16:48.731 --> 16:51.644 That you have so much code and that it's hard to debug. 16:51.644 --> 16:57.200 You just put that in production and then it becomes difficult to debug and maintain in production. 16:57.200 --> 17:00.083 So those are the kinds of problems that customers come to us with. 17:00.083 --> 17:04.948 What changes are you seeing in product management and in story breakdown and these kinds of things? 17:04.948 --> 17:05.567 Oh yeah 17:05.542 --> 17:05.923 absolutely. 17:05.923 --> 17:07.826 mean, Claude can help with that as well. 17:07.826 --> 17:18.682 So essentially, you think of an inception in ThoughtWorks where you kind of get gathered at a client site and brainstorm and come up with these ideas and the output of that whole 17:18.682 --> 17:21.314 Declan inception is a master story list, right? 17:21.314 --> 17:26.489 Where you have like a very refined set of stories that your team can get started with. 17:26.489 --> 17:28.373 I think Claude Code and 17:28.373 --> 17:34.726 ChatGPT and tools like that can really help in breaking things down, given a set of criteria. 17:34.726 --> 17:45.602 Like we did this for one of our clients where we said, this is a product requirements document, and I want you to break it down into features where you capture the story 17:45.602 --> 17:51.794 description, the volatility of the requirement, any questions, a high level estimate based on what you've done. 17:51.794 --> 17:52.614 and things like that. 17:52.614 --> 17:56.567 And it's really powerful, and it can actually do a really good job there as well. 17:56.567 --> 18:00.140 And so generally, think you can be very fast. 18:00.140 --> 18:06.783 I mean, you can move really fast when it comes to understanding, like breaking things down and things like that. 18:06.783 --> 18:09.683 ah 18:09.683 --> 18:10.904 fast reader. 18:10.904 --> 18:21.309 It's funny when you're, when you're talking about this, I'm thinking about an inception Kroger in fact And a previous, sort of waterfall process had generated like an 80 page, 18:21.309 --> 18:26.021 you know, product requirements document, you know, business analysis, you know, package. 18:26.021 --> 18:31.744 like job number one was reading that thing, you know, and it wasn't necessarily internally consistent with itself. 18:31.744 --> 18:32.815 Like what, what it was saying. 18:32.815 --> 18:38.257 And, and, know, we did, we, we sat down and plowed through it and tried to find, okay, this is, this is. 18:38.257 --> 18:45.164 where it says this way that way and says it that way over here and like which one is right and Claude would have done a very quick job of that. 18:45.164 --> 18:53.962 it would have done that in minutes you know and it it took us hours so but when you when you say like breaking a PRD down into stories you guys probably wrote the PRD in the first 18:53.962 --> 18:56.674 place I guess or tell me how that goes. 18:56.871 --> 19:04.146 Yeah, it was written manually by somebody from the client side, a product owner from the client side. 19:04.386 --> 19:14.464 But the interesting thing is also that with persistent memory with Claude a lot of this can be captured as specifications that Claude will actually remember. 19:14.464 --> 19:17.677 It can be checked in as part of your code base. 19:17.677 --> 19:22.380 And it can be used as, I mean, we used to acceptance tests that ensure that. 19:22.631 --> 19:23.588 like that. 19:24.067 --> 19:30.939 so the whole idea of breaking a PRD, which to begin with is a document that people shouldn't be writing, think. 19:30.939 --> 19:35.092 That's not how you should be doing it. 19:35.853 --> 19:37.493 But it's what a lot of people do. 19:38.174 --> 19:47.458 So Claude can really help in breaking that down into not only stories, but stories with acceptance criteria, which can be converted into a bunch of acceptance tests. 19:47.458 --> 19:48.288 Or your 19:48.378 --> 19:57.020 specifications which you adhere to right like and these can these can change over time and you can go and edit them and yeah 19:49.056 --> 19:57.692 Yeah. 19:57.692 --> 19:58.373 Yeah, I love that. 19:58.373 --> 19:59.915 Especially the acceptance test part. 19:59.915 --> 20:00.376 That's neat. 20:00.376 --> 20:01.068 I hadn't thought of that. 20:01.068 --> 20:03.561 So my next question is usually architecture and planning. 20:03.561 --> 20:08.658 And you've talked about that a little bit, using Claude to generate ADRs and things like that. 20:08.658 --> 20:10.580 uh Anything else coming to mind? 20:11.153 --> 20:12.175 No, think so. 20:12.175 --> 20:14.151 We had to choose between 20:15.122 --> 20:15.693 systems. 20:15.693 --> 20:19.560 Like basically, we looked at 20:20.232 --> 20:25.687 kind of forget the names but back in Kroger we used Shiro if you remember 20:25.687 --> 20:28.019 yeah, yeah, yeah. 20:28.499 --> 20:31.922 Are you thinking of like Auth0 and OKTA and things like that? 20:31.922 --> 20:32.802 Got it. 20:32.823 --> 20:34.544 Yeah, yeah, key cloak. 20:35.445 --> 20:36.138 Yep. 20:36.138 --> 20:44.334 loaded a bunch of these options and yeah, I mean, as a technical architect, you kind of think of these parameters that you assist these options by, right? 20:44.334 --> 20:54.019 Like how many users can you, mean, how many concurrent requests can you make or what is the feature set or how stable is it and what is the odd option? 20:54.327 --> 21:00.492 generally and what is the support like for these open source solutions or things like that. 21:00.492 --> 21:09.719 So as an architect, you think of these things, but I think Claude, when I've used it, has been very helpful in uncovering other aspects that you would have missed. 21:09.719 --> 21:14.022 um And also it can do research for you. 21:14.022 --> 21:20.667 So basically it pointed me to a benchmarking blog, which was kind of authoritative, which kind of said, OK, this is how 21:20.995 --> 21:28.335 this tool performs with this kind of framework performance and this load, you know, like under these conditions with this hardware. 21:28.515 --> 21:31.675 So just from a research point of view as well, it's very useful. 21:32.315 --> 21:45.656 Oh, actually, I forgot to mention perplexity, which is kind of what we use for research I it's heavily, I think even Gemini does that for now, but it is citation based, right? 21:45.656 --> 21:49.680 Like any question you ask, they kind of ensure that they... 21:49.748 --> 21:54.194 cite very authoritative sources of information in coming up with an answer. 21:54.194 --> 21:59.181 I think uh even perplexity is something that we can use for coming up with good ADRs. 21:59.181 --> 22:02.223 ah 22:02.223 --> 22:03.403 I'll have to look at that. 22:03.403 --> 22:04.323 That's cool. 22:04.323 --> 22:07.823 I'd heard the name, but nobody had actually described to me what it is. 22:07.823 --> 22:09.663 So yeah, you're the first. 22:09.683 --> 22:10.323 Fantastic. 22:10.323 --> 22:11.103 Okay, cool. 22:11.103 --> 22:12.583 So perplexity for research. 22:12.624 --> 22:15.684 obviously software development itself, it changes. 22:16.424 --> 22:24.444 But what comes to mind for you when you think about how teams operate and how does software development change with these Gen.AI tools? 22:24.444 --> 22:25.144 Yeah. 22:25.396 --> 22:26.816 Software development change. 22:26.816 --> 22:29.418 I think it can go a lot faster. 22:29.418 --> 22:41.622 I think we have to be mindful and go slower maybe because basically the idea is that you can do so much in like a day that you can also incur a lot of tech debt. 22:41.622 --> 22:48.010 just in terms of software development, think like we were discussing before, you have to 22:48.010 --> 22:51.474 ensure that you have the right sort of guardrails in place. 22:51.474 --> 22:54.658 You are actually understanding what Claude is doing. 22:54.658 --> 23:03.584 And although Claude is writing code, you're still accountable for that code written and you're still accountable for that working in production environments, right? 23:03.684 --> 23:09.588 from a software development point of view, I think you need fewer people maybe to get the job done. 23:09.588 --> 23:12.918 ah 23:12.918 --> 23:24.536 pairing, I don't know if is relevant anymore because yeah, like basically Claude Code I found that it can be a very good pair when it comes to debugging issues in production or 23:24.536 --> 23:26.117 debugging issues anywhere. 23:26.238 --> 23:37.271 In fact, it's kind of so helpful and it can weed out issues that you would, it would take you like a day to figure out that it's some sort of a DNS issue somewhere, but it. 23:37.271 --> 23:38.472 can do that very, quickly. 23:38.472 --> 23:41.433 So that whole loop, that's very powerful. 23:41.573 --> 23:45.435 I mean, there are downsides to it, like we discussed, where it can go into its own rabbit hole. 23:45.435 --> 23:50.481 But ah I think, overall, your development process, can go a lot faster. 23:50.481 --> 23:58.995 it's still, I mean, I'm still to speak to somebody who's done this on a project for like six to eight months and put something in production and seen it work. 23:58.995 --> 24:08.215 with a lot of users and I don't know, I it just seems like everybody is jumping on this bandwagon and people want to start using it. 24:08.555 --> 24:22.836 I think a lot of the learnings over the last 10 or 15 years that we apply from an engineering practices, engineering principles standpoint still applies, right? 24:22.836 --> 24:25.416 Like we still need to... 24:25.449 --> 24:28.663 we still need to learn to design systems well. 24:28.663 --> 24:30.604 We need to think of edge cases. 24:30.604 --> 24:34.151 We need to be mindful of how much we building, things like that. 24:34.151 --> 24:42.788 But yeah, in the right hands, uh Claude Code or a coding tool like this can be very, very powerful. 24:42.788 --> 24:42.970 Yep. 24:42.970 --> 24:43.962 Yeah. 24:43.962 --> 24:45.198 No, for sure. 24:45.198 --> 24:45.809 Totally agree. 24:45.809 --> 24:49.092 interesting, I normally ask about pair programming and mob coding too. 24:49.092 --> 24:50.733 And yeah, I'm so torn, right? 24:50.733 --> 24:54.048 Like I really came to love pair programming, working at ThoughtWorks. 24:54.048 --> 24:58.412 Like we were pair programming six, seven hours a day, you And it felt so powerful at the time. 24:58.412 --> 25:02.586 But to your point, it doesn't feel as needed anymore. 25:02.586 --> 25:04.057 But doesn't that cause new problems? 25:04.057 --> 25:07.310 Like around like team alignment, team cohesiveness around the code base? 25:07.310 --> 25:10.091 Like, do you have any concerns there? 25:10.342 --> 25:13.985 uh I think some of those can be addressed if everybody is using. 25:13.985 --> 25:25.082 So one of the really good things with the previous way of working pre-AI was that, like if you remember in Kroger, all of us used the same set of tools, right? 25:25.082 --> 25:31.297 And you could switch to any developer's machine and you could simply start writing code without kind of... 25:31.297 --> 25:34.520 And we had a set of good uh guardrails in place. 25:34.520 --> 25:37.062 Like we had static type, static code checking. 25:37.174 --> 25:43.019 static uh code analysis, things like that, a right set of safety nets in place. 25:43.019 --> 25:44.241 And I think that still applies. 25:44.241 --> 25:55.374 So even when you're using AI, if everybody in the team is using the same set of instructions in the form of Claude MD file, or the set of instructions that you pass on to 25:55.374 --> 26:01.152 the LLM, which can be a skill set, which can be your memory, lot of that can be shared between team members. 26:01.152 --> 26:01.627 Yeah. 26:01.627 --> 26:04.429 and checked in as part of the code, right? 26:04.429 --> 26:10.335 So that way a lot of the context is shared and the coding agent becomes an active team member. 26:10.335 --> 26:16.897 ah So I think that does solve part of the problem that you're talking 26:16.897 --> 26:17.548 I think you're right. 26:17.548 --> 26:24.089 I think some of that, like keeping the code cohesive is maybe not. 26:24.211 --> 26:26.822 It may be our responsibility, but not our job anymore. 26:26.822 --> 26:28.663 We're accountable, not responsible, something. 26:28.663 --> 26:33.425 But I also like what you're getting to there around like checking things in. 26:33.425 --> 26:35.803 That's been a question that's bugged me a few times. 26:35.803 --> 26:37.411 It's like, should we check everything in? 26:37.411 --> 26:39.930 Like from my experience, I'm normally working solo. 26:39.930 --> 26:43.372 So I can just check everything in and not think about it, not worry about it, you know? 26:43.372 --> 26:50.365 But on a team where you might have multiple developers going, does it make sense for, okay, so yes, check in your Claude.MD I agree. 26:50.365 --> 26:52.337 To me that sounds like totally right. 26:52.337 --> 26:54.940 Check in skills, yes, right? 26:54.940 --> 27:02.327 Wait, does it make sense for there to be guru skills and team skills and maybe Beskar wide skills? 27:02.327 --> 27:03.229 Like, I don't know. 27:03.229 --> 27:06.391 Is there anything you shouldn't check in that can you think of? 27:06.643 --> 27:13.853 Yeah, I mean, you can have personalized skills, I think, override what you're checking in. 27:13.995 --> 27:15.321 Yeah, yeah. 27:15.321 --> 27:20.815 but can you does it make sense to or should we or does it make sense to go towards sharing everything? 27:20.815 --> 27:24.728 Like, can you think of a what would be outside of a share everything? 27:24.980 --> 27:27.131 Yeah, I can't think of anything that... 27:27.612 --> 27:32.208 I mean, basically, there's no data issue here, right? 27:32.208 --> 27:34.429 There's no data privacy issue here. 27:34.429 --> 27:37.902 It's just a set of guidelines and interacting with the LLM. 27:37.902 --> 27:43.946 And it's good for a team to share those guidelines so that you're... 27:44.187 --> 27:44.688 so that... 27:44.688 --> 27:51.254 it's a little bit deterministic, at least, the way Claude answers your questions, or, you know, like the way it goes about solving the problem. 27:51.311 --> 27:52.070 Yep. 27:52.070 --> 28:03.297 and with persistent memory where it can retain that context across coding sessions that kind of becomes important because there's uniformity there and every developer works their 28:03.297 --> 28:12.353 own way and they can have their own set of skills or their own style uh but from a yeah yeah 28:12.353 --> 28:13.014 or something. 28:13.014 --> 28:19.458 So maybe, but nothing, so far I have not really seen anything where it's like, no, never check in your plan files. 28:19.458 --> 28:21.169 Like no, check in the plan files. 28:21.169 --> 28:24.114 of course, people, yeah, you want to be able to see them. 28:24.114 --> 28:25.243 Okay, cool. 28:25.243 --> 28:34.449 I mean, of course, should be avoiding checking in files with sensitive content in it, some sort of a key, API key or password or whatever. 28:34.449 --> 28:37.726 As long as you're avoiding that, I don't see any. 28:37.726 --> 28:42.958 file filled with social security numbers, like yeah, PII, things like that, no, for sure. 28:42.958 --> 28:45.869 that would have applied before, there's nothing new about that. 28:45.869 --> 28:48.210 yeah, okay, cool. 28:48.415 --> 28:50.886 You talked a little bit about testing earlier. 28:50.886 --> 28:57.579 You talked a little bit, I love how you talked about like a PDR can become an acceptance test suite, for instance. 28:57.579 --> 28:58.580 Really, really cool. 28:58.580 --> 29:01.061 Are you seeing testing change in any other way? 29:01.313 --> 29:05.036 I think testing like development has become pretty easy. 29:05.036 --> 29:15.264 uh It's become very easy to write integration tests because Claude can do a lot of the heavy lifting where you spin up a bunch of microservices using Docker compose, get that 29:15.264 --> 29:17.656 running in GitLab CI or wherever. 29:17.656 --> 29:23.048 It's become so easy to do that right now because if you would spend a of effort would go in. 29:23.124 --> 29:27.045 to getting an end-to-end integration test to work, but that has changed. 29:27.286 --> 29:39.563 And from a unit testing point of view, think we discussed this sometime back within Beskar but basically the idea should be to have Claude generate your tests and be very thorough 29:39.563 --> 29:46.476 with it, review them, and essentially ensure that they're testing the right set of things. 29:46.629 --> 29:59.016 there is this whole argument of classist versus mockist sort of testing and essentially ensuring that you're not actually testing you should be testing your the values and 29:59.016 --> 30:00.237 function returns 30:00.661 --> 30:08.081 rather than going and testing that the function is calling another, making another function call with an object within it, sort of a thing. 30:08.101 --> 30:12.061 basically, good tests are important. 30:12.521 --> 30:14.681 So spend time reviewing that. 30:14.681 --> 30:21.041 And then ask Claude to pass those tests with, do the red-green refactor. 30:21.041 --> 30:23.241 Make that a skill, sort of thing. 30:23.241 --> 30:28.861 So the whole red-green refactor implementation could be a skill, a workflow, where 30:28.935 --> 30:39.914 you it generates a set of tests for you you review those tests and then you invoke a skill to kind of say okay hey you have this sort of test go ahead and write code to pass them. 30:40.194 --> 30:41.005 interesting. 30:41.005 --> 30:42.265 Okay, cool. 30:42.265 --> 30:42.575 Cool. 30:42.575 --> 30:43.557 I hadn't thought of that. 30:43.557 --> 30:46.047 So, okay, but let's see, let's get beyond the test suite. 30:46.047 --> 30:51.399 Let's get beyond like integration testing and unit testing and sort of the literal QA, right? 30:51.399 --> 31:01.267 And so for instance, personally, I'm still desk checking work, you know, like when, especially if it's like an experience improvement. 31:01.267 --> 31:02.327 Like I want to look at it. 31:02.327 --> 31:07.260 I don't want to just tell Claude, like make the experience better in the following ways and just cross my fingers. 31:07.260 --> 31:08.790 You know, I want to see it. 31:08.790 --> 31:12.722 want to, you know, on, some level that helps me feedback into product design. 31:12.722 --> 31:14.884 You know, that's where some of the creativity happens. 31:14.884 --> 31:26.491 so how are you seeing that change in terms of like how you interact with UI And how do you like the manual testing, the error guessing, you know, sort of the, the, the, the actual 31:26.491 --> 31:26.941 active quality. 31:26.941 --> 31:28.824 quality assurance, not just the test suite itself. 31:28.824 --> 31:30.205 That's an interesting question. 31:30.205 --> 31:34.928 I've not actually seen tools come across tools that can actually solve that problem. 31:34.928 --> 31:39.741 I think you still rely on Selenium or tools. 31:39.741 --> 31:40.872 What is the... 31:42.152 --> 31:43.013 Playwright, yeah. 31:43.013 --> 31:44.175 Playwright tools like that. 31:44.175 --> 31:50.079 But yeah, I think you can still use Claude or Codex or coding agents like that to... 31:50.277 --> 31:52.850 invoke those tests, but it's a good point. 31:52.850 --> 31:57.988 basically how would you, I don't know, like I'm not, they've not done that sort of thing. 31:57.988 --> 32:04.181 think mostly there are still automatable wire coding agents, I would say. 32:04.181 --> 32:04.562 Nice. 32:04.562 --> 32:06.183 like, I like you're sticking to automation. 32:06.183 --> 32:07.012 Like I like it. 32:07.012 --> 32:07.665 That's a good thing. 32:07.665 --> 32:08.166 Okay. 32:08.166 --> 32:09.948 What about so integration? 32:09.948 --> 32:17.555 What about the, and, on, and on some level, think integration here, I mean, obviously you continuous integration, continuous delivery, those kinds of things. 32:17.555 --> 32:21.709 But I think also just the literal sort of act of integrating multiple code bases. 32:21.709 --> 32:22.722 Like if you. 32:22.722 --> 32:31.683 suddenly need to start using an API service that's part of one of your enterprise customers that you didn't use before, you know, and now you need to integrate with it and 32:31.683 --> 32:32.374 you weren't before. 32:32.374 --> 32:40.034 Like, that sort of literal act of integration in as much as like continuous integration and builds and all that. 32:40.034 --> 32:41.935 Yeah, what are you seeing changed there? 32:42.187 --> 32:44.591 I think it's kind of remained the same. 32:44.591 --> 32:49.216 basically, mean, sorry, could you kind of repeat the question? 32:49.444 --> 32:51.456 it's an awkward question. 32:51.456 --> 32:58.587 keep trying to, I'm maybe trying to conflate multiple things, but I think of it this way, like Kroger Pharmacy, right? 32:58.587 --> 33:03.789 One of the things that they had to do was integrate with an existing enterprise database. 33:03.789 --> 33:09.081 You know, and they couldn't really change it, you know, and that was always one of the big risks for that project. 33:09.081 --> 33:13.093 You know, we kind of knew like, yeah, here's the PRD for this code that we have to write. 33:13.093 --> 33:19.156 We're going to write it, you know, and we're going to write, you know, the, backend and we're going to write the UI and they're going to talk to each other. 33:19.156 --> 33:28.020 So there's not a lot of integration risk because we're coding the backend and we're coding this, but we also needed to talk to this other thing over here that we could not control. 33:28.020 --> 33:33.724 And so I kind of mean that when I'm saying integration right now, you know, the, the act, the literal act of integrating was something 33:33.724 --> 33:41.862 and not just sort of integrating code that, you know, person A and person B on the same team wrote, but like integrating with stuff which is external to you. 33:41.862 --> 33:43.885 Like, has it changed any of that? 33:44.136 --> 33:45.776 I don't think there's a lot that has changed. 33:45.776 --> 33:48.156 You still rely on your contract testing. 33:49.036 --> 33:52.076 basically, I think it's become very easy to mock things out. 33:52.076 --> 34:02.136 Like if you had to, I just generate a mock that maybe meets this sort of, if you have like swagger docs, just generate mocks. 34:02.316 --> 34:12.716 And I think there are some third party services that support that, but building something like that in-house has become so easy that you can. 34:12.925 --> 34:17.262 I mean, in fact, that's true for a lot of other things as well, if you want to. 34:17.262 --> 34:17.623 Yeah. 34:17.623 --> 34:18.358 So I think. 34:18.358 --> 34:26.080 that just reading that swagger, it's gonna go from here's the swagger URL to here's a working client so fast. 34:26.080 --> 34:33.368 So yeah, think it does speed, just inherently, just its ability to read documentation speeds that up quite a bit. 34:33.368 --> 34:34.044 Yeah. 34:34.044 --> 34:34.993 yeah, absolutely. 34:34.993 --> 34:39.096 might not necessarily see that there is like a different paradigm on the other side of that. 34:39.096 --> 34:47.553 Like it might just sort of create a mapper and not very thoughtfully convert our object to their object or something like that and not really realize that, this, ugh. 34:47.553 --> 34:51.257 So I still think you need to like think through the domain a little bit with it. 34:51.257 --> 34:55.840 But the literal act of reading a swagger and building a client, like it can do that incredibly quickly. 34:55.840 --> 34:58.021 Yeah. 34:58.021 --> 34:58.922 Cool. 34:59.540 --> 35:03.001 What about when it comes to deployment and infrastructure? 35:03.001 --> 35:04.584 Are you seeing any changes there? 35:05.250 --> 35:14.313 We actually, I was heavily involved in getting infrastructure automated for one of the previous clients and Claude did a fantastic job. 35:14.333 --> 35:17.355 He used Terraform and Terragrunt mostly. 35:17.355 --> 35:23.348 I read somewhere that these coding agents are not really good at coming up with an initial cut. 35:23.348 --> 35:24.134 uh 35:24.134 --> 35:26.785 from a code quality standpoint that they're not great. 35:26.785 --> 35:33.048 And you have to tweak and tune and set it up properly for it to do a good job. 35:33.048 --> 35:37.090 for IAC, I saw that it was actually doing a really good job. 35:37.090 --> 35:44.354 Like basically it broke things down into modules, stacks, and that TeraGrunt configs for every environment. 35:44.414 --> 35:45.424 That was so neat. 35:45.424 --> 35:49.020 Like basically, I think it was very kind of. 35:49.020 --> 35:50.500 nice to work with. 35:50.540 --> 35:57.660 something that would have taken like a week or two weeks, you can kind of do it in two days. 35:57.800 --> 36:06.220 So yeah, I mean, it really works beautifully with infrastructure as code And it kind of makes it easy to test it as well. 36:06.580 --> 36:12.678 So one of my things, and I think this is mostly from Sharan his... 36:12.678 --> 36:15.841 Go to tool for any kind of build activity is make, Makefile right? 36:15.841 --> 36:25.199 So you kind of put everything in a Makefile and any kind of build or test or integration test, or run anything, you do it via make. 36:25.340 --> 36:34.428 And I asked, I kind of put that as a guardrail and so you can just say make TF plan or make TF plan this particular stack of services. 36:34.428 --> 36:37.571 And it doesn't really, I mean, it just goes ahead and does that. 36:37.633 --> 36:43.395 So yeah, mean, it's really crazy how efficient you can be in building some of these things. 36:43.395 --> 37:00.191 yeah, so it does understand terraform I mean, again, you can actually inform about various environments and which AWS profile to use for which environment and configs like that. 37:00.191 --> 37:01.557 You can put that in your Claude.md 37:01.557 --> 37:07.997 You just have to ensure that the right set of instructions are in place, then it's your buddy can actually do lot of things for you. 37:08.078 --> 37:09.540 nice, yeah, totally. 37:09.540 --> 37:14.004 I normally ask about debugging next, but you already talked about that, and I agree with you, it's incredibly good. 37:14.004 --> 37:16.010 Refactoring, we've already covered. 37:16.010 --> 37:17.421 uh dependency management. 37:17.421 --> 37:19.014 Are you seeing anything there? 37:19.014 --> 37:27.338 I have a team right now that's got a dependency on Quarkus, you know, and it's a little bit out of date and now it's time for them to go update it, you know, and they're going 37:27.338 --> 37:35.072 through the motions of making that safe, you know, and so that kind of dependency management maybe if you've taken a dependency on a framework and you now need to update 37:35.072 --> 37:35.242 it. 37:35.242 --> 37:41.735 But of course, we're also seeing like Shai Hulud and the recent worms kind of running through uh 37:41.735 --> 37:44.958 GenAI tooling and all of this. 37:44.958 --> 37:48.200 How are you thoughtful about dependency management these days? 37:48.233 --> 37:58.293 I mean, in the projects that I've worked on, we've not done anything special apart from using a skill just to ensure that you're on the latest version or, know, like basically 37:58.293 --> 38:05.614 ensuring that, I mean, actually with vulnerabilities, AWS itself does a good job, right? 38:05.614 --> 38:16.514 Like, because you get those alerts and they run these scans that tell you that you're using a package that is actually has this vulnerability. 38:16.653 --> 38:28.852 So yeah, think writing a workflow, I mean, you can do these hooks as well with Cloud, where you can say that, hey, uh this part of the code base is kind of sensitive. 38:29.253 --> 38:37.681 basically, if anything changes there as part of a commit, I want you to run this script just to ensure that the dependencies are in order and things like that. 38:37.681 --> 38:39.864 I can think of those kinds of ideas. 38:39.864 --> 38:41.406 But yeah, basically, 38:41.406 --> 38:48.703 Some of these things can be built into the Claude tool set just to ensure that you're on the latest possible version. 38:48.703 --> 38:49.023 Yep, 100%. 38:49.023 --> 38:53.328 Okay, so in you working 38:53.328 --> 38:59.661 Working at a company that focuses on custom software development for clients, think this, and you've kind of hinted at this earlier, right? 38:59.661 --> 39:02.294 You started to say like, you need fewer developers. 39:02.294 --> 39:08.757 So, but a question I often ask is, what are you noticing about team roles and team ratios, right? 39:08.757 --> 39:15.381 Like in the old days, we might've said like, six pairs of developers, you need two BAs and you need one QA, you know? 39:15.381 --> 39:19.383 that wouldn't be work for everything perfectly, but that was a decent, 39:19.383 --> 39:21.464 you know, get started template. 39:21.784 --> 39:24.905 What are you seeing these days for like ratios? 39:24.905 --> 39:33.649 Do we need more of a certain role and less of like fewer developers like you've said, but what other impacts are you seeing on teams and how they get built? 39:34.273 --> 39:48.014 Yeah, so I think more of like basically you need somebody who is good at design and who is kind of good at system design, architecture, somebody who's a senior developer and like a 39:48.014 --> 39:52.830 bunch of mid-level developers who can actually use Claude code effectively, right? 39:52.830 --> 39:55.251 Guided by the senior team member. 39:55.251 --> 39:56.170 So, I mean, 39:56.170 --> 40:07.734 Yeah, like basically I think for junior developers especially it according to me it looks like it's going to be a problem because Claude Code can act as one of your junior dev 40:07.734 --> 40:13.916 pairs where it guided in the right kind of way, can do a lot of things for you. 40:13.916 --> 40:20.820 I think that's where teams mostly will cut costs in kind of hiring freshers or 40:20.839 --> 40:21.301 Yeah. 40:21.301 --> 40:22.762 lot of experience. 40:22.762 --> 40:25.635 there are still things to figure out there. 40:25.635 --> 40:36.478 Like for anybody who wants to enter the IT industry right now, it's going to be very difficult because you have something that's solving a lot of the problems that they would. 40:36.478 --> 40:38.300 And people are being productive as well. 40:38.300 --> 40:39.948 So the entry barrier is high. 40:39.948 --> 40:42.690 Yeah, yeah, getting higher. 40:42.751 --> 40:44.774 Yep, yep, agreed, agreed. 40:44.774 --> 40:46.616 Okay, but what about other roles? 40:46.616 --> 40:50.429 earlier when I was asking about testing, leaned into we automate everything. 40:50.429 --> 40:54.103 So I'm not hearing a lot of like, we have more QAs because there's more software. 40:54.103 --> 40:55.484 Okay, I'm not hearing that. 40:55.484 --> 40:55.956 But. 40:55.956 --> 40:57.999 Are you hearing anything around like product management? 40:57.999 --> 41:01.204 Do you need more access to SMEs, Subsect Matter Experts? 41:01.204 --> 41:05.409 Do you need more access to product management to answer questions? 41:05.409 --> 41:09.576 Or does that mean that you need differently shaped developers that can sort of do that on their own? 41:09.576 --> 41:12.959 how is it impacting like the other roles on the team? 41:13.223 --> 41:23.199 Yeah, definitely from what I've seen, There is a need for developers to think more from the product angle because a lot of the implementation has been taken off their hands, 41:23.199 --> 41:23.439 right? 41:23.439 --> 41:33.327 So essentially, product managers and somebody who can guide or have a clear vision and roadmap for a product, that becomes really important. 41:33.327 --> 41:37.890 I think the term is product engineers, I really don't like actually. 41:37.890 --> 41:40.052 But I prefer software engineers. 41:41.193 --> 41:42.768 yeah, that's kind of beginning to. 41:42.768 --> 41:44.257 I mean, with this whole thing, 41:44.105 --> 41:53.049 I've actually gone through some sort of an existential crisis, to be honest, where you kind of think what the hell is going on and where is this all headed, sort of thing. 41:53.230 --> 41:59.391 I think we are headed where people need to start thinking product more. 41:59.391 --> 41:59.757 Yeah. 41:59.757 --> 42:01.987 product engineer thing is really kind of happening. 42:02.236 --> 42:02.966 Yeah. 42:03.217 --> 42:04.078 Yeah, a hundred percent. 42:04.078 --> 42:11.438 I keep telling people, I don't know if I'm right, but the thing that I keep telling people instinctually is talk to Claude about the domain, right? 42:11.438 --> 42:15.898 If this is quick service restaurants, talk to Claude about the hamburgers, right? 42:15.898 --> 42:17.738 Talk to Claude about the point of sale system. 42:17.738 --> 42:20.118 Talk to Claude about, talk to it about the domain. 42:20.118 --> 42:23.038 Don't talk to it about the curly braces and where to put them. 42:23.038 --> 42:24.458 It doesn't make any sense. 42:24.458 --> 42:29.678 The more it knows about the domain, the smarter it's going to be at writing software for you. 42:29.678 --> 42:31.672 So yeah, that totally 42:31.672 --> 42:35.634 lines up with what I'm seeing, what you're talking about, that we all, you want to be talking about it more. 42:35.634 --> 42:36.895 So that means you need to know it. 42:36.895 --> 42:37.916 You can't avoid it. 42:37.916 --> 42:48.362 You know, I feel like teams used to have sort of, you know, a developer, maybe two that were so focused on the code and not so much focused on the actual domain. 42:48.362 --> 42:52.084 And I feel like there's less room for that person right now. 42:52.084 --> 42:52.634 Yeah. 42:52.634 --> 42:52.975 Yeah. 42:52.975 --> 42:53.705 Yeah. 42:54.072 --> 42:54.966 Yeah. 42:54.966 --> 42:55.776 Okay. 42:56.649 --> 43:02.149 So I've kind of seen that lot of people are coming up with these Claude skills, right? 43:02.149 --> 43:07.249 Like where they have their own set of roles or that sort of thing. 43:07.249 --> 43:14.449 And I've seen one of those repositories where it was really interesting where a CTO was a skill. 43:14.789 --> 43:15.829 I mean, basically. 43:15.929 --> 43:18.649 So any feature that needs to be implemented, 43:18.795 --> 43:28.445 it has to kind of go through that and you can kind of see, you can vet with the CTO or CEO just to kind of see if it's in line with the product vision or not sort of a thing. 43:28.445 --> 43:38.023 So like you said, basically that can be made a part of your whole code base and you can kind of ensure that you're doing the right thing. 43:38.024 --> 43:38.618 Yeah. 43:38.618 --> 43:39.720 Yeah, you could. 43:39.720 --> 43:44.348 That could be a git hook or not a get hook or Claude hook, for instance, or git hook either. 43:44.348 --> 43:44.909 Yeah. 43:44.909 --> 43:45.930 Ooh, I like that. 43:45.930 --> 43:46.850 Yes. 43:47.031 --> 43:47.392 OK. 43:47.392 --> 43:52.758 Has it changed how you hire staff, train, mentor when you think about those aspects? 43:53.198 --> 43:59.602 Absolutely, I think it definitely has changed the way you onboard new folks to the team. 43:59.683 --> 44:09.853 Because now you can have all of this context with the powerful features that Claude Code have added, with memory, with skills and stuff like that. 44:10.074 --> 44:16.399 It's become very easy to ask questions about the project, the problem you're solving, as well as the code base. 44:16.639 --> 44:17.645 Like you saying. 44:17.645 --> 44:25.094 what are we trying to solve from a business model standpoint, and how is the code based solving the problem? 44:25.094 --> 44:28.320 And it's become very easy to ask Claude to teach you something. 44:28.320 --> 44:33.726 So you can say that, I'm new here, and just tell me how I can do this. 44:33.726 --> 44:37.934 So I've paired with an in, like a. 44:37.934 --> 44:47.587 funny side note where I've paired with ThoughtWorks folks where they kind of grab a new developer joining a project, they just grab the keyboard and they say that, just tell me 44:47.587 --> 44:48.718 what to do and I'll do everything. 44:48.718 --> 44:53.460 I want to get ramped up on the project and that's really annoying. 44:53.460 --> 44:54.960 like gone are those days. 44:54.960 --> 44:56.497 You don't have to do that. 44:56.497 --> 45:00.823 Like Claude can be your onboarding agent and can answer all the questions for you. 45:00.823 --> 45:05.704 think onboarding people has definitely changed. 45:05.954 --> 45:06.396 Cool. 45:06.396 --> 45:07.079 Okay. 45:07.079 --> 45:09.361 How are you enabling teams with these tools? 45:09.361 --> 45:11.682 you said you're pretty stable headcount wise. 45:11.682 --> 45:17.027 So maybe you're not, you're onboarding people to projects more than you're onboarding people to Beskar. 45:17.027 --> 45:20.948 did you go through a phase where you were kind of deciding what to do? 45:20.948 --> 45:28.803 You talked earlier about how we sat down and talked about it and we agreed that, you unit tests should be classist not mockist and things like this. 45:28.803 --> 45:30.884 What different processes do you have? 45:30.884 --> 45:34.629 an organization around how you decide how to use these tools? 45:34.629 --> 45:37.742 Do you enable people in order to get them to use the tools better? 45:37.742 --> 45:38.973 Those kinds of things. 45:39.589 --> 45:50.428 Yeah, mostly who are joins kind of goes through, used to go through this object bootcamp sort of thing where they learn about professional programming principles like this whole 45:50.428 --> 45:55.373 classist versus mockist or why immutability is so important, things like that. 45:55.373 --> 45:59.946 So you have sort of these really good engineering programming principles that you have to learn, right? 45:59.946 --> 46:03.929 And I think those are still relevant to some extent. 46:04.001 --> 46:08.102 Because you as an engineer, need to learn that, according to me. 46:08.102 --> 46:11.184 My personal take is that. 46:11.184 --> 46:17.027 anybody who's joining Beskar, right now we have a set of clear opinions on what tools to use. 46:17.027 --> 46:27.374 And when it comes to GitHub Copilot versus Claude Code way of working, how you should go about managing context, context windows, and things like that. 46:27.374 --> 46:30.038 how you should manage your cloud coding sessions. 46:30.038 --> 46:37.467 Basically the idea is that should be able to exit a cloud coding session and enter it or start a new session without a lot of fear. 46:37.467 --> 46:45.049 Basically it should not be like you have a huge conversation that you lose if you exit a session. 46:45.049 --> 46:52.573 your Claude Code setup should be such that a lot of that is already built in and it's able to pull stuff from memory pretty easily. 46:52.573 --> 46:58.417 So we have some opinions internally and I think Baid is like a result of that sort of thing. 46:59.158 --> 47:04.561 So yeah, basically that's kind of how we are doing it right now. 47:04.761 --> 47:12.143 Interestingly, we have not actually, I mean, we have a plan, but we've not hired anybody and we've not had the need to do that. 47:12.143 --> 47:13.634 But yeah, that's the plan. 47:13.634 --> 47:22.644 Essentially, go through an object boot camp and also this set of training sessions where you learn about opinions on some of these things and how to use them. 47:23.652 --> 47:24.043 yeah. 47:24.043 --> 47:25.605 Maybe tell us a little bit about Bait. 47:25.605 --> 47:26.936 We kind of skipped past that, right? 47:26.936 --> 47:32.404 You guys do a lot of custom software delivery, but you also have this product, this generative AI. 47:32.404 --> 47:38.627 Do you guys use it internally, or do you guys have customers using it, and where does it fit in the landscape? 47:38.804 --> 47:40.775 Yeah, actually, it's still evolving. 47:40.775 --> 47:44.439 We started a podcast series recently. 47:44.439 --> 47:57.660 And uh basically, uh we have a discussion every week on new stuff that people have found interesting in the AI space and things like that. 47:57.660 --> 48:00.936 And the way it is kind of a result of that, where we thought that 48:00.936 --> 48:04.927 Like OpenClaw, OpenClaw was a very kind of interesting development, right? 48:04.927 --> 48:11.489 So the idea of a Bait, like the Bait bot, we call it Bait bot, and that's one aspect of it. 48:11.489 --> 48:21.152 But the idea is to build this autonomous bot that you can chat with from WhatsApp or Slack, which can help a team member, which can act as a personal assistant to a team 48:21.152 --> 48:21.952 member. 48:22.252 --> 48:27.989 So there is a a lobby bot that helps you onboard. 48:27.989 --> 48:35.089 you know, to Bayport and then you can ask it to be your personal assistant. 48:35.089 --> 48:42.049 So you can kind of say that I'm a DevOps engineer, I want to kind of monitor that this particular service is behaving properly, just go ahead and do it. 48:42.049 --> 48:50.449 And you can chat with it from WhatsApp, can kind of say from Slack, you can kind of add it and say that I want to do this. 48:50.449 --> 48:54.309 So it's controlled actions with persistent memory. 48:54.677 --> 48:59.397 So essentially, have a bunch of these agents that specialize in things. 48:59.477 --> 49:02.537 And they can help you do day-to-day work. 49:02.537 --> 49:06.297 They can do things like that. 49:06.297 --> 49:14.377 the idea is to also try and see if we can build proxy team members. 49:14.377 --> 49:16.737 So let's say you're not able to attend a meeting. 49:17.329 --> 49:18.879 your proxy does that, right? 49:18.879 --> 49:24.891 And that proxy agent has all your context because it has all the conversation history. 49:25.111 --> 49:26.993 So it kind of knows you as a personality. 49:26.993 --> 49:32.985 essentially with OpenFlower, there's this concept of a soul for every bot, right? 49:32.985 --> 49:40.008 So we are kind of going with the same kind of approach where every soul in the team, basically every team member has a soul. 49:40.008 --> 49:43.657 And then you kind of can make, I mean, this is just a... 49:43.657 --> 49:49.002 theory that we want to start, but it can take notes for you and then action items can be passed on to you or things like that. 49:49.002 --> 49:59.481 So the bot can attend the meeting and I think Granola is a startup that's solving a lot of these problems around meetings where it can take notes and stuff like that. 49:59.481 --> 50:05.845 But our idea is to extend that and make it more interesting. 50:05.845 --> 50:09.869 like if you had a, if it's a meeting where you're doing a technical 50:10.241 --> 50:21.885 architecture review and you want to quickly ask, I mean you've been having these recurrent meetings and you want to kind of go back in time and figure out why you decided on a 50:21.885 --> 50:22.727 certain path. 50:22.727 --> 50:29.174 You can ask this bot for it and say hey babe by the way tell me why we chose to do this or whatever. 50:29.174 --> 50:30.681 So that's the general idea. 50:30.681 --> 50:36.226 to your point, if your entire job is going to meetings, then one of these meeting note taker AIs is very helpful. 50:36.226 --> 50:40.409 But if your job is composed of many things, then it's less helpful. 50:40.409 --> 50:44.443 So this kind of ties that together with anything else that you might have asked it to do. 50:44.443 --> 50:45.143 that's cool. 50:45.143 --> 50:47.675 OK, now I want to use Bait. 50:47.926 --> 50:49.849 Yeah, mean, it's still in beta. 50:49.849 --> 50:52.532 We're still trying to, still testing it out. 50:52.532 --> 50:54.896 But as soon as it's out, I'll you know. 50:54.896 --> 50:55.456 Yeah. 50:55.456 --> 50:59.531 No, definitely I'll appreciate your feedback on some of this. 50:59.531 --> 51:05.296 so we are, I definitely appreciate your time today and we are coming close to the end. 51:05.296 --> 51:12.251 There's three more questions I wanna ask and these are kind of getting more into personal and less about the generative AI. 51:12.251 --> 51:17.375 In fact, the first question I'm gonna ask is something you've had to change your mind on, right? 51:17.375 --> 51:19.296 Or something that was unintuitive. 51:19.483 --> 51:22.427 And so far, almost every guest has said, I was wrong about AI. 51:22.427 --> 51:24.320 I didn't think it would be as important. 51:24.320 --> 51:28.424 So, okay, that's fine, but try to think of something else also. 51:29.406 --> 51:35.093 So here's the first question, something you've had to change your mind on or that was unintuitive at first. 51:35.937 --> 51:37.737 I was going to answer the same. 51:37.989 --> 51:39.220 Yeah, I think. 51:39.471 --> 51:42.912 You mean from an AI standpoint or just generally? 51:42.912 --> 51:44.446 light perspective, yeah. 51:44.697 --> 51:48.631 think of I mean, general life perspective, don't think of anything. 51:48.631 --> 52:00.482 But yeah, I mean, from an ESF standpoint, like I said, I've always been a skeptic and I've always played the devil's advocate and I think Lord Code, I've definitely come around now. 52:00.482 --> 52:08.009 Yeah, it seems like and it's getting to a point where you like I've seen developers who don't even have intelligent idea or any other idea open. 52:08.089 --> 52:09.599 All they have is terminal. 52:09.599 --> 52:11.159 yeah, that's me. 52:11.959 --> 52:13.039 yep, that's me. 52:13.039 --> 52:13.659 Yeah, it's funny. 52:13.659 --> 52:17.999 At first I thought, oh, this is like when Stack Overflow showed up. 52:17.999 --> 52:20.259 Stack Overflow felt like cheating for a little while. 52:20.259 --> 52:22.039 It was like, wow, this is great. 52:22.899 --> 52:28.919 And before that it was like internet mailing lists felt like cheating, because you didn't have to open the manual. 52:29.170 --> 52:31.350 And at first I thought, oh, this is just like that. 52:31.350 --> 52:33.610 But I'm starting to think, no, it's much bigger. 52:33.610 --> 52:36.810 It's almost like before the internet, kind of bigger. 52:37.276 --> 52:40.417 So yeah, think, yeah, it's a good answer. 52:40.417 --> 52:40.848 All right. 52:40.848 --> 52:44.247 So next question, something you are most proud of. 52:44.499 --> 52:54.130 I think, yeah, mean, leaving the regular job sort of thing and being kind of brave enough to go and start something on your own, yeah, I think that is something that, can be proud 52:54.130 --> 52:54.890 of. 52:56.511 --> 53:03.776 Yeah, basically to, and to start with no clients and just to kind of quit one day, not having anything else. 53:04.029 --> 53:05.612 to have the courage to start something. 53:05.612 --> 53:08.417 Yeah, I think that's something that's interesting. 53:09.438 --> 53:11.554 what's something that's giving you joy right now? 53:11.805 --> 53:17.351 I've started gymming I have a good routine now and I'm really enjoying that. 53:17.351 --> 53:19.867 So it's both gym and running, actually. 53:19.867 --> 53:22.569 That's kind of giving me joy, apart from the work. 53:22.569 --> 53:23.349 Yeah, yeah, yeah. 53:23.349 --> 53:23.940 Yeah, of course. 53:23.940 --> 53:24.891 Of Yeah. 53:24.891 --> 53:26.872 We have to we have to do other things too. 53:26.872 --> 53:27.532 Awesome. 53:27.532 --> 53:29.574 Guru, thank you so much for being on the show today. 53:29.574 --> 53:30.373 Really appreciate you. 53:30.373 --> 53:31.535 This is great. 53:31.786 --> 53:34.026 Thanks, Kyle, fun being here 53:34.624 --> 53:34.794 Okay.