Secure and Scalable AI Experimentation with Cloud-Native and Containerized Environments

Emrul Islam is the Chief Innovation Officer for Kasm Technologies.

Connect with Emrul:  

Kasm Workspaces is a platform that provides secure, browser-based access to desktops, applications, and web services by leveraging container streaming technology. It utilizes Docker containers to deliver these environments, allowing users to access them directly through their web browsers without the need for additional software installations.

[00:00:00] Todd Robinson: Hello. Hello everybody. So excited to be here today with Emrul from Kasm Technologies or, and we’re gonna be talking a little bit about Kasms Workspaces today and taking a couple different approaches to talking about AI. And so I’ll throw out a little bit about OpenMetal’s background in AI and then if you don’t mind, Emrul will follow me up right after this with your guys’ perspective.

But OpenMetal is a pretty fundamental provider when it comes to AI. So a lot of what we do is work in the hardware space. So we’re helping customers understand, okay, if I’ve gotta do inference workloads or training workloads, what kind of GPUs do I need? If mostly if we’re talking about training workloads.

But if we’re talking about inference workloads, which is a lot more common when it really goes into production, we also provide advice and hardware specific to if you are gonna be doing small tokens per second type jobs, that you can do it on CPUs certain types of CPUs that we run. And then you step into A100s H100s, H200s and things like that.

And so on our side, from the OpenMetal side, yeah, we’re pretty fundamental. So we’re excited to have Kasm here today ’cause I think they’ve got a, a different perspective. And so I get a chance to learn a few things that I’m always interested in the AI world. And I won’t get too sidetracked, but I always say to myself like, the, the AI space changes so much.

So some days I’ll be like, oh, I’m a five out of 10. Like, I know what I’m doing. And then something new comes out. You know expert, some, some, some various new ways to do this. And I realize, oh, actually I’m a three out of 10, and then maybe I build myself back up to a seven. So for me, I’m just yeah, excited to have Kasm here today to kind of talk through a part that I’m less familiar with, and hopefully it’ll be some education for the, for the crew out there watching.

So, again, Emrul welcome and yeah, if you don’t mind, give us some background on Kasm.

[00:01:47] Emrul Islam: Yeah, sure. Thank you, Todd. So Kasm started off as a US DARPA project for the US military. And essentially its goal is to facilitate secure collaboration. That that’s what the objective was, right?

So how do you do that? And what it is today, what came of all of that is a VDI replacement. That’s how people see it. But under the hood what we do is we instead of with a VDI technology like Citrix or whatever, you have like a virtual machine. That you are basically remoting into with Kasm When you request a session, we spin up a Docker container for that session and we stream that container to the end user, right?

So there’s a number of advantages to this approach of doing things. For start, it’s much lighter weight where you have a virtual machine just to run an operating system might need four gigs of ram. A Docker container really needs nothing except whatever applications are running in it. And then the speed that you can boot up a Docker container is much faster versus starting up a virtual machine.

And they’re much lighter weight, so you can tear them up, turn them down. And then we should also mention that the, the way that you manage these containers is much easier. We so a lot of us have had pain of having to build go golden images for VDI or even sort of AMIs on Amazon. Keeping them up to date is quite difficult and it’s time consuming.

With our approach, it’s all very modern, it’s all Docker file based. We support all sorts of modern integration tools like Terraform and Ansible. So we’ve basically made the VDI lighter easier, not as cumbersome and just a bit more enjoyable to work with.

[00:03:26] Todd Robinson: Yeah. Very cool. That’s what I said. I was excited to talk to you guys ’cause it is such a different take on it and I, and, and it makes sense of course, based upon where the business came from.

Yeah. So I’m, I’m excited. I guess how, where do you want to jump to in this, because I think that you’ve got a set of. You got a set of things to teach us. So if you, if you don’t mind, maybe give us the reason, like maybe some of the concerns that you guys have when you’re trying to support customers using this.

Yeah. And then I suspect you’ll, you’ll teach us a lot about how it actually is happening with, with your product and the way that you guys are approaching this.

[00:04:01] Emrul Islam: Sure. So I mean, we are, we are not a natural AI company, right? We sell a VDI products, but what’s, what’s happened is because of our background in facilitating secure collaboration that actually ties over with many of the AI requirements that, in the marketplace today, right? So we are speaking to a number of customers healthcare providers, et cetera. They want to do AI training. They want to support researchers, having access to potentially sensitive information. And then there’s all sorts of issues around how do you get the data to the researcher, or how does the researcher operate on the data whilst maintaining your security or obligations with respect to regulations that you might be operating under.

How do you do all of that? And especially how do you do it at scale, right? Because if you’ve just got, you know, a team of one or two researchers. That’s not a problem. But if you’ve got 20, 30 researchers and they all need different libraries, they all need different data sets. How as a security person do you even make sure that you’re adhering to everything that you need to adhere to?

And that’s where we can help. And we help just because that’s where we’ve always been able to help without product. ’cause of the isolation that it offers. The, the fact that all that runs on the end user machine is a standard web browser. So they don’t need to download, install, or run anything.

And they don’t need to download any data that might be sensitive.

[00:05:20] Todd Robinson: Yeah, no, that makes that makes total sense. I mean, there’s this constant battle where to be honest, really good, good engineers or developers will pull things down into local environments. It’s just such a habit. And in many cases there’s good reasons efficiency wise to do it.

Some of it’s historical. But that’s really interesting that because you’ve been able to combine essentially this new concerns that people have with AI. So yeah, you have a proper, giant training set, maybe. Where is that supposed to live? That’s not supposed to live in random places, right? That should live in a controlled structure.

And following yeah, that, that process. Yeah. That’s really cool that you guys support it that way. Yeah. I keep, keep going. If you don’t mind. Is there example use cases, like I had mentioned, Hey, there’s gonna be a training data set, there’s gonna be maybe a giant labeled data set or something like that, that is Well, yeah, yeah.

Take us through maybe some examples.

[00:06:12] Emrul Islam: Well, one that’s really easy to understand, I mean, researchers and, and AI developers, they’re, they’re pretty good with Jupyter or Collab notebooks. They get a lot done right? But there are some class of problems where you just need a desktop environment or having a desktop environment would reduce the friction in their day-to-day work.

Right. And one of those areas is, for example, reinforcement learning, right? Mm-hmm. So reinforcement learning where you take like where you’re teaching a computer or an AI how to, to solve a problem without giving it labeled answers, right? A lot of those problems can be visual. I know, I know in the LLM space, it’s not visual.

But if you are like trying to train an AI to automatically maneuver a drone or a racing game play that sort of environment where it’s visual, it kind of need to see what’s going on. As it’s progressing and a lot of problems. I know when I started this out a few years ago, I was, I was learning to play doom.

I don’t know if you guys ever played Doom. I think everyone has

[00:07:10] Todd Robinson: like, oh yeah, yeah, build, build, build worlds and doom. Why not?

[00:07:13] Emrul Islam: But it’s a fun little exercise trying to get doom to like an AI bot to play doom, right? And what I found is it would keep getting stuck. And I couldn’t work out just by looking at the logs, why it was getting stuck, and then I actually played a video of what it was doing, and it was actually getting stuck in the corner of a room and it couldn’t navigate its way out.

So being able to see that visually is a huge eye opener for certain use cases. And when we look at AI, we are not just talking about LLMs anymore, we’re talking about video we’re talking about audio, we’re talking about multisensory inputs and outputs. And the researchers need a good way to be able to interact with that.

And it’s, it’s all great if they’ve got a laptop or a desktop where the graphics card and everything else, but. You know, a lot of people are remote these days and they don’t, and then you’ve got the data problem. So I actually have a quick demo to show you about what I’m talking about here.

It’ll take maybe 90 seconds but you’ll get a quick idea. So if, if you want, we can jump into that.

[00:08:08] Todd Robinson: Yeah, yeah. No, that sounds great. And definitely, I, I relate to that especially if you’re trying to do things at scale, having different members of the team using lots of different environments that may or may not be configured the way that they actually anticipate it instead having a standardized environment.

Yeah. It’s really cool to see this coming together, the VDI world with this. Right. And it makes absolute sense. Right? Yeah. Yeah. So yeah, go, go for your demo, but yeah, excited to, to, to hear about this stuff. Yeah.

[00:08:43] Emrul Islam: Okay. Give me a second. So this is a Kasm workspace. I’ve set this up. It took me about five minutes to set up earlier today. So in the demo today we’re gonna be looking at this library. It is called reinforcement learning. Baselines three zoo. So they have a bunch of reinforcement learning examples.

This is for educational use. More than anything, it’s, it’s not doing anything practical. But you can train agents and you can view their training results and you can plot the results, et cetera. So we are gonna be looking at one of these agents that we’ve trained. And I’m also gonna show you what I have to do.

So if anyone is interested I just cloned the repository. I, I installed the requirements. And then I’m gonna run this command here. Okay? So got the command ready to go here. And when I run this command it’s gonna load up an AI model that’s already been trained in this case to do something very simple, which is to train a cheetah, how to walk or simulated cheetah, how to walk in a veteran environment.

And while it’s not walking as gracefully as a real life cheetah would, you can see it’s able to move along. And this is an example of, Hey, I’m able to do this visual demo. I can see what’s going on. If the model was limping along, I’d have some idea as to when my training I went wrong. Whereas if I’m looking at logs, I wouldn’t have such an idea.

So this is what Kasm enables me to do and to run this, I didn’t need to install anything on my local machine. Instead I just logged into Kasm. And you can self host it in the cloud or in prem in your data centers. And it’s able to use things like the Nvidia Coda acceleration that’s on the host machines to basically facilitate my use case here.

[00:10:27] Todd Robinson: Yeah. Very cool. Yeah. If you don’t mind, tell us a little bit more about what is on the other side of that then? It just, yeah, it figures, you know, I’m in the hardware side, so right away I’m like, oh, okay, wait a minute. What is on the other side that allows to support both what you guys have to do in the streaming side but of course to, to run the workload itself.

[00:10:45] Emrul Islam: Mm-hmm. Mm-hmm. Let me show you. It’s a really good question. So first of all, lemme show you.

Yeah. Yeah. So Kasm actually can all run on a single server, and we have a free five user version, or five concurrent session version that anyone can install. Home labs sort of tinkerers evaluation. So everything can run on one server, right?

But essentially if we, if I look at the multi-server installation, so I can tell you what’s going on Kasm essentially has a control plane and then it’s got these agent servers, which are either virtual machines or bare metal machines that you set up with Docker. And when a user logs in and starts a new session Kasm will direct the request to one of the agent servers and it will spin up a Docker container and stream it to you using our VNC technology.

So I’m gonna give you a an example of what that looks like. Here are all my workspaces. I was showing you one earlier that I was already booted into. So let me go into a Cuda enabled image. I can choose some options. All of this is customizable. So I can say that I want Python three in here and that I want PyTorch.

And let’s say I want a bunch of PIP packages. And then I just run this and this is when the API is connecting to that agent server. It’s spinning up that docker container and it’s streaming it, and you can see that that happens within a second or two. Yeah. Are

so we are going to launch one now. So let’s just pick this one. And I can pick like some configuration parameters. These are all adjustable. You don’t need to have any at all. But this particular one that’s for ai, lets you pick what version of Python you want, which a ML framework, if you want some additional PIP packages.

And then you launch a session. And when you launch the session, that’s, that’s when it goes through the agent server and spins up that duck or container that’s just for me, and it’s streaming that Docker container straight to my machine. And now I’m in here and I can have all sorts of apps. I can have a fully blown instance of Google Chrome in here.

I can also have Visual Studio if we’ve got some of the Jet Drains products in there, if that’s what people prefer. So it’s all, customizable and it’s all like having your own laptop, but in the cloud. Okay. Yeah. Yeah. And with those configuration parameters we talked about, you know, how do you do things at scale?

When you, when you’re doing sort of AI research it’s a bit of a difficult problem because researchers need access to a wide variety of tools and libraries and resources, and they necessarily don’t know them in advance. But then in some environments, which are air gap to lockdown, how do you even facilitate that?

With the integration we’ve made for ai you can actually have a set of python builds that are instant, like set up with the packages that the researchers will need. And there’s like a shared pip cache on the host so you can tap into that. All of that is powered essentially by. You know, Docker features that allow you to basically integrate nicely with the, the containers that users are running.

[00:13:45] Todd Robinson: Yeah. Very cool. No, no. Like definitely this is a common problem in all kinds of places, whether it’s developers and whatever they’re working on AI development. Take, take me through then, if you don’t mind. I think in this case you’re running how, how does it work on the backend? What is actually behind the scenes?

This is designed to connect into a particular set of LLMs that are running on a particular set of GPUs or CPU that you’ve picked out? Yeah. Or how does that Yeah. How does that work?

[00:14:16] Emrul Islam: Yeah. Let me show you. Let me just go back. Okay. So this is the Kasm administration panel. Okay. And then if you look at infrastructure, you see all of the hosts.

That are available. So those are the agent servers. You can have as many as you need and each of those hosts will have some capability. So if I look in here this one if I look at GPU info you can see this one has an RTX 30 90. Okay. So all of that hardware that’s on those Docker hosts can be made available to end users when they’re running sessions.

Okay. So if I look at the workspaces we looked at this Cuda enabled Noble the one we just ran a second ago, and if I look at the configuration, I can see that we’ve said it needs one GPU. Okay. So what Kasm will try and do is when I launch a session with this workspace, it will try and place it on an agent that has one GPU available that is, that there isn’t another user using it at that time.

Okay, so this is really handy again for lab environments research environments where you may have eight to 10 GPUs or a little cluster. And by all means, no one needs them all the time, but someone needs to be able to log in and use one. So this allows efficient utilization of the underlying resource because the resource is only allocated when the session is required.

[00:15:39] Todd Robinson: Yeah. Got it. Nice. If you don’t mind, ke ke keep into the demo here, but if you don’t mind, describe a couple, a little bit more about the customer types that, that the, you know, the businesses or researches or, you know, maybe schools, things like that, that really are using the systems or that you’re targeting to use the system.

I think that’ll also help me and probably the rest of us watching. Yeah, yeah. Yeah, get a, get a sense of really, is this for me? Who’s it for? Yeah.

[00:16:04] Emrul Islam: Got it. Got it. Okay. Well, I mean, let me just show you very quickly. We have so many different applications. A lot of these are gonna be familiar to a lot of people.

We have video editing, we have image editing, et cetera. It gives, and, and these are all applications that our users are using which is why they’re here. Right. So in terms of who, who uses this, obviously we have a large number of customers in the defense sector. And we’re looking at at customers in the health industry as well.

There’s also people that just want they run this in their home labs. It’s very popular because it’s free for five concurrent users. You can get like a. A sort of desktop at home just by running this in your, in your home lab and be able to access your resources from anywhere in the world essentially.

So the use case is pretty broad. It ranges from people that just want a VDI replacement to people that wanna do specific types of desktop streaming. The, there are some people that use this just for browser isolation because you can run a native version of Chrome straight in this with with any Chrome policies that you have.

So it’s a, it’s a good way of doing that. And lastly we have a number of customers in the Osynth space. So they, they like Kasm because they can originate the traffic from for, for their session from anywhere in the world. So that basically means that they can leave footprints in a way that’s suitable for what they’re trying to do.

[00:17:27] Todd Robinson: Yeah, very interesting. Do you have in the, for, for the AI kind of discussion or the ML discussion what are you seeing now? Yeah, as, as users and then what would be a thing that I would buy if I was that user? From you all? Yeah.

[00:17:43] Emrul Islam: Well, let, let me show you something. There, there’s a couple of things here, so let me just get rid of that session.

So I, I’m not sure whether you would buy this, but I, I use this it kill so if I launch this one, oops.

Okay, not that one then. Let’s close this one. So sometimes you wanna forget, forget actually just AI researchers and developers. Sometimes you just wanna use ai, so you want to use an LLM. Right. And that, that has its own sets of challenges because where does the data go? Even if you trust open ai commercially how do you make sure your users are securely handling the data that they’re sending it?

So within a Kasm workspace you can actually set up, in this case anything LLM, so that all of the interactions with an AI can stay within the workspace. It’s like a little sandbox. So you can see here we’ve got this let me just get, grab my Gemini credentials. One second. I’m just gonna plug this in off screen.

There we go. Okay. So let’s say we’re gonna use the latest Gemini.

All right, so all of this stuff is gonna setting, setting it up within a couple of minutes. Here we go. So now I can connect to any LLM provider with this tool. It’s, it’s, it’s fantastic. But I’ve chosen to connect to Google, Gemini, say as a, as a company or an organization. We have a contract with Google.

We trust them to provide LLM services et cetera. But as a user, I also need to interact with it. And there’s a lot of, you know tools available to do that. Let’s ask a quick question. What, what is Kasm?

Maybe it’ll do a better job than I do. Gotta switch to a different model. Okay. I didn’t know why that happened. Bear with me.

[00:19:47] Todd Robinson: Yeah, you’re in, you’re in an unscheduled demo, so, but it does seem like you, you probably are using this quite a bit yourself, so

[00:19:54] Emrul Islam: yeah, I think when you use these things, you appreciate why, what the pain points are.

Okay, so

let’s see. Right, so. Again, this is all running in its native application. It’s talking to an LLM that’s in the cloud, but the application itself is running in somewhere secure. So if I have embeddings, et cetera, I can store them within my own private network, et cetera. It’s a succinct answer.

I think I elaborated a bit more, but not bad. Not bad.

[00:20:29] Todd Robinson: Yeah. Yeah. I think you beat, I think you beat it out. Yeah. So five outta 10 LLM got beat out by the expert. Yeah, I, it’s pretty, like, two things that are kind of popping into my mind is that, you know, there, this is somewhat, you know, all this is a little bit uncharted territory for everybody, but it feels like you guys are, you’ve figured out the combination of both the v the VDI business but where this can also take you, what, I guess describe a little bit to me maybe about how.

You all are listening maybe to the customers and developing your next things. Yeah. What might be, I guess, gimme some insight as to where this might be taking you all. Again, very cool stuff, but very always interested as to how you get your, what’s your muse is and how you’re kind of picking things to work on next.

[00:21:25] Emrul Islam: Yeah, so we get a lot of feedback from customers. We have a very active Reddit community as well. And we, we get a lot of YouTubers sort of mention us on, on their, you know, tech channels and whatever that that’s build us a huge community, which we’re really grateful for. And they’re always feeding back sort of requirements, asks, et cetera.

So the sorts of things we’re seeing right now we’re, we’re seeing more demand for automation tools. So we, we did, terraform not so long ago. We’re gonna be adding Lummi soon, et cetera. Mm-hmm. We’re seeing people also ask for more control over the network. So that relates to, Hey, I want to use a Kasm with A VPN, right?

So when I’m in Kasm, I wanna have a VPN that’s directing my traffic through for whatever reason, Paris. Okay. So we’ve integrated all of those sorts of features into the products. There’s a variety of use cases. It continues to surprise me even today, what use cases people have. So if you have a global marketing team they have a website, they need to see what that website looks like in different parts of the world.

How, how do they do that? And before Kasm showed up, they had vPN company, and they were switching VPN profiles and that’s a bit clunky. Now they can just basically log into different Kasms from different parts of the world and see what that website looks like. It’s, it’s much simpler.

Little things like that is, you know, we build them because customers have a pain point. We address them. So automation, networking we’re doing more around general sort of performance improvements. So believe it or not, multi-monitor is a big, big ask. So we are looking at a, a streaming, and I’m, and I’m showing you, you know, I’m streaming one desktop, but if I have two monitors and I wanna stream two.

Two virtual desktops, how do I do that? So we’re working on that. Okay. And that’s, again, something that’s quite, you know, for most of us, we don’t do it, but if you are a power user, you definitely want, we’re gonna need things like that. So that, that’s the sort of thing we work on. We’re adding new images all the time new applications mm-hmm.

To our repository. And lastly, we’re just making it easier for people to install Kasm itself. So we’re gonna have launches in the AWS marketplace, for example. So you can just click a button and install it and get it going very quickly.

[00:23:40] Todd Robinson: Alright. Yeah. Yeah. So you had several things inside of there.

So interesting. The, the, the multistream, I guess that you call the dual, dual monitors, or sorry, I forget exactly what you said, but yeah, that, that is an interesting challenge. Just, you know, this is a slight segue and then I see we have a question. We’ll come, we’ll come to the question. But what, what are the different skill sets of your team then?

Because you’ve got, that sounds like quite a few different expertises, engineering wise. Can you an idea of what the team looks like? And I know it’s a little bit different probably than where you thought I might ask a question, but it just ’cause you guys are covering so many different things. It sounds pretty exciting.

[00:24:21] Emrul Islam: Yeah, I mean, I, I’ve, I’ve been at Kasm now just about three months and the team is phenomenal. It’s, just very broad range of expertise is all I can say. And very open people as well. I think a lot of them are connected in some way to sort of sort of US government military work.

And that probably is why they understand the problem we’re solving very well because they’ve, they’ve got the experience of, Hey, we’ve gotta do things in a way that protects data from being exfiltrated, or we have to do things in a way that basically means people can secure certain assets in a, in a, in a very structured way.

Right? So that, that discipline and also that knowledge that they bring to the product it definitely, it definitely shines through. And that’s, that’s kind of a bit about the makeup of the team and then, you know, Kasm as a company very much. I. Curious. I would say curious is the word.

You know, if, if I wanted to, if I went to my boss and said I wanna try something I don’t know if it’s gonna work, he’d probably say, okay, we are curious. Let’s give it a go. You know, as long as it, as long as it doesn’t cost us a lot of money, like, you know, take a lot of time, let’s try it. And that, that’s definitely helped us because it allows us to try things.

Obviously when you do that, certain things don’t work, other things do work and we can basically mix and match. I see a question here about the type of models available. So yeah, good, good question. So. We support any of the LLM tooling that exists today. So again, Kasm is a streaming solution, so we’re not restricting you to whatever models you may or may not have.

But one interesting thing is if you’re running a private model, either say you’re running a private Azure open AI model, or even if you’re running an OLIMA model you can actually file those off and only allow them to be accessible from inside the Kasm workspaces, right? So it just makes the deployment and security of those things a lot easier because you know the front door that every user is gonna be coming through.

And that’s gonna be a, a web-based interface that’s streaming pixels to access your service. That means you don’t need to open firewall ports, you don’t need to worry about people literally copying and pasting the data straight out. ’cause it’s not easy to do when it’s on a video stream.

And with Kasm we’ve got some DLP control, so you can even even log down things like the clipboard. So if you don’t want people to copy and paste from a workspace and that can be very useful in the legal field if you’re doing m and a activity where people are allowed to inspect information but not necessarily take it away we can facilitate those sorts of use cases and all of that obviously has some bearing on ai.

’cause in ai you don’t want people to walk off with your HIPAA data, right? Even copy and paste is gonna be disabled. So we brought all of that to our tooling and, and that’s why it’s available here. I,

[00:27:05] Todd Robinson: yeah, very, very interesting. And I wonder, part of that question might be is to like what by, if I was just getting started with you all, just to get a feel for what might be there may, from that side, is there certain models that you, that are already Yeah, we know how to run them.

Here’s either the cloud provider or the bare metal that we would recommend, and here’s how you can just get going with that.

[00:27:28] Emrul Islam: So we don’t get into the specific models that a customer’s gonna be running themselves. I think that’s more of a technical aspect. I think that’s more OpenMetal territory.

But what we, what we will do is basically tell you that we’ve got, you know, a huge category log of software that everything that I’ve showed you today is available out of the box with Kasm, right? So if you need PyTorch, if you need TensorFlow, if you need any of the pi charm tooling and jet drains, all of that’s available.

Anything LM Easy diffusion, all of those models, tools are available to use. And further, we publish all of our Docker images. So if an administrator wants to customize it, adapt it for their own requirements, have their custom image, that’s totally possible as well.

[00:28:10] Todd Robinson: Got it. Alright. Yeah. Yeah. No, so understood.

Yeah, I think on our side. So we’ve, a lot of the kits that are out there now in a good way, we’re able to just use, utilize the kits. And so it’s relatively straightforward now to test which models work well on different hardware. And so yeah, I think at this stage it is, any of the modern ones have now caught up and whether Nvidia is putting out a NIM for it.

Or you can find it on hugging face and you can trivially bring it in on top of o on top of CPUs and GPUs. So we, the GPUs are obviously a more typical thing that you see. But in our case, we also recommend that if it’s relatively light inference tasks. Like development stuff, you can absolutely do it on CPU.

And so being able to configure that in a way that gives you sufficient kind of tokens per second. Mm-hmm. Especially on the modern the, we, for us it’s the gen four and the gen four introduced it for Intel. The Zon, the gen four, and then the gen five improved it. And I know that Gen six has also got additional silicon in their power or the.

The, the, the higher end one, there’s, they’ve now separated it into the, to the two different types. Mm-hmm. The P cores and the e-courses where the P cores, we’re still working on that. We actually haven’t adopted those yet, but we have published for gen four and Gen five compared to like a one hundreds and, and et cetera, so that you can get a sense of what that will be.

So yeah, I guess if, if what you’re saying is is then that really is the choice of the company that’s implementing it, what they have decided to have on the backend both including the models and then how the models are actually being run. They’re adapted. Yeah. Yeah, yeah. Yeah. On

[00:29:51] Emrul Islam: CP architecture, we support ARM and X 64.

So arm CPUs are sometimes more efficient, most of the time, a bit cheaper as well in the cloud. And on the GPU side you know, getting, nvidia Coda accelerated applications to work inside a container. It, it has a few steps involved and we’ve pretty much automated all of it with our images.

So as long as you have on your host machine, the Coda driver is installed, the Nvidia driver is installed on the, on the Docker host, then they should just work inside the container.

[00:30:22] Todd Robinson: Yeah, very. Yeah. Very cool. We won’t ask you how you did that, but yes, it can be a challenge. We do have familiarity with this as well, right?

We, we have been through that process to get those things working, but especially it’s nice to be able to build off of someone else’s versus going like, why is that not working? Our, our, our world in OpenStack? There wasn’t as much of that out there publicly available information. And actually, yeah, a couple of our team members have spent a good amount of time making sure that those do work correctly.

And even knowing that some of the. This is all happening right now. So a model may be released and it doesn’t run properly on top of the Intels, but does run on the, on the, the A sorry, the Nvidia. And somebody’s gotta figure out how to move that over so that they can actually run on the CPU. So yeah.

Very cool stuff. Let me think here what maybe you might have some questions for me as well. But I’m thinking through just because the way that you guys have approached this, I think is great. Like it makes total sense that driven by the security needs that you all have, that you’ve had to take this approach.

I’m trying to think of other things that it crosses over into. Do you. This is a left, maybe it’s a left field question, but do you run into things that require confidential computing and that you have had to start down that path to know that, hey, when these sessions are spun up, where are they? I’m kind of curious just ’cause you guys are so much in the security world as well, you’ve mentioned,

[00:31:45] Emrul Islam: well, like I said, some of, some of them environments that customers use Kasm in and they’re air gaps environments.

So they are very confidential to the extent that there’s no outside connectivity at all. So yeah, like our software works in those environments. I would say they’re pretty extreme. Not everyone has requirements like that, but the fact that we can operate like our software can operate in that, and the fact that our cust company knows how to deal with customers with such requirements means that we can help them along.

And that’s. A lot of the reason why I think a, some of the AI companies are talking to us because the requirements they have are sort of bordering on that kind of specialist requirement around sensitivity. The other thing I should mention is yeah, the data now lives in many places. You know, you, we talked about it at the beginning of the call.

We don’t want it on the end user’s laptop. That’s all clear and good. But if you’re a multinational company and you’re, you are dealing with health data from different geographies each of that will have different requirements. You cannot take health data from the European Union and try and process it like it was American healthcare data.

The, the regimes are similar, but not the same, right? Mm-hmm. And you’ve gotta have a parallel structure that follows the regulation of each of those. Well. Again, you know, not everyone, I mean, not every solution can be solved with technology, but we definitely provide the tools to meet those requirements.

So with Kasm you can mount a cloud storage or private storage into the workspace, so it doesn’t need to be stored on the agent server. You can have like a private channel from the agent server into your cloud or your data center where you allow certain bits of data to be access accessed.

That sort of bespoke ability to manage data storage is a big part of at least you know, the data security aspect of it. When you’re dealing with potentially unsanitized data or unredacted data you’ve gotta make sure that you, you’ve got a very narrow path that it travels through and you know who is accessing it, what they’re doing with it, and that they can’t do anything else.

And then Kasm really just makes it easy to do that sort of requirement.

[00:33:51] Todd Robinson: Yeah, that’s pretty interesting. And so I, it may, and you don’t have to do any more live demos but you could maybe speak to it, but what are some of the o overseeing tooling or, or screens, dashboards that you all would have that help?

Understand, I guess, what is the utilization of the platform by the different team members when what’s all, what’s all going on? Like, you had showed off a little bit about like, hey, here’s how you know what you’re connecting to and you’re using this GPU, et cetera. Mm-hmm. But I would imagine there must be also some tooling to help the kind of the administrator understand the utilization of the system, what’s going on.

Is that data inadvertently doing things it shouldn’t be.

[00:34:29] Emrul Islam: Yeah. So we provide a lot of metrics and logging, and we have like a little at the moment is a, is a bit of a dashboard that’s just showing usage. But going back to your question of, of how do we know what to work on well, one of the things we’re working on is a, a sort of more in detailed, view of what’s going on inside these sessions, right? So we already have things like screen recording so you can record user sessions if you want. Cool. Yeah. But a lot of people don’t want that. They just wanna know what’s going on without actually having to watch a video of it. So Yes. So that’s a lot of oversight.

[00:35:01] Todd Robinson: People don’t Yeah, that’s, nobody likes that.

[00:35:04] Emrul Islam: Exactly. But we have, we have, because it’s Docker, remember at the end of the day, Docker is not, virtual machine technology. It’s not a hypervisor. It’s, it’s nice, it’s a isolation technology. And because it’s running on the same kernel as the host, you can run a lot of standard Linux security tools on these agents and they’ll all just work.

For example, CrowdStrike Sentinel one those agents, you run them on the host and they’ll protect all of the containers. And you’ll get visibility into what’s happening in all the containers. So there’s a big part of the Kasm story, which is, it just works ’cause it’s Linux. We don’t need to do anything special, any of the existing tooling.

Maybe with some config changes, we’ll happily play along.

[00:35:42] Todd Robinson: Yeah. Yeah. Very cool. So though I, I could imagine that it must be pretty fun, I guess, to be the creators. And, and give me, so each one of those, when you have that dashboard up there and I’m, I’m picking it. There’s a bunch of Ubuntu on there.

That is the workspace or what? Yes. Give me the nomenclature. Okay. Yeah. So yeah, workspace. So there,

[00:36:01] Emrul Islam: there, there’s like, there’s a registry you can go see. There’s hundreds of apps that are available. And when you pick one, it downloads, it downloads the Docker image onto your machine onto your agent server.

And when you launch the workspace, it starts the docker container and starts streaming it to you, right? Mm-hmm. I didn’t mention, but we even support things like game card pass through. We’re gonna add support for Smart card pass through, ’cause some people are using smart card authentication. And you can stream YouTube on it.

I mean. You wouldn’t want to necessarily do it, but I come from a background where I tried doing that in some other VDI solutions and it just broke. So I was really wowed when I came to Kasm. I’m like, this just, just works really well.

[00:36:45] Todd Robinson: Yeah. Yeah. No, it’s pretty, yeah, it’s pretty brilliant how all that, that has come together.

So if you, if you don’t mind that, who do you, who, who buys this from you guys? Like what, who is not the company necessarily, but is it is it a CTO that’s coming in first? Is some infrastructure person, is it the head of development or, or a research, a head of like a principal engineer of a research team?

Or like, who actually says, I like this stuff. We, we should buy this.

[00:37:11] Emrul Islam: Yeah, it, it varies a lot. If a customer or an organization is doing a VDI replacement project, then it will be usually some member of the tech team that’s tasked with looking at what options there are, right? So the VDI space is not that big.

There’s a set number of vendors and, and we’re obviously one of the, the upstarts. So they come to us, they’ll try it out and then they’ll be talking to us. That conversation obviously will then evolve and there’ll be a, a bigger stakeholders meeting and discussion around that. So that’s one way we engage.

Other times, you know, it’s just a tech team and they, they just want a developer workspace for their 10 engineers and they just want a 10 user license, no problem. And e even within Kasm, we sort of eat our own dog food. So all of our own development happens through Kasm. So we launch mm-hmm.

A Kasm, workspace that’s got all of our development tooling, it’s got our GitLab, everything connected all the authentication is done and we do all of our work through that.

[00:38:10] Todd Robinson: Got it, got it. Alright. So, but much, much of it is at this stage is the teams looking for a better VDI. Right. And, and that they’re very Okay.

And then, so I, I guess when you guys are thinking about the, on the AI side, you’re then, are you taking this back into the current customer base or are you looking for new customers this way? Or eventually, obviously new customers, but probably in the beginning going back to existing customers to get the ideas of what they need.

[00:38:41] Emrul Islam: It is a bit of both. I think a lot of our existing customers are looking at AI and they’re asking the one question, which is, how do we even use this in our organization? Like they, they’re not AI developers, they’re not gonna be producing models, they’re consumers. But how do they do that, you know, with all the other things that they have to do.

So we mention what we have and like the, anything LLM image, et cetera, they, they quite like that sort of stuff and it allows them to get started with AI in their organizations very easily. Right. The other thing we have is like a partnership with Oracle. So you can basically get a Kasm desktop in the cloud with all the AI tooling available.

And it’s just you don’t need to, I wouldn’t say it’s serverless, but it’s pretty much. Just the desktop in the cloud with AI tooling. So depending on what people are looking for, it will impact who’s coming to us and, and what their, what their approach is gonna be. But we have ways of engaging with, with everyone.

[00:39:39] Todd Robinson: Yeah. Yeah. We got, I got it. Yeah. No, I think that the, again, this, for me, it was gonna, I knew it was gonna be very interesting ’cause it was such a different space from where, from where we exist. And, and I think in our world obviously en enabling a system like this can be run on OpenMetal. And for us, I think we’re, we’re finding it is.

People are working it out. So much of it right now is, is that people are working it out as to how they’re gonna implement this. Yeah. And and yeah, for us, I think we’re just excited to be able to start offering things where it’s, again, you can do your and it’s mostly inference workloads. Whenever we, I think, start to cross over into thinking in the training world, these are, these tend to be very very specific purpose use cases and not for long periods of time.

I think in our case, and probably in your case, is that these are gonna be, I. In, in utilization, not in training. And I think, I think that’s what you were, you guys were just saying what would be, do you, do you guys know like what are, what are the most popular workspaces? It sounds like the one that you showed off right there was that everything. LLM. Is that what you called that one?

[00:40:48] Emrul Islam: Yeah. Oh, that’s a new one. Anything LLM that’s a new one. We just added that last week. Honestly, I, I think that there’s no telemetry in our products, so we don’t actually know what people are doing with it. Okay. But we do, I mean, these are all images that are hosted on Docker hub.

So we can look at which images are being pulled the most. Right. And across Kasm, I think we’ve had over a hundred million pulls via images which is kind of staggering. And from there you can see, we know Yeah. Audacity seems to be quite popular. Who does, who’d have thought?

[00:41:20] Todd Robinson: Yeah. Yeah, I was just gonna say, I think I don’t know.

We had a, we had a link up there a little bit ago, which I think is what, I think that’s what we were linking. No, we were actually, no, it wasn’t a Docker hub. I don’t, I forget what was up there, but yeah, I’m just looking it up right now just to check it. But we’re in that same note. Sometimes it this like our, what are the most popular things, but it does sound like some telemetry might be in order.

You might you might poke the team there, but knowing already that, like you said, if you’re getting that level of polls and you can see the specific ones that’s already given you some pretty good ideas.

[00:41:52] Emrul Islam: Yeah. Yeah. And I think just to circle back, Todd, I think where, where we can work together, Kasm and OpenMetal is essentially that you guys are far more technical.

So customers come to you, like you say, they wanna know tokens per second. Okay? But they come to us with as much softer set of problems, right? How do we achieve something, right? And I think the. The, the, the challenge for companies like us is we have to be able to work together and help customers sort of find a joint solution, right?

So we are not the experts on the hardware side of, you know ai, right? But we would definitely talk to companies like you to help our customers get to whatever objective they’ve got, and we can certainly work together on the technical side of security and also on the actual operational side of security to make sure that the customer actually gets the solution that they need.

[00:42:46] Todd Robinson: If you don’t mind, then, yeah, this might, this may help it out. Hopefully it’s of interest too, is give me a little bit more information about the VDI I side. I mean, we, we’ve been kind of bouncing back and forth between what, what they’re actually doing, but what, yeah. I guess gimme a little more insight into the VDI tech. ’cause I think, like I said, it’s powering things you didn’t think possible, like you could actually stream YouTube properly. Yeah. Maybe give, yeah, give us a little background on the tech, the, like a little more grassroots Kasm, VDI tech. And then, and what what do you need typically on the cloud side to make sure that that part of it’s going?

[00:43:16] Emrul Islam: Yeah, sure. So the, the, the core tech is, is open source. You can go to our GitHub and you can actually get the source code for the VDI streaming solution. It’s all based on VNC. But we’ve modified the VNC protocol slightly to basically support a few optimizations that we need to do. And those optimizations are essentially allowing us to stream with lower latency and higher higher definition.

And then that pretty standard software is actually just installed into every docker image that’s in our workspaces. So when you start a docker. Instance we start the workspace that VNC server within that container is started as part of the boot cycle and it starts streaming to the end user.

So the tech is, it seems very like effective at what it does. It seems really great, but it’s also pretty standardized. And anyone it amazes me that anyone can go and look at it, use it. You can even run those workspaces just with box timer Docker. You don’t need to install anything from Kasm.

You can pull an image and you can run it on your local Docker instance and you can visit it in your web browser and actually access that desktop.

[00:44:24] Todd Robinson: Right, right. Okay.

[00:44:26] Emrul Islam: So you think about what you can do with that sort of thing. You know, we, we are playing with technology around process automation, right?

Mm-hmm. So we’ve seen like browser control demos with ai. We’ve all had about robotic process automation where you can basically automate certain tasks. If you can boot into a desktop that spins up in under a second and you can run native apps and point and click it changes the game in terms of testing automation.

You can basically do all sorts of workflows that you couldn’t easily do before, right? And then we haven’t even spoken that much. But we also support Windows, right? ’cause windows in the VDI space is huge, and unfortunately, windows containers do not have any desktop components. So you can’t do that.

So for Windows, we do something a bit different. We use the IDP protocol and we have an agent that runs on the Windows server that’ll basically simulate a session.

[00:45:16] Todd Robinson: Interesting. Yeah. Co cover that a little bit more. I is that, I’m guessing it is a pretty good chunk of the business the window side, but yeah, if you don’t mind, I guess explore that a little bit more and then, and then I know in the last couple minutes I got a couple of the things for you.

[00:45:28] Emrul Islam: Yeah, sure. So I mean the, the, the Windows side Yeah. Is, is a big chunk of the business. Obviously any, any VDI project would be windows centric. What we’re seeing though is we have customers that need access to a Windows box for all sorts of reasons that you wouldn’t have assumed. So we have one customer they need to make available a Windows application to their end users without actually requiring them to install anything.

So they just basically got Kasm. They run their software inside Kasm and they give their users access to it. Simple. That way their users are able to, access windows apps through a browser. Nothing to install. And the com company well our customer runs everything inside themselves.

Just seeing a question on the pricing model. So we didn’t talk about pricing but I think I mentioned Kasm is free for five concurrent users. So you can download it today, you can install it, you can load up 200 users into it if you want. But five can log in at a time, obviously subject to your hardware, being able to support that quarter load.

And then beyond that we offer two pricing schemes. One is per concurrent session or you can have per named user, right? So most companies will, will deal in concurrent users. They’ll say, you know, we need 3000 concurrent users. And they’ll get a license accordingly.

[00:46:50] Todd Robinson: What do you, do you have kind of kits that you’re or recommended as they, as they start down that path to say, Hey, we’re gonna need these con concurrent users. And like you said, the part of it is driven by what’s the infrastructure behind it? Maybe explore that. ’cause it ends up like you’ve kind of got two things you’ve gotta do.

You gotta get them understanding their software, but they also have to have the infrastructure and it’s set up, configured correctly to support a volume like that. But yeah. Is there some kits or some ideas Yeah. That viewers can take?

[00:47:17] Emrul Islam: So we do everything with partners. So we have like a number of hardware partners that will, that we work with, and they will advise the customer on what actual hardware they need and they’ll be responsible for making sure the hardware is compatible.

And at the same time when it comes to cloud environments like Oracle Cloud or AWS GCP, we can advise customers on which mix of hardware that they need. Based on whatever sizing they give us. So we can assist with that. And we also have autoscaling agents so customers that have demand that’s maybe elastic we can accommodate as well.

[00:47:49] Todd Robinson: Got it. Okay. Yeah. Very cool. Sometimes, and we don’t mind, we don’t mind mentioning the various different clouds you might be on here with OpenMetal, but in our world we love seeing workloads go into the cloud. And, and this is clearly some cloud native things because what ends up happening for us is it’s a little bit more about OpenMetal is a, after a, a company has really stabilized infrastructure wise to say they’re cl they’re a cloud native company and they have built up and, and are using this large system.

What they typically find is they hit a tipping point where you’re like, oh, that. You know, it’s great in the cloud that gives you this instant ramp, but the fact is, is once you’re spending, I don’t know, you pick your number a hundred thousand dollars a month, you don’t go from zero to a hundred thousand.

You go from like 95 to one 10 to maybe one 20 or back down and somewhere inside of there. And, and in our space, we facilitate people companies that have hit a tipping point when they say, oh, wow, look, this is, I’m always at a hundred. Boy, that’s, I didn’t realize it was gonna be so expensive, but I am cloud native and I love that I can, I can basically hook these different things together, have my giant object storage, right?

And then I have my compute running and these are all hooked together nicely, kind of like an API first mentality. Hmm. It, yeah. And so for us, i, public cloud is great. In fact, we sell to public cl companies that, that, that turn around and sell the OpenMetal platform out as a public cloud. And of course, so we love that.

There’s definitely a time in place between this tipping point that occurs. Yeah. And so I think we’re, we’re excited to, again, for me it was great to learn additionally what, how you guys work in the, and again, this the place that you have inside of the industry. And again, just that mentality of like, look, we can provide this secure AI approach because we’re kind of, that’s our mentality.

Kasm was born that way, and so you’re carrying that in. So, yeah, I think for me, I think I’m excited. I was trying to think if there was any other things in our last couple minutes here that we wanted to get out there. You got the pricing out. I don’t know if that, I think you, you talked about how it works.

Yeah, yeah. Anything else that you think we didn’t cover?

[00:49:57] Emrul Islam: Well, I can pick up on, on the, the whole, OpenMetal and public cloud. Yeah, we, we definitely, I, I personally have seen that with a number of customers that I’ve worked with in, in various roles. They went a a bit sort of, they, they dived in too much into the public cloud.

And they thought it’s, it, it, you know, when you go into I don’t know, like a snack shop and like a fast food shop, you don’t look at what you’re buying, right? You just buy ’cause it’s cheap. It’s relatively cheap, right? Yeah. And then you realize it’s not that cheap because you put a lot of stuff and you don’t need all of that stuff.

And now the biggest problem you face is you dunno how to unpick the stuff that you bought with what do you actually need, right? It’s so, it’s so intertwined, like, okay, I don’t know how to unpick this. And I think, yeah, there, there is definitely a big push to multi-cloud, a big push to companies bring some BA stuff back into on-prem.

The AI just exacerbates it, right? Because some data really they don’t want in cloud anymore. And it is too sensitive to go anywhere, right? And so working in that climate means you need software and hardware that can bridge between them. So I think OpenMetal’s great. The fact that you can provide cloud style primitives inside customer’s own environment is fantastic.

And then from our perspective, you know, we’re happy, completely self-hosted, completely cloud native. It’s, it’s one of the selling points of our product that it kind of works anywhere. We didn’t architect our software to only work in a public cloud because obviously it wasn’t where it was born.

And that means that we can now help our customers put things where they make sense for them.

[00:51:34] Todd Robinson: Yeah. No, no. I definitely appreciate that. And I do always tell people is you, you should stay in the cloud, I guess probably stop buying so much of the fast food stuff. I haven’t heard that analogy before. I like it,

[00:51:46] Emrul Islam: but it definitely, yeah, s3 buckets is so cheap, you know?

[00:51:49] Todd Robinson: Yeah. And you bought so much and then you’re like, whoa, what do I even do with this? ’cause I bought it. Yeah. But I don’t want to eat it anymore. Exactly. So, yeah. Alright, well, yeah, I think we’re, we’re at the end. I don’t think there’s any outstanding questions that we can get to now.

But yeah, for me, I definitely appreciate the time and it was very educational for me and I kind of, again, like I was saying at the beginning, I appreciate the different perspectives that are out there. And yeah, I’m just glad you guys could join us today. And thank, thank you so much. Think it, we sign off here much.

Yeah, definitely appreciate it. Again maybe I don’t think we have time to flip the OpenMetal on back to, to Kasm, but again, it’s ca it’s ca let me just make sure I’m giving the right names out here. I’m on the Docker Hub one, so yeah. Anyways, guys, check out Kasm Technologies, Kasm works workspaces and and like you said, it’s the five, I think you said GI give.

The last thing is how, how do I get into it really easy?

[00:52:40] Emrul Islam: It’s five up to five concurrent users, five concurrent users. Kasmweb.com you can just go get started and install it on any virtual machine.

[00:52:49] Todd Robinson: Perfect. All right. Appreciate it, Emrul. All. Thank you. Thanks everybody. Thanks for tuning in.