Great infrastructure isn’t just about uptime — it’s about trust, ownership, and making change feel safe. Kristjan Elias shares how Pipedrive balances platform reliability, cloud efficiency, and AI experimentation at scale.
In this episode of Ship Happens, host Per Krogslund sits down with Kristjan Elias, Director of Engineering Infrastructure at Pipedrive, for a conversation about what modern infrastructure teams actually own — and why that ownership matters.
Kristjan reflects on nearly a decade at Pipedrive, spanning the company’s growth from roughly 100 employees through multiple stages of scale. He explains how infrastructure engineering differs from application development, with a longer-term focus on migrations, lifecycle planning, secure defaults, and building trust through stable, repeatable platforms.
At Pipedrive, the infrastructure team owns cloud assets end-to-end, communicates upgrade timelines clearly, and balances a “golden path” with flexibility for team needs. Kristjan shares how north-star metrics like four-nines uptime, latency, cost per seat, and security shape the team’s decisions — and how platform quality directly impacts customer retention.
Per and Kristjan also go deep on the technical stack: a long-lived PHP monolith, microservices fully on Kubernetes, a large MySQL footprint, Kafka, Elasticsearch, and infrastructure as code through Terraform and Ansible. They discuss lessons learned from AWS migration, including the realities of elasticity, cloud cost management, and cultural change after large infrastructure shifts.
The conversation then turns to AI in production. Kristjan shares how 85% of engineers are already using AI coding tools, what it looks like to run open-source LLMs on GPU-backed Kubernetes with sglang, and why teams still rely heavily on external model providers. They also explore the uncertainty of AI unit economics, build-vs-buy decisions, supply-chain hardening, insider risk, and why autonomous agents need strong governance before they can be trusted in production.
This episode is a practical look at infrastructure as a long-term systems discipline — one grounded in responsibility, customer value, and engineering judgment over hype.
(00:00) Why AI Models Keep Changing
(00:43) Welcome and guest intro
(01:46) Kristjan’s origin story
(02:47) Why he stayed at Pipedrive for 10 years
(05:25) What infrastructure engineers really do
(07:14) How to run migrations without disruption
(08:04) Ownership and accountability in platform teams
(10:22) The infrastructure team’s north star
(12:56) Uptime, latency, cost, and security metrics
(14:28) Culture shifts after AWS migration
(17:46) Pipedrive’s stack: Kubernetes, MySQL, and more
(21:28) AI transformation and coding agents
(23:59) What AI agents mean for SaaS
(25:27) Agents as microservices
(25:59) Trust, risk, and responsibility
(26:34) Running LLMs on Kubernetes
(28:01) Model sizing and infrastructure costs
(28:53) Build vs. buy for LLMs
(31:25) AI unit economics and pricing questions
(35:38) Supply-chain security in the real world
(40:04) Insider threats and access control
(41:26) Governing autonomous agents
(43:37) Infrastructure lessons and wrap-up
(45:01) Closing and sponsor
Kristjan Elias is the Director of Engineering Infrastructure at Pipedrive, where he leads platform and cloud strategy across a fast-growing SaaS environment. Over nearly a decade at Pipedrive, he has helped guide the company through major phases of scale, infrastructure modernization, cloud migration, and platform evolution. Kristjan’s work sits at the intersection of reliability, cost efficiency, security, and developer enablement, with a growing focus on how AI tools and autonomous systems can be introduced responsibly into production environments.
• Kristjan Elias — LinkedIn
• Per Krogslund — LinkedIn
• Learn more about Pipedrive
• Learn more about Docker
[00:00:00] Kristjan Elias: It's a constant figuring out, because obviously the, the models they keep constantly changing, right? For now, all of our use cases have not proved that we need these local hosted models, but we still have them for other reasons, for like internal reasons, uh, for educating people, for testing. That's always the, the challenge with, with services and products, right?
[00:00:19] Kristjan Elias: So you have to do it well enough for people to want to pay for it. And, and we have had this, uh, idea in the back of our mind that, that if, uh, if, uh, we are, uh, like a. Service provider for, for companies or, or, or usually, uh, b, B2B, uh, that, that if, uh, if the sales person or the customer is willing to pay for it out of their own salary, that's kind of the, the key, right?
[00:00:41] Kristjan Elias: So you have then, then you have a good product.
[00:00:43] Per Krogslund: Welcome to Ship Happens, a podcast where I sit down with some of the smartest people in the industry to talk about the skills, techniques, and mindset that helps you ship amazing software. Today I'm talking to Christian Elias, director of Engineering Infrastructure at Pipedrive, which is an online CRM provider.
[00:00:59] Per Krogslund: I'll let Christian [00:01:00] talk a bit more about that in a second. So Christian has worked in infrastructure at Pipedrive for more than 10 years, and before that he held data warehouse and CIS admin roles at Skype and Sweat Bank. So with that intro, it's a pleasure to welcome you to the show Christian, and thank you for taking the time to talk to me today.
[00:01:15] Kristjan Elias: Thank you per for having me.
[00:01:27] Kristjan Elias: I am really happy to, uh, join your podcast. It doesn't really, it doesn't happen too often for me, so,
[00:01:33] Per Krogslund: well, it's, it is honestly also new for me. Uh, I am, I am not a, uh, seasoned podcast host, so it is, uh, kind of new for both of us. So we, uh, we are gonna wing it and just have a, a, a good conversation, hopefully.
[00:01:44] Kristjan Elias: Alright. Alright, let's do it.
[00:01:46] Per Krogslund: We gonna start with you, um, just to kinda get your origin story. Um. You have a lot of years in infrastructure. You've been an ic, you're now a director, and, um, used to be a, a CIS admin as well, which I, I guess is SIT admins is what we call [00:02:00] infrastructure today in most companies.
[00:02:01] Per Krogslund: Um, and so, so I think one thing when I looked at your, your LinkedIn profiles that you, you've been a pipe drive for 10 years. Um, and I've been in the industry as well for, for a long time. And I think people routinely change roles like two or three years change jobs. Um. But you stuck around with pipeline.
[00:02:20] Per Krogslund: What is, what's special about this place? Like,
[00:02:23] Kristjan Elias: well, yeah. Um, so maybe to touch your first point, uh, c admins, our kind of yesterday's, uh, is how we call the infrastructure engineers yesterday or in the previous generation. And back then, uh, the, the focus was not on, uh, on this kind of automation as much as it is today, or it was, was, was now, um, Pipedrive.
[00:02:47] Kristjan Elias: It is a special place because it's 10 years ago when I was, um, um, switching jobs. Uh, I was not actually looking for a new job. There was pipe drive where I wanted to go because some of my, my [00:03:00] previous colleagues and, and friends were already there. And Estonia being, uh, a small co country, 1.3 million people in the IT industry.
[00:03:09] Kristjan Elias: Everybody knows everybody basically. So, so they, I was working in Skype, uh, some of my friends left for Pipedrive, looked at Pipedrive. I want to go there and, and got the opportunity. And then, uh, throughout this 10 years, well, I joined when the company was a hundred people. A small startup at that time still go, uh, going into it.
[00:03:32] Kristjan Elias: What was it? B, b round, uh, BSer, uh, funding round. And, uh, from, from there on, the company has gone through many different life cycles. Or I can, I, I can honestly say I have been in four different companies at that time, during this time. 'cause we have had this stages of, of growth. So pipeline maybe. Luckily has been one of those, uh, startups that succeeded, you know, for every [00:04:00] 10 startups, maybe one gets to succeed.
[00:04:02] Per Krogslund: Yeah.
[00:04:02] Kristjan Elias: And with five, so
[00:04:03] Per Krogslund: did they, did they start off by doing something else entirely than they do now, or
[00:04:07] Kristjan Elias: No. Exactly the same. Okay. We started, uh, as the CRM, uh, sales, CRM built by salespeople for salespeople. The, the, the special there was that it was. Founded by five Estonians, uh, uh, half of them working in sales themselves, they needed a tool, so they find, found some tech people and built a tool.
[00:04:29] Kristjan Elias: That's how we were there.
[00:04:33] Per Krogslund: So, so what is it about it, like you, you came there and then you stayed for 10 years. You started as an IC and now you're a director, and so you, you've had like a, like a full career at Pipeline. What is it that makes you stay there? What keeps it interesting?
[00:04:46] Kristjan Elias: Yeah. Uh, I've had this for year to year.
[00:04:50] Kristjan Elias: This question that, uh, for myself that is it ever going to get, uh, easier or, or kind of boring and, and it hasn't so [00:05:00] far because again, we have to found, uh, find a way to reinvent yourself. I, I started as a fourth infrastructure engineer. Uh, and now, uh, the team is, is about 30 people, uh, different teams, different disciplines.
[00:05:14] Kristjan Elias: Over time, we have had to go through periods of hypergrowth and, and, and dealing with it, or periods of finding efficiencies dealing with it. So it's, it's always a challenge.
[00:05:25] Per Krogslund: So maybe we can talk a little bit about what is it that makes infrastructure engineers different than just you, you know, your standard, you could say application developer of, of, or, um.
[00:05:34] Per Krogslund: Service, uh, people, building services, web services, and so on. Um, my personal experience with infrastructure teams is something that they work with extremely long timelines that they are not just like we deploy once a day or like multiple times a day, but you, you more have like a long view of things. You have a little bit of a different way of looking at things.
[00:05:52] Kristjan Elias: Yeah. Well, the area o obviously changes constantly. The infrastructure, which we are talking about can be also called. [00:06:00] By another name like cloud operations or Cloud cloud engineering, which is basically what it is nowadays. And then the step further, or a step in another direction, it's also already an necessary engineer.
[00:06:11] Kristjan Elias: So what makes the job, um, um, specific is, is maybe this long term. Uh, skill building and ownership of one area. So you have to, you, you, you're kind of in a place where a lot of other people have to trust. You have to have a lot of trust that, that this, uh, service or this base component is being well maintained in a secure manner, in a repeatable manner.
[00:06:39] Kristjan Elias: Your kind of, uh, basic jobs are to, to to build the trust to for others. Yeah. And, and, and, uh, that is maybe something that, that. While a service builder, application builder, they, they, they have a project, they go to the next project, uh, they, they iterate fast. Um, [00:07:00] that's where the, the timeline change or a difference, I think is so infrastructure engineer or cloud engineering, they have to own their, uh, domain and build their knowledge over a long time, uh,
[00:07:12] Per Krogslund: yeah.
[00:07:13] Kristjan Elias: In that area.
[00:07:14] Per Krogslund: And I guess it'll get, get rid of the old parts. Like you're, you're always moving and there's like things that just gets outdated and you need to get rid of it and migrate people off. And that's gonna be a Yeah. A year long exercise.
[00:07:24] Kristjan Elias: My migrations are, are a constant, uh, migrations and upgrades are something that, that just comes with territory all the time.
[00:07:31] Kristjan Elias: Right. So we have to, uh, and the easier may, maybe one of the, the jobs for these kind of engineers is also to, to handle these things. Uh, as, as a, with, with as small as impact for others as possible, right? So that, uh, you don't have to go and involve the application developers. Uh, you can main, uh, do it on your own.
[00:07:52] Kristjan Elias: You don't have to interrupt the customers. You find a way how to do it without interruption. So it's kind of, uh, to take care of the [00:08:00] platform as, as, as. Inly as possible, right?
[00:08:03] Per Krogslund: Yeah. Yeah. And I guess that's always a balance, right? I've seen different takes on this from, from different infrastructure teams where you always have this balance of you can say the individual autonomy of the individual developer wanting to do something, and then there's, you know, then there's the other part of like, you have a golden path or you have a standard, or this is the way things are done.
[00:08:21] Per Krogslund: And there's like the responsibilities of the platform team of, of creating like a. A general only like healthy platform.
[00:08:27] Kristjan Elias: Yeah.
[00:08:27] Per Krogslund: And so when is it you as a platform team step in and like, no, we, we are gonna update your service now. And like that's our responsibility. And when, like in some, some other organizations are like, no, developers are responsible for updating it, otherwise they can't guarantee uptime.
[00:08:40] Per Krogslund: Right. You have this, this thing of like, where is the responsibility and accountability? I don't know. How do you do it at Pipedrive?
[00:08:47] Kristjan Elias: In Pipedrive, uh, we have, I, I know that in, in some of the companies, the, the, the ownership of different infrastructure assets is, is very much distributed. So
[00:08:57] Per Krogslund: yeah,
[00:08:57] Kristjan Elias: you have even to a degree that, uh, [00:09:00] everybody gets their own a s account.
[00:09:02] Kristjan Elias: So set up your, your, your things and, and, and run with it. Uh, in we have chosen to. Uh, to have all of the asset ownership of the, of the, of the cloud assets, uh, owned by infrastructure department themselves. So, okay. Uh, all the base components to every single virtual machine even is owned by the infrastructure engineers.
[00:09:23] Kristjan Elias: And, and it, it, the asset has a customer that has, has a, as a, as a. As a feature which is owned by our product tribe as a customer. And, uh, all of these components that are built, they have kind of a, some kind of a life, life cycle, some kind of, uh, um, end of life set as well. So when we know that there will be, um, an upgrade that needs to be done, a migration needs to be better done.
[00:09:48] Kristjan Elias: It'll be planned with a timeframe that it, they will, it'll be announced there. Uh, grace time is given. And then it's executed, basically. But it, it is usually, uh, still [00:10:00] based on agreement. So if there's pushback, we, we manage this.
[00:10:04] Per Krogslund: Okay.
[00:10:04] Kristjan Elias: Uh, depending on, on, on the, on the level. Or the scale of the change, obviously.
[00:10:09] Kristjan Elias: Okay. We have had these, uh, huge changes, uh, with the whole platform, which, which kind of need, even need, have needed to have a full shutdown of the, of the, of the product. So this is,
[00:10:20] Per Krogslund: I guess it's always exciting. So for your, for your team today and like. In, well, it's 2026 now. I was about to say 2025. But, uh, for, for like a high performance like platform team, what is, what is the North Star like, what is that?
[00:10:35] Per Krogslund: You can say this is the thing we are doing and we are doing it really well. Like what is your, like, ultimate reason for being? Hmm.
[00:10:42] Kristjan Elias: So, uh, it's a good question. Because we have, uh, we have to think of it ourselves, uh, constantly. And then we are kind of sit getting into this goal setting time again of the year.
[00:10:54] Kristjan Elias: We have to set your annual goals, uh, as a company, pipe drive and, [00:11:00] and many others. We have obviously goals of, of growth, but companies grow. We have, we need to have revenue growth, uh, which is, uh, difficult to directly. Um, influence with, with, uh, platform teams such as, uh, such as ours. Uh, so, so we have opportunities on the other end, uh, for example, on the EBITDA side.
[00:11:22] Kristjan Elias: So you have to build efficiencies and try to find efficiencies, efficiencies in the platform, which you have done over the last couple of years. Quite extremely, uh, found different ways of, uh, decreasing our like unit economy. Uh, but the main thing we kind of have to keep an eye on is this, uh, is this trust and the quality of the platform so that, uh, uh, we, if we cannot, uh, directly affect the growth, we can affect the churn.
[00:11:56] Kristjan Elias: If it's, if it's related to the product quality or, or the,
[00:11:59] Per Krogslund: okay.
[00:11:59] Kristjan Elias: The platform [00:12:00] quality. So taking care of, uh. Of reducing, uh, the amount of incidents or, or, or performance issues, basically.
[00:12:08] Per Krogslund: Gotcha. Oh, so when you say churn, you say customer churn. Customer churn, yes. Customers are leaving because the, the platform is similar, stable enough, they see an outage to get fed up.
[00:12:17] Per Krogslund: They can't get data.
[00:12:18] Kristjan Elias: Yeah. So,
[00:12:18] Per Krogslund: gotcha.
[00:12:19] Kristjan Elias: I mean, that, that can happen or that has happened in the past. And then this is something that is, uh, directly avoidable by, by the team of, uh, the teams such as ours.
[00:12:30] Per Krogslund: Yeah, but I guess you're not fully in control, right? Because you, you're running the platform, but like a developer might have pushed like a terrible piece of code that breaks down and, and falls over.
[00:12:41] Per Krogslund: Uh, you know, that can also happen.
[00:12:42] Kristjan Elias: Yeah. Obviously it's not a hundred percent, uh, on, on the platform side. They always have, uh, different conditions that can come into play, but that we can always own what we can. Uh, yeah, what can come from, from our side.
[00:12:55] Per Krogslund: Yeah. So this is boiled down to specific metrics that you have.
[00:12:59] Per Krogslund: This is, this is [00:13:00] how we know it's going well. That we have like, yeah, we have an uptime of like 99.99, nine, whatever,
[00:13:05] Kristjan Elias: four. We usually four nines is our Okay. Four, nine targets.
[00:13:09] Per Krogslund: Yeah. Yeah.
[00:13:11] Kristjan Elias: And also latency. Latency in terms of, uh, of the, like the main product endpoints. Uh. Millisecond late.
[00:13:20] Per Krogslund: Yeah, yeah, yeah. So the, the, the customer in the end experience, this is a fast thing to use and it's responsive and they get the stuff they need and, and right away and so on.
[00:13:28] Per Krogslund: Yeah, yeah,
[00:13:29] Kristjan Elias: yeah. And the, and the, besides those which I mentioned before, is basically the, the, the unit economy or we use, uh, cost per seat as our historical, as our main, main, uh, financial metric. Yeah.
[00:13:43] Per Krogslund: Yeah, so always bringing down the cost per seat. Like you have a billion users and they cost 2 cents an hour, or whatever you have, but you have, you have some economics behind it.
[00:13:50] Per Krogslund: Yeah. So it makes sense. Yeah. Yeah. So you, you're, you're balancing efficiencies and stability costs. And I guess security as [00:14:00] well.
[00:14:00] Kristjan Elias: Absolutely. Security is, is, uh, is one of the, like next to some of these other main things like business continuity, continuity and all that. The security is, is, uh, something you, you cannot really compromise on.
[00:14:13] Kristjan Elias: Right. So yeah, the, the, our InfoSec department and uh, our kind of closest friends to work together with
[00:14:21] Per Krogslund: that is good. It's good to be close friends with forsake. It's like, uh, so, um. I spent a couple years in at Solando and also at Spotify, and so like they, they have like very opinionated platform teams in these people as their places.
[00:14:35] Per Krogslund: Those are like very specific places. They're enormous and so they do need to have like a very specific strategy, but I, that's one of the things I always found interesting about infrastructure is that you can, you can kind of shape the organization. Like the way people work with, with how you provide infrastructure.
[00:14:49] Per Krogslund: Well, you can kind of compare it to like how people build roads and that's how we get around. Um, so do you have like a, a specific idea for Pipedrive or [00:15:00] like how, how are you shaping people and the way you provide infrastructure?
[00:15:05] Kristjan Elias: Yeah, there, there is, there kind of are stories behind it because Piper, uh, migrated to a WSA couple of years ago.
[00:15:14] Kristjan Elias: Okay. From a, from a private cloud environment we had Okay. We had a private cloud setting in, in Rackspace using, uh, using, uh, cloud, uh, layer called OpenStack open source layer.
[00:15:25] Per Krogslund: Oh yeah, I
[00:15:25] Kristjan Elias: know. And, uh. And then coming from that, uh, or transitioning into a pay as you go model from, from a very fixed, uh, fixed, uh, contracted infrastructure, the you people really had this mindset that, uh, that everything is built for peak because we don't have to scale.
[00:15:43] Kristjan Elias: If you have fixed infrastructure, it's always. At peak, so you don't have to think about, uh uh, but the services being too efficient, uh, running with the resources that actually need. So that was, that was, uh, a time where we had, when [00:16:00] we, when the transition happened, we had to really, really go and, uh, reshape how people think or how they approach, uh, approach their work, uh, which was really.
[00:16:11] Kristjan Elias: A challenge, I would say. But nowadays, uh, we are, we're very elastic. And then, and, and that is on top of, uh, people's minds quite a lot. And that is a, like a sword with two edges, because now it may be, might be that people are too fixed on, on that the service is being efficient so that that sometimes, uh, silly things have happened because of efficiency or, or trying to.
[00:16:39] Kristjan Elias: Fine tune it until the ne until the last zero to get the, the last descent out of it and then cause incidents because of that. So, yeah. So you, you have this, it's a very, very much of a two, uh, edge sword kind of a view.
[00:16:52] Per Krogslund: Yeah. And I think this also, like what it happens in, in a lot of organizations that moves to cloud, they, they start off by just over provisioning and they just have a lot of [00:17:00] overhead ready for anything.
[00:17:01] Per Krogslund: And then, yeah, I think the exercise you've gone to, I think it's, it's rare that I hear about it, but I think in these, these years now where. There's a little bit more financial pressure on companies. Um, going through that exercise is hard, like finding that, like that right level between you not spending too much money, but we are also not under provisioning and we can't deal with a, with a certain spike.
[00:17:21] Kristjan Elias: To, in the end of the day, we have to, uh, uh, take care of the customer's experience first, and, and then the, the cost is always behind it. Uh, just that have to balance these things, but the customer experience needs to come first. That's our business. I mean, yeah, that's our, our our, uh, that's how we look like it's a face.
[00:17:41] Kristjan Elias: If, if, if it doesn't work, then the customer can leave. That's
[00:17:44] Per Krogslund: it. Yeah. Yeah. So how does your, this might be a hard que hard question to just answer, like with a, with a, a simple answer, but more like, how does your, like your landscape like application level, are you, uh, Kubernetes microservices, like what, what kind of like [00:18:00]
[00:18:00] Kristjan Elias: mm-hmm.
[00:18:00] Per Krogslund: Structure are you using?
[00:18:02] Kristjan Elias: Um, pipe is, is a microservice house. Uh, okay. Well we started 15 years ago, so we obviously have a PHP mono monolith still there as well.
[00:18:11] Per Krogslund: Everyone has a PhD monolith, someone doesn't wear.
[00:18:15] Kristjan Elias: Yeah. But, but, but we started with the microservices before. There was a, a, like a huge push to containerize the environment here.
[00:18:23] Kristjan Elias: Okay. Like our, our first microservices were in, in Linux server as no, no pro processes. Okay. So that was, uh, interesting. But, uh, over time we, we went through. Uh, the talker, uh, talker, swarm, uh, phase oh talker. Swarm, U-C-P-U-C-P phase.
[00:18:43] Per Krogslund: Yeah.
[00:18:44] Kristjan Elias: And, and now, uh, we are fully in Kubernetes. All of our microservices, uh, are hosted in in Kubernetes.
[00:18:50] Kristjan Elias: Uh, we have a. A very huge fuel footprint of MySQL, uh, server or CRM is a, is a, is a essential [00:19:00] database, right? So,
[00:19:00] Per Krogslund: yeah,
[00:19:01] Kristjan Elias: so we have a, a, a very huge, one third of our footprint is, is MySQL. Uh, machines built in easy tools. Uh, we don't, uh, a s we don't use RS 'cause we already had a very well working, uh, database system, which we could kind of move over or, or, or recreate in the AWS environment.
[00:19:20] Kristjan Elias: So all, most of our, our services are, are. Vanilla ourself built. Uh, recently we have, uh, migrated or kind of slowly migrating from vanilla Kubernetes into, into ES clusters to
[00:19:31] Per Krogslund: yeah,
[00:19:32] Kristjan Elias: to remove that other layer of, of, of, um, required maintenance for this virtual machines. But yeah. Um, mostly, or not, mostly all of the infrastructure that we have is, is built with automation.
[00:19:48] Kristjan Elias: We have, uh. Very high level of infrastructure as code with, uh, Terraform and, and Ansible behind it. Everything is, is done with code, no manual [00:20:00] work at all. Um, and lots of like supporting infrastructure services as well. So most of it is in Kubernetes. There are databases and then there is in the middle Kafka clusters, Elasticsearch a bunch of other things as well.
[00:20:17] Per Krogslund: Cool. So yeah, I can, I, can I get, I start to see why it is been, been interesting for 10 years since you've gone from like PHP, monolith, Docker, swarm, Kubernetes and so on. So you had many iterations of doing this the right way.
[00:20:30] Kristjan Elias: Uh, well, we had many. Yeah. Uh, I would say that the, we may have also had, uh, a bit of like too much freedom or, or time to do it how we wanted to do it.
[00:20:42] Per Krogslund: Yeah.
[00:20:42] Kristjan Elias: And some other companies that, that maybe had, uh. Uh, different pressures, uh, from time to time. Maybe didn't have time to do it correctly. Yeah. But we had some, some, some more of that opportunity. Uh, so compared to. Some of [00:21:00] our local Estonian competitors, maybe like we are in a different place or in a different readiness of, of the platform.
[00:21:07] Kristjan Elias: So
[00:21:07] Per Krogslund: yeah,
[00:21:08] Kristjan Elias: quite happy about that.
[00:21:09] Per Krogslund: Yeah.
[00:21:10] Kristjan Elias: Yeah.
[00:21:11] Per Krogslund: I can't remember what the name of the law is, but this thing of like a, a, a complex or advanced system doesn't start off as like an advanced system. It that like a simple system and then it's grown out like you, it's very hard to build like a, like an advanced system from day one.
[00:21:26] Kristjan Elias: That's true. That's true.
[00:21:28] Per Krogslund: So if we, if we look a little bit ahead instead of like, you have, you have this thing now, and now we, like, everyone seems to agree that 2025, this like year of AI and everything is changing and people are working in new ways. Like, how do you feel that on the infrastructure side, like how is it impacting, like, I'm guessing your team is probably using some AI agents, but also like, are you feeling the way that you have to like support developers?
[00:21:50] Per Krogslund: Like, um, has this, has, is this thing going to change? Do you feel the pressure for that?
[00:21:57] Kristjan Elias: Well, absolutely. Things [00:22:00] constantly change and need to, uh, that, that's, that's totally fine. I, uh, I have also been very close to the company's, uh, AI transformation during, um, like this year and half of the year lately. Um, we, we, as being, uh, private equity led the company.
[00:22:20] Kristjan Elias: We we're kind of in a, in, um. Rut maybe for a while. And then saw this ai AI transformation happening outside. And I'm like, oh, let them do it right. And, and then at some point woke up that, oh, oh, oh, oh hell, we are, we are staying, like leaving ourselves behind. And from that, we have had a really, like, uh, I would say a rush, but a successful rush.
[00:22:44] Kristjan Elias: Nowadays, uh, 85% of the engineers use, uh, the coding systems constantly. Uh, infrastructure engineers included. So my department has maybe been [00:23:00] from, from day one, one of the early adopters of the, of these different, uh, um, capabilities. Uh, before there were cloud code for request, uh, reviews. We all had these already ourselves for infrastructure.
[00:23:15] Kristjan Elias: So, so. As for the future, I think this is a kind of an opportunity. So if we have infrastructure as code at the moment, which is like we, we, we are, we're declaring infrastructure. We're building infrastructure, uh, um. In the future, this maybe allows us to, to have infrastructure software, uh, that, that we leave agents to run this software.
[00:23:39] Kristjan Elias: So right now it, it is, maybe there is a, there is a psychological barrier for people that, hey, I have, I have declared this, this kind of piece of resources and I need to kind of own it or manage it. But if, if you abstract it all of it away into a software and then the agent managing is also software.
[00:23:59] Kristjan Elias: [00:24:00] Then the next layer already comes. Right. So it's, yeah, it'll all, it'll all change, uh, over time, maybe in another 10 years. Uh, we only have two people,
[00:24:15] Per Krogslund: but also I think like it is doing with, with coding agents, it seems like that's the first use case where we really see these come into play and be efficient and really helpful. And they're really starting to move a lot of things, but. Do will we, do you think also we will start seeing that like something like Pipedrive will start having some AI functionality that you as an infrastructure team needs to take care of?
[00:24:36] Per Krogslund: Like hosting, lms, or like supporting hosting an agent on your infrastructure and so on, so you can start offering as features to customers? Or is this like still far away?
[00:24:46] Kristjan Elias: Well, this is real already today. We, so, uh, as a hosting lms, we actually do that, uh, already. Okay. Uh, this, uh, open source LMS for specific product use cases and, and, [00:25:00] uh, and working on.
[00:25:02] Kristjan Elias: Uh, like an agent platform for our product is, is kind of very much in the works at the moment. Yeah. Uh, it's a, like a brave new world or, well, in the end it's still, it's still all software. It's just, it's a matter of changing, uh, the approach or the thinking that, that you have to kind of abstract some things away to the agent level and then, and.
[00:25:24] Kristjan Elias: Then we'll see what happens with all of that.
[00:25:27] Per Krogslund: Yeah, I think there's a good, I think there's a good perspective in thinking about agent, maybe Agents are the new microservice. It's just a small unit that does something. So you give it an input, you get an output, uh, and maybe these, these services will start pinging each other and you have a, all of a sudden you have a mesh of agents as you would just use ku.
[00:25:43] Per Krogslund: I don't know if Kubernetes is gonna be the answer for that, but it seems like Kubernetes has been the answer for everything the last 10 years. So.
[00:25:50] Kristjan Elias: Maybe. Maybe we, we'll see. Right. But, but yeah, I think that, uh. Oh, in all [00:26:00] of it there is the question of people. So PE people have their psychological barriers in, in, into this.
[00:26:06] Kristjan Elias: So it's the problem of trust with all of the LLMs and uh, and then the AI capabilities from from last year. This year. It's just a problem of trust. Can you? Yeah, so it's, it's capable already. But how, how far can you trust it if it doesn't have eyes, like go as a person? Yeah. In that sense, it's, it's the IBM question, right?
[00:26:29] Kristjan Elias: So if, if it cannot be responsible, it should decide. So,
[00:26:34] Per Krogslund: so actually, since you have LMS running in production, um. I think maybe just to take a little bit of a magic away, because people have a tendency of, of looking at these language models as a, a magical back box. And like, how do you run a magical black box in production?
[00:26:48] Per Krogslund: Like, is it, is it actually hard to do? What do you need to do?
[00:26:54] Kristjan Elias: It's not very hard. It's hard. So essentially what we, what we have, it's, uh, we [00:27:00] have Kubernetes clusters, uh, running on GPE two node. Uh, yeah. And, uh. With, uh, this component called sg l uh, we orchestrate the, an LLM downloaded from hugging face, let's say, is it a LAMA or, or something else.
[00:27:18] Kristjan Elias: So it's, it's, it's rather in end it's easy 'cause the sg l is just creates the rest endpoint. And then it's just a service. So.
[00:27:27] Per Krogslund: Perfect. See, thank you for taking a little bit of the magic away. I think just for people listening to this like, oh, that doesn't actually sound hard. It's not, not that magical. You can actually just use an easy 2D PU instance.
[00:27:38] Kristjan Elias: Yeah.
[00:27:39] Per Krogslund: Boom.
[00:27:40] Kristjan Elias: Scaling of course will be, will be a problem if you want to do it in how large, hard, like, really big scale. But, but, but, uh, uh, this is essentially the same as, as, as the big, big, uh, companies are doing it.
[00:27:53] Per Krogslund: Yeah,
[00:27:54] Kristjan Elias: they will do it on, on bare metal, but, uh, but yeah.
[00:27:58] Per Krogslund: Yeah, but you can run it [00:28:00] easy too. So, um, yeah.
[00:28:01] Per Krogslund: And I guess that goes back to the under provisioning. Over provisioning. Like how big of a model do you wanna run? Is like a small good enough or do you really want to go big because you need like a more complex use case? And I guess that's still something that we need to figure out. Um.
[00:28:15] Kristjan Elias: Well, yes, it's a constant figuring out because obviously the, the models they keep constantly changing, right?
[00:28:20] Kristjan Elias: So you have your different op, op options, opportunities, uh, choose which one works for what. And, and that's the key, right? So you, you, if, if you end up in a situation where you use the most powerful model for everything, maybe that's. That's too simple or too over or, or oversimplified. And you will pay for it.
[00:28:42] Kristjan Elias: So, so what's the magic or, or the magic area? You have to just find the correct use case for the correct, uh, resource.
[00:28:49] Per Krogslund: Yeah. Or
[00:28:50] Kristjan Elias: match. Match these things.
[00:28:52] Per Krogslund: Yeah.
[00:28:52] Kristjan Elias: Between themselves.
[00:28:53] Per Krogslund: Have you done any kind of cost calculation of saying, okay, if we run this thing on our, uh, cloud setup [00:29:00] versus getting tokens from open ai.
[00:29:03] Per Krogslund: Something that's like kind of similar or at least good enough. Uh, have you made like a cost calculation of saying, yeah, this, this actually does make sense? Or are you more doing it because you need specialized models or smaller models or like something very specific?
[00:29:16] Kristjan Elias: Well, we started doing it last year, um, more, mostly because of, uh.
[00:29:24] Kristjan Elias: Unknown, uh, situation. So we, we, we ha have and had ambition to build this, uh, AI based features into the product. We didn't know what will be the impact, uh, or kind of, we couldn't even, even, uh, forecast to, well, what would be the impact in terms of costs. So we prepared ourselves by creating also this opportunity to, to have locally hosted models.
[00:29:48] Kristjan Elias: Or ourselves, uh, hosted these lums. Um, but until now, uh, all of the main use cases have, have remained with, uh, external providers because, uh, we have been [00:30:00] able to choose or there, there has always been these opportunities to use this many or small models where, which make just sense. It's just easier or it's just, uh, it's you basically by using the third party, you, you offload a lot of the responsibility Yeah.
[00:30:14] Kristjan Elias: For it. So it's it for now. All of our use cases, uh, have not proved that we need these locally hosted models, but we still have them for other reasons, for like internal reasons, uh, for educating people for testing. Uh, I'm sure as we progress, uh, however, uh, that this will come into play, uh, with, with this very specific.
[00:30:40] Kristjan Elias: Specific use cases for the product.
[00:30:43] Per Krogslund: Yeah.
[00:30:44] Kristjan Elias: But for that to happen, uh, we have a ways to go to figure out kind of this, uh, uh, workflows how to we, how to like identify the correct, like there is a use case, but what's the correct, uh, [00:31:00] or the most reasonable model. And then the way of using the model is, which is a bit of a, off a topic at the moment.
[00:31:10] Per Krogslund: Yeah, I guess so. I guess just like figuring out, okay, using this model with this like context size and these settings, that means that for my cloud bill, I guess you have like an additional, like an additional layer. You're not just thinking about tokens anymore. You're really thinking about like, what does this cost me in DPU and ram?
[00:31:24] Kristjan Elias: Yeah. This is actually a, a, uh, a, a valid topic also if, if you think about the unit economy layer, uh, of, of, of a product such as ours. So if you take a. A non ai uh, system. Then we have this kind of a fixed, uh, unit price because the code or the machine or resource is always there. You can ask how many times you want.
[00:31:49] Kristjan Elias: It'll still cost the same because you have these hours on this resource. But then, and, and the customer for us, let's say it's a dollar per seat per month. So it's, it's easy. [00:32:00] Uh, then if you, if you, if you take into account this, uh, agent, uh, systems or software where actually every kind of a request or, or, or movement costs, so then.
[00:32:17] Kristjan Elias: Uh, like the, get this out the door or Basically we don't know what it'll happen because
[00:32:22] Per Krogslund: Yeah,
[00:32:23] Kristjan Elias: based on the load or based on the, on the user's interactions. Or let, let's take emails. So emails come in, there is a agent evaluation of the emails. So before that, we, let's say we have a algorithmic, uh, evaluation.
[00:32:38] Kristjan Elias: It'll always cost the same so we know how much it costs. But if, if it's agent, everything goes to an LLM and LLM is third party or cloud bill based, then well, it, it can, it can have a crazy cost behind it. Yeah. The unit economy goes out of the door and, and there is the place where. Uh, like the, the, the local [00:33:00] LM LMS or, or, or et cetera, or these various purpose built solutions will come into play and, uh, or also likely u usage based pricing for the pro, for the products or the services or,
[00:33:12] Per Krogslund: yeah.
[00:33:12] Kristjan Elias: Outcome based pricing. So it's. Brave New World again.
[00:33:18] Per Krogslund: Brave New World again. Yeah. Yeah. Interesting perspective, like how it also gonna Yeah. Impact, like actually usage, pricing for end users to actually start. Because right now people kinda expect that everything is free and I'm, I'm guessing that's not gonna, like, for like you, you can see you go to Google and they just already have like an AI answer for you, but it's like, that must.
[00:33:37] Per Krogslund: That must also mess with their uni unity economics. Right? They, they spend an enormous amount of money to always just default to an AI answer, and they can't be making that much of like search query compared, considering how much it cost to run this thing.
[00:33:50] Kristjan Elias: There was at some point, I have forgotten already, where, where I see saw it and what was the actual unit, but, uh, but the, the, the comparison between a [00:34:00] normal Google.
[00:34:01] Kristjan Elias: Uh, request. Yeah. And they, I based Google request was something like 10 times or something.
[00:34:05] Per Krogslund: Yeah.
[00:34:06] Kristjan Elias: In terms of cost.
[00:34:07] Per Krogslund: Yeah.
[00:34:07] Kristjan Elias: So it, it cannot be, it cannot, uh, endlessly be like this. So it has, ah,
[00:34:14] Per Krogslund: if, if you have Google money, I guess you can do it for a while, but like, you know, a, a company like Pipedrive or like any other company in the world, basically can't afford to do that forever.
[00:34:22] Per Krogslund: So I guess at some point we'll start seeing the world even out. So it becomes like. Oh, it is actually reflected in the cost that you do as a user. Then it's, then it's interesting if it's good enough that people actually wanna pay for it.
[00:34:33] Kristjan Elias: Yeah, yeah, yeah. That, that's always the challenge with, with services and products, right?
[00:34:38] Kristjan Elias: Yeah. So you have to do it well enough for people to want to pay for it. And, and we have had this, uh, this, uh, idea in the back of the mind that, that if, uh, if, uh, we are, uh, like a. Service provider for, for companies or, or, or usually, uh, BB two P. Uh, that, that if, uh, if the sales person or the customer is willing to pay for it out of their own salary.[00:35:00]
[00:35:00] Per Krogslund: Yeah,
[00:35:00] Kristjan Elias: that's kind of the, the key, right? Yeah. So you then, then you have a good product.
[00:35:04] Per Krogslund: That's a dream. That's the way you wanna get to. But, um, but I guess it was like, I think it was like, uh, I think it was Cursor or one of the other ones, it was like they saw like this, that initial success was that, that, uh, developers start buying licenses for it, even though the local enterprise procurement unit didn't allow it.
[00:35:21] Per Krogslund: So they just bought it themselves and then they kind of knew that they had something essential that people
[00:35:24] Kristjan Elias: wanted. Yeah. Yeah, we, we saw that as well. And then we had this, oh no, don't, don't do it. Don't,
[00:35:31] Per Krogslund: don't, don't, don't, don't do shadow it, please. Yes,
[00:35:36] Kristjan Elias: exactly.
[00:35:38] Per Krogslund: So actually, speaking of shadow it, I guess, um, it's also been a year of like, this, this makes it sound like a New Year's show is like looking back at the year, but it, it is been a year of like a lot of supply chain stuff, like a lot of supply chain, uh, attacks and um, um.
[00:35:54] Per Krogslund: You, you are, I guess, partly responsible for security. You have an InfoSec department. I guess from the platform side, you [00:36:00] also need to ensure that you provide like a secure baseline. Like is this something you can see is getting worse? Are we getting worse at dealing with this as a industry or is it just like business as usual?
[00:36:13] Per Krogslund: Just more of it?
[00:36:18] Kristjan Elias: Uh, you, you mean like are we getting at worse in responding or are the, the situations getting worse overall?
[00:36:24] Per Krogslund: Yeah, I think it's like, I think there's like the response and then I guess there's like, why is this happening? Because we have, we've been producing a lot of software the last couple decades.
[00:36:32] Per Krogslund: We're using more and more softwares. It's kind of natural that we also see more and more flaws and like, no, no software is flawless. Right? So it's. I guess it's kind of just natural that we see more and more supply chain attacks and, you know, like issues, uh, with the depend dependencies we have. Um,
[00:36:48] Kristjan Elias: well, in, in some cases I would say that, uh, we have ourselves also left, left the door open too much.
[00:36:56] Kristjan Elias: I mean this, with this MPM attacks, uh. [00:37:00] Our, we, we could, could have mitigated it, uh, ourselves before, uh, already. Uh, but we didn't because that wasn't, uh, it didn't hurt enough. Maybe that was the, that was the case. But we have gone through, um, a lot of hardening in Pipedrive again, maybe throughout the year.
[00:37:19] Kristjan Elias: Uh, we are a part of the, or we have a. Private equity lead. Vista Equity Partners is our, our, our main investor at the moment. And they have a lot of huge pro like portfolio of other companies, software companies. Yeah. And then, um, from there we, we get our fair share of, uh. Of, um, guidance or, or, or more like, or marching orders, like how far you have to, how far you have to, uh, uh, harden the systems.
[00:37:52] Kristjan Elias: And then we have, we have, luckily, I would say I. Before it has happened, being able to, [00:38:00] like, for example, re IP restrict all of our AWS environments, all of our Google Cloud environments or our GitHub, uh, et cetera, et cetera, before the, some of these attacks happened.
[00:38:12] Per Krogslund: Yeah.
[00:38:12] Kristjan Elias: And we're able to mitigate it already there, uh, but we as easily wouldn't have done it.
[00:38:22] Kristjan Elias: Also, I mean, it's like where does the push come from? And then many other companies got hit with, with these things.
[00:38:30] Per Krogslund: Yeah.
[00:38:30] Kristjan Elias: Um, I think, uh, in we are getting better, better and better at it, but you see that, uh, there are these, uh, interests in play, you know, so if you listen to the security people, a security department, we're very bad, right?
[00:38:45] Kristjan Elias: So.
[00:38:48] Per Krogslund: I used to work in security and those, I I I, I also understand security people. Uh, they are like the risk people and they do need to see the problems before they happen. Need to, they need to push that other way. Uh, but
[00:38:59] Kristjan Elias: you [00:39:00] know, the most secure system is, uh, is a shut down system. So
[00:39:04] Per Krogslund: Exactly. If you just turn it off and delete all the code and go home, go live in the forest, it's completely fine.
[00:39:09] Per Krogslund: Then you don't, you don't have CVS in the forest really offline. That's, that is true.
[00:39:13] Kristjan Elias: Exactly.
[00:39:13] Per Krogslund: Exactly. Yeah. Yeah.
[00:39:15] Kristjan Elias: But yeah, overall the focus needs to be there. Uh, it's, it's just again. It. One case is enough for all of the revenue to go, so yeah, it
[00:39:28] Per Krogslund: is. Yeah.
[00:39:29] Kristjan Elias: Be really careful and mindful about, about this.
[00:39:34] Per Krogslund: Yeah.
[00:39:35] Per Krogslund: Yeah. There's been some cases over the last decade where it's like, oh yeah, that broke that company. And I guess a lot of people have looked at that and learned, okay, that's, that's a risk we do need to deal with. And then it's more like finding the right balance of like, we don't need to be like completely locked down.
[00:39:49] Per Krogslund: We don't need to burn the servers to resecure. That's not the, that's not the answer either, right? You need to kind of find the balance between saying, well, this is secure enough and we are not like creating too much friction as [00:40:00] well. Like people can still build interesting things.
[00:40:02] Kristjan Elias: Yeah. I think one of the, something I've struggled with myself in terms of thinking about it is that, uh, what is the biggest kind of threat to a, to a, to a company such as ours?
[00:40:16] Kristjan Elias: It is an internal malicious sector, right? So if you have, uh, somebody that has enough privileges to do something like, let's say in my team infrastructure department who have, uh, those privileges and, and they have, they become malicious for some reason.
[00:40:32] Per Krogslund: Yeah.
[00:40:33] Kristjan Elias: So that is kind of something, uh, how do we protect from something like that?
[00:40:38] Per Krogslund: Yes. And
[00:40:40] Kristjan Elias: basically at the same time,
[00:40:41] Per Krogslund: for something like your team, that it really has the key, the keys to everything. Like you can, you can do a lot of damage if you're in infrastructure.
[00:40:47] Kristjan Elias: Yeah. Yeah. And, uh, you, you can obviously, like, so you have people building it, something, and you, in, in an, in an extreme case, you have location, you build it, but then gimme the keys [00:41:00] or I take away your keys.
[00:41:01] Kristjan Elias: So you kind of, and then you, if you want to do something, I give you the keys. So there is a, there is a watcher. But then who watches the watcher and, uh, endless. But, uh, all of this, this is, is again, like constant friction. So we have to deal with it. What's the, what's the balance there? So what's the speed?
[00:41:21] Kristjan Elias: It's, it's just constant compromise, right? So you have to find the correct way.
[00:41:26] Per Krogslund: Yeah, you do. And I think, I think it's really, I think it's really interesting how, you know, those two topics are kind of like. Converging. You have the, you know, the AI agents can, you really trust what they're doing? It is like, but then you have the other side of it.
[00:41:38] Per Krogslund: Well, if, if you had, if you had like really clear like, you know, data access policies, you had really clear, like managing keys and like, you know, then you could, can let an agent go and do the thing because Okay. It has certain permissions because you have your permission layer under control and it doesn't have access to data that it shouldn't have.
[00:41:55] Per Krogslund: So it couldn't go and delete it by accident. Like we, we've seen these things where like my, my [00:42:00] coding agent deleted my production database. And I guess the answer, the question is that, well, why does it have delete access to your production database? That seems insane. Um, yeah. So I guess it's just gonna continue being this, like finding the right balance.
[00:42:13] Per Krogslund: Especially considering if we start having these, like you can say more, you could say autonomous agents maybe start doing things that where we really need to be in control of what they can do and that like permission layer and the role roles based access controls and all that. So
[00:42:29] Kristjan Elias: yeah, I think in, in those cases.
[00:42:32] Kristjan Elias: This, this granularity Will, will, will work or will help. Yeah. There is, you don't have one agent that does, does it all for you. Yeah. You have, you have a, like a mesh that, that controls itself or kind of, yeah. Governs itself or you have a always an arbitrary component there that constantly controls and then as an as a software, it can be quick so you don't have to have this human control part.
[00:42:56] Kristjan Elias: Just, uh, have one LLM fight another and then that. [00:43:00] See how how it end.
[00:43:02] Per Krogslund: Yeah. And just, just see, uh, just see your LLM hosting costs go up as they fight over like small, random tasks. It's gonna be interesting,
[00:43:10] Kristjan Elias: but it, this fighting part, it's interesting. We had, uh, uh, had some cam fight, uh. Trading sessions within in the team as well, where one of the task was to kind of get two LMS to fight or kind of argue and see how, how, how they do.
[00:43:25] Kristjan Elias: So it's, it's, it's an interesting thing to watch.
[00:43:28] Per Krogslund: Yeah.
[00:43:32] Per Krogslund: Again, it's gonna be, it's gonna be interesting. It's a, it's a brave new world also in infrastructure. Is there anything you wanna plug as the kids say, or like recommend, is there any, like, any final last words of wisdom from the infrastructure side?
[00:43:48] Kristjan Elias: Yeah. Something I have maybe mentioned in, have been in another podcast, podcast as well a while ago.
[00:43:55] Kristjan Elias: Uh, is that the, the, the, the aim or, or the, the main [00:44:00] part of the business is not to run infrastructure is, is to get the business working. So, so, so the, the focus needs to be customers, customer value. And, uh, the minimum that needs to happen for the infrastructure to work. And then when you have time, when you have the opportunity then to make it, uh, make it perfect.
[00:44:18] Kristjan Elias: Um, I would say that, uh, often companies may be focused on the wrong thing and then fail, so
[00:44:25] Per Krogslund: yeah,
[00:44:26] Kristjan Elias: that's, that's not.
[00:44:27] Per Krogslund: Or they focus on a billion things,
[00:44:32] Per Krogslund: will it ever be, will your, will your platform ever be perfect? Now you try doing it like five or six times. Will it ever be perfect? I.
[00:44:40] Kristjan Elias: No, I, I think the, that's, it shouldn't be so if, if it's, if it's ready, then it's a bit weird. So, yeah. Something hasn't changed then.
[00:44:51] Per Krogslund: Yeah.
[00:44:52] Kristjan Elias: Or, or it gets too static. Uh, yeah. Uh, people will rest.
[00:44:56] Kristjan Elias: No, it cannot be that way. Things have to change.
[00:44:59] Per Krogslund: [00:45:00] Things have to change. Alright. And with that, that's, uh, that's a wrap for today. So again, thank you so much again for joining me. Christian, it was a pleasure talking to you. And if people want us to know more about you, where can they do that? I think most people say LinkedIn today.
[00:45:12] Per Krogslund: No one says Twitter anymore, but I guess LinkedIn as well. They can find you.
[00:45:16] Kristjan Elias: Yes, it's LinkedIn. Twitter doesn't exist anymore, right?
[00:45:20] Per Krogslund: No. Twitter doesn't exist anymore. It's gone.
[00:45:23] Kristjan Elias: It's gone.
[00:45:25] Per Krogslund: Uh, one thing that still does exist is dock, and that's the, that's the open platform for building, shipping and running apps, and that's what's, uh, support the show.
[00:45:32] Per Krogslund: Thank you for tuning in, and we'll see you next time.