Fuzzball & Nvidia NIM: Speech-to-text Webinar

Um a lot less subject matter expertise in terms of you know David's level of involvement in these things. But I'll let David introduce himself in terms of his background and and where where he fits. Yeah, thanks Jim. Um yeah, my name is Dave Godlo and I used to be a scientist, a neuroscientist working at the National Institutes of Health. Um and in that capacity I became interested in high performance computing while I was there and I joined um the Bowolf team which is the uh that's the large kind of cluster that is available to uh scientists at the NIH. I joined that team as a staff scientist.

I worked there for several years and I was around during kind of the big bang of containers within HPC about 10 years ago and I've been kind of around that community and a member of that community ever since. Awesome. See, I told you all PhD. I'm just some lowly marketing guy. I don't know. Um, but I look forward to it, buddy. Thank you for joining. Um, and we really look forward to this as as always with you, buddy. Um, you know, you you talk about HBC, you talk about Bowolf and all these things and I think anybody who's been in the that industry for a while kind of knows what that is.

And yeah, I think there was a fundamental shift as we kind of moved to containers just in gerally even outside of HPC. I guess it was like 10. Oh my gosh, it's 10 years ago now. 10 years ago for the HPC shift, I think even maybe 12 or 13. It was 15, I think. Like, yeah, it's crazy. Um, you know, is that kind of like an HPC 2.0 thing? I mean is that kind of is is that the world we live in David? I mean yeah it it totally uh changed the way that people think about applications and HPC and you know how how users install them and run them and and you know it's just it's a normal tool now.

It's like people talk about Singularity and Appainer the way they talk about using you know slurm or using the shell or SSH or something. It's just something you use in HPC. Yeah. It's like in my world it was like well there was this cool Docker thing and now it's just like uh Yeah. Docker comp you you know a doc you know what I mean so like kind of a yeah just I think it's just part of the world so um you know David uh you know we we are have been heavily involved with you know open source projects like Warwolf and Aptainer um but I

think our kind of like you know the flagship that we look at is fuzball uh and what we're going to talk about a little bit here today is fuzzball so David just to give people context you know what is fuzball it's the like the how about the most generic question I could ever ask you Yeah, I like to answer that question usually by telling a short story and so I'll do that really quickly. Um, so we we just got done talking about 10 years ago.

We had HPC users, their DevOps friends were using Docker and they were like, "Hey, we got to start using this with an HPC and it wasn't really a good fit for a variety of reasons." Um, but we decided instead of like trying to fit this tool in and make it work where it doesn't really belong, let's create a container solution that does work with HPC. That was Singularity uh which became Appainer. Now we're having the same thing. We have users that are coming to us and saying, "Oh, I don't just want containers. I want container orchestration, right? And I want you to install Kubernetes in HPC." And you know, it's it's you can do it.

It's uh it's not necessarily the greatest fit. It's built for kind of a different purpose. And so what we're saying is let's purpose build for you for HPC a container orchestration solution which will allow you to you know maximize the utilization of your HPC resources. And that's exactly what fuzzball is, right? And I think it's interesting too, David, the way it's architected in that, you know, I think, you know, my first misunderstanding of fuzzball, and I say my first was that, oh, it's all Kubernetes. It's actually not, right? Like we're using Kubernetes in a way that's like intelligent in the context of HPC, correct? Like it's Can you just talk a little bit about that architectural choice?

Yeah. Yeah. Yeah. So, in order to understand that, you have to understand that there's kind of two different ways to run compute. There's services and there's jobs. And so essentially services are these things that are going to run in the background. They're going to run forever and they're going to interact with stuff. You know, you're going to make requests of them and they're going to spit things back. Jobs are like they have a start and an end. They usually take some input and produce some output. Um they they you know process files, they create graphics, they do stuff and then they get they get the stuff done and they're done.

So, Kubernetes is really good for services and it's al also really good for what we call microser architecture where you take a bunch of services and you put them together like Legos to make an application. Um, and so lots of applications are built on top of Kubernetes as these with this microservices architecture including Fuzball. So the the fuzzball orchestrate application itself is this big microservices architecture that takes a bunch of little microservices and glues them together like Legos and makes the application work. And that's a that's a really great way to create an application these days. That's right. However, when it when Fuzball reaches out and talks to a compute node and runs a user's job, that doesn't really have anything to do with Kubernetes.

Now you're running underneath of the fuzzball orchestration system itself which is orchestrating your stuff for you and is not really got anything to do with Kubernetes. Yeah. So we're using services to orchestrate the jobs basically right the compute and all the stuff that's going on which I think is the exact right way to do this um because of the complexity of like the control. Uh it's almost kind of like well we're using there's a control plane in Kubernetes but it's kind of like orchestrates this kind of larger control plane that's helping us kind of run sometimes internally we refer to it exactly as that. Yeah. Exactly.

That's right. And I think it's I it's a I think it's a genius architecture. Uh one that I had not seen interesting enough because I think once people get like really deep into Kubernetes like well let's just use it for everything. Yeah. You know let's use it for what it's good at. You know what I mean? And I think it's a I think it's the right way to do it. But you you you David you talk about jobs. Um, what are the jobs that people are running? I know we package a lot of these things and we talk about them like what just give me an example.

We're going to talk about one in particular today, but like what are they? Yeah. So, um, I mean you can think of anything that that is that is HPC focused. So, we're talking about things like um you know big fluid dynamics codes that need um that need MPI to run. So they got to run across multiple different nodes and utilize this big massive cluster in order to you know calculate a result. We're talking about weather modeling you know we're talking about a lot of uh biomed like um you know doing uh sequencing genomic sequencing aligning you know alignment uh things like that. Yeah. Uh modeling right?

Yeah. Earthquake seismic stuff climate modeling you know weather stuff all sorts of things basically. and and so that's a lot of high performance computing. There's also um sometimes it's called like in the HPC world we call it like the long tail of high performance computing or sometimes it's called high throughput computing. Yep. And what that is is like I've got 10,000 files that I need to just analyze. I need to run this thing on them that takes 10 minutes per file and it's going to take forever if I try to do it on my, you know, laptop. And so instead, I'm just going to get a thousand computers spun up, a thousand nodes, and I'll do 10 on each one of those computers, and then I'll be done, you know, really, really quick.

And so that's another thing that's, you know, parallel. What's going to spin up all those computers, make sure it's all running, distribute the job, blah, blah, blah. Fuzzball. Right. Exactly. And so that's that's kind of where it fits y'all. So if it's, you know, that's a really good example, a really simple one, David, of kind of how that how where fuzzball fits. So So let's talk a little bit about, you know, the topic today. and NIM. Uh I guess first of all like you know we talked about jobs like what is NIM? It's kind of a service like h how does that interact here? Yeah. So I think that we're kind of um we're embarking on a a a little series of webinars where we're kind of highlighting different types of workflows that we can do with Fuzball.

And NIM is a really interesting one because um so a little while ago Nvidia put out this thing. Oh yeah, go ahead. I was going to say what does NIM stand for David? Because Yeah. Yeah. Yeah, we're just be engineers and talking acronyms all day. So, Nvidia put this thing out that they called the Nvidia uh inference services or Nvidia inference microservices. Yeah, I talked a little bit about microservices before. It's a big buzz word, you know, and so essentially it's just a process. A service is just a process that's going to run in the background and take requests. And so what NVIDIA has done as a you know kind of service to the AI community is they've prepackaged a lot of models in containers like inference models that you can use uh through something like Kubernetes.

And so when we say inference models, what we're talking about is like you can um spin up a a server and then you can generate you know um images based on text input for instance or you can you know you can have a chatbot that you interact with or something like that. These are all inference models and NIM is Nvidia prepackaging some of these and containerizing them for you and then giving them to you in a way that you can spin up easily and then interact with. Yeah. Well, yeah. You I wonder why they did that, David. I mean, because it uses GPUs, perhaps. I don't know.

Well, it's such a course. Of course, it's useful to them to do that. It's useful for them to have more people using GPUs. So, yeah, it's super valuable to everybody. Win-win. Yeah. And Yeah. And so, I've talked a lot about like how um you know, fuzzball is great for jobs. So, we're kind of starting off by flipping that script around a little bit and being like, but you can run there's nothing saying that you can't run services in them, too. And so, um even though uh nim was uh created primarily for the idea was to use it with Kubernetes, we wanted to show you today that there's nothing keeping you from using it with fuzzball either.

And so, that's kind of, you know, what we wanted to demonstrate today. Well, let's do it, man. Let's uh stop looking at our faces and start looking at terminal, dude. Absolutely. The less you look at my face, the better. I've got a great great face for radio. All right. So, yeah. Um, so, you know, I just kind of want to show you there might be folks on the on the on on the webinar who who are watching this who have not seen Fuzzball. And so, before we jump into the actual specific use case, I thought it might be good to just kind of give you a little bit of a tour just to give you an idea of what Fuzball looks like.

Right. So, here is the web interface for Fuzball. Um, there's also a command line interface that you can also program against. So, it's really super flexible. But, we're we're looking here at a bunch of different workflows. This is kind of like the landing page when you first jump into Fuzzball. And because I've got a shared account, um, you know, I can go through and I can see all the stuff stuff that I've am running and stuff that I've run in the past and stuff that my colleagues um are running or have run in the past. And this is pretty cool because it's really collaborative, right? So, you can go in here right off the bat and I can say, "Okay, my friend, my uh colleague Brian Fan, I'm gonna shout him out.

Hope hopefully he's not running anything he doesn't want people to see." So, it looks like he's running something that has the X um XFCE desktop in it and it says ANIs. So, I think he's doing something with Ancis. Um but yeah, you could go in and um you know, I'll run I'll look at one of mine. I'll just look at my workflows. I I ran this silly little thing called cows moose the other day and I this could be somebody else's workflow too and I can go in and not not only look at it but I can look at the actual file that was created to run it and I could also open that up in the workflow editor.

We've got this workflow editor that allows you to just drag and drop um to create jobs and then just like link these jobs together to create dependencies and you know give these jobs commands like and you know whatever. And when you when you do these things, what you're actually doing behind the scenes is you're creating this YAML file. And it just makes it super super easy and super collaborative because I could start off with a YAML file that one of my colleagues has. I can just go in and look at it and I can start editing it myself and then I can run it based on that.

So yeah, that's kind of like at a real high level. Um, let me go ahead and delete this one. I don't want to run that one. But yeah, and now if I want to run this after I've created that, all I got to do is hit the little start start workflow. Oh, it says there's an unsaved job. Let me go ahead and save that. Yeah, me uh not sure what it's upset about there. Okay, let me go ahead and run the workflow. I'll call this K2. It's just upset that you're doing a live demo, David. you know that the machines know when it's a live demo and they know to like do things and so knows.

Yeah. And so like without getting too far into the weeds just a highlevel overview of what's happening here. So this instance of fuzball is running on AWS. And so we've got this server that's running like it's serving this this um this web guey and this web UI. I'm not supposed to say guey. That's an old person term. The web UI. Um, and it's it's also serving like the orchestration engine, all this other stuff. And it's waiting for me to hand it a YAML file, which I just did through the web interface. And once I hand it that YAML file, it's going to look at it and say, okay, what jobs do you have?

What are the dependencies between jobs? What resources did you ask for? I'm going to go out to the cloud and spin up, in this case, the cloud. It could also be on a local cluster, but I'm going to go out and either spin up in the cloud or allocate in a cluster the compute resources that you need. I'm going to grab the containers for those jobs, land those on the resources. I'm going to start the commands inside them, and then I'm going to, you know, get all the dependencies between them uh up and running. And that's what's happening behind the scenes here. And then when we're done, now we're done.

Everything's already run in the AWS environment. it's spinning those resources back down so you're not stuck paying for them in your account. Of course, on prem it would just release those resources so that other users can use them for other jobs. So that's that's kind of all the at a very high level what's happening behind the scenes here. That's a big deal. Just efficient use of plat of compute, right? Like massive big deal. Cool. And yeah, so I it's a moose because it's a cow with some antlers. Um yeah. Yeah. If we go back to the I changed the definition here a little bit. I said I wanted a moose.

See that David? You're you're um your dad got tired of cows. Really? I miss you and your dad jokes. So that's at a really high level what fuzzball is all about. It's very and I was trying to highlight too that it's very collaborative. And I think that um you know one of the things we like to talk about today is that we're taking collaboration to the next level with Fuzball. So you know it used to be that you first log into Fuzball, you see all these workflows. Um if one of your colleagues has created a workflow that's kind of generic and kind of fits your needs, you could look at that workflow and you could edit it to be more what you needed.

Well, we started thinking about that and we started saying,"What if CIQ built for you already, a bunch of really generic workflows for a bunch of different popular applications and what if those shipped with Fuzball so that you could automatically have access to them already as soon as you as soon as you install Fuzball?" And so we did that and we're calling that and we're we continue to do that and we're calling that the workflow catalog. And so now when you first get fuzzball, it installs already with all these different, you know, pre-created workflows for all these different popular HPC applications. And we're creating these workflows in such a way that they should be push button.

They should run just from the start. You shouldn't have to really change anything about them. You should just press the button and they go. And then, you know, this is the pattern that most scientists like, you know, so then you get something that works already and you just kind of tweak it and edit it. Maybe you change a path so it it accesses your data instead of the sample data. Or, you know, maybe you say, "Okay, well, I want to change the container so that it uses a little bit, you know, more advanced software and I want to up the the CPUs from four to 16 or something." You just sort of tweak stuff until it, you know, works for your particular workflow.

So the the thing I just showed you um just kind of highlight this. Of course, you know, I work for CIQ. So we got to have a a CIQ application cows, right? Because I created LOL cow. So if we go in here and we we have a look at this. This is the thing I just showed you essentially. So we've got this little um we have this little uh YAML file. It is a a template and then it has a couple of variables in it. one for text and one for the type of cow. And what this allows you to do is when you actually want to run the application, you end up with this this really nice little popup that just says, well, here's the defaults.

You could stick with the defaults or if you want, you could change the defaults. And what I did before is I just changed it to moose because I've got all these different cows I can I can choose from. Uh, let me change it to Stegosaurus instead. And then once I do that, I validate it and I press continue and I press So, so basically I didn't have to do anything as a user, right? I just I just basically changed some text in a box and I click some buttons and now I've got a workflow running. Um, and of course this is a really simple kind of toy workflow, but in a minute we're going to show you a much more advanced workflow um, using once again the NVIDIA's NIM.

So that's great, David. I just just to tie off on the workflow catalog, um, I like when you're working within a team, any anybody can publish to the workflow catalog, right? It isn't just the stuff that we have. have like a team could say here's the thing and then like people could pick up and they could branch that thing. I mean is there a concept of branching or is it just like well there's a new one over there and this guy created it and you know what I mean and you know what I mean is is is that concept here yet? Yes. I'm so glad that you brought that up because I totally whiffed on that and forgot about it.

So, not only do you get like pre-baked workflows that CIQ has created and the CIQ has verified and like I, you know, I should I should have showed you that some of these workflows have like a CIQ badge on them to show that they've already gone through the verification process and these are going to be baked in when you install. Um, but also a lot of these workflows are stuff that we're developing right now that we're just playing with and anybody on your team can go in and create these workflows. Other people can see them, other people can use them, you can branch them. It ties into Git.

So you can have GitHub repos that have these workflows described as template files and then you can just, you know, publish them out to uh to Fuzball through your CI process, your continuous integration process. That's great. So it just works within the whole delivery supply chain you already have, which I think is a big deal. So yeah. All right, let's see what a That's what a Stegosaurus looks like. All right, look at you with a top hat. Cool. All right, so enough silliness. Enough silliness, Jim. That's enough. Let's let's let's talk about let's talk about NIM. So, um using the workflow catalog, you know, I'm going to call out Brian Fan again.

He's the man of the hour today because he he did a lot of work to create uh it wasn't a lot of work, but it was he did he did the work to create um this Nvidia uh inference microservices workflow that I'm going to be demonstrating today. So, he should get the the you know, he should get the props for that. But once again, the same thing I just showed you, but this is, you know, it's a lot more complicated. There's uh a lot more um richness as far as different um things that the user can change. And so if I go ahead and I say, you know, run the application, I'm not going to run it because I already have it running, but um you know, you've got this Docker image.

And in this case, we're going to be looking at a model called Parakeet. And as you sort of said at the beginning, um this is a speechtoext uh model. So the the type of inference we're doing here is somebody's going to be talking and then we're going to be having text printed out somewhere. Um and so this is already, you know, pre-baked, created by the good folks at NVIDIA. And essentially to run this, all I had to do was I you actually you have to sign up to use Nvidia NIM. You got to go and you give them, you know, your email address and stuff and then they okay you as a developer.

It's free and everything, but you got to be registered and then you can create some some keys to be able to access these. So I had to create some keys and I had to change these out with my keys and that was it. And then things ran. And so um I'm not going to run it right now because as you might have seen if you were looking at the workflows, I actually have one down here that is running already. And that's exactly, you know, what what this is is that it's it's one of those um those NIM models. Here's the NIM. I was going to bring this up and talk about it, but um it's one of these models parakeet running.

And so if we look and see what that what's that what that is doing, it's running this this um this inference micros service that I can connect to. And after I connect to it, I can take speech and translate it into text, which if it's still working, of course it is, is what I'm doing right here. So basically everything that we're saying um during you know this this this webinar um is is being transcribed uh not here at my on my local computer but it like my local computer is connected to this AWS instant instance that's running through fuzzball and every time I talk too much it yells at me.

There we go. Um, so it's it's connecting to this uh this instance that's running up on fuzzball and it's transcribing in real time as we talk. You know what we're saying in a terminal. It's awesome. I want to I want to play Look at it's it's catching what I'm saying as well. This is I want to be uh I want to play war games here. Shall we play a game? Uh I'm aging myself. If you know it, you know it. Yeah, of course.

19 1985 war games I think 85 I'm not sure if that's true or not but in that in that general area so so yeah so just to recap to make sure everybody understands what's going on here um so Nvidia created and containerized um something called parakeet which is a trained model which will take your speech and transcribe it um we spun that up within fuzball I was able to spin that up within fuzball and get that running as a service. And then here on my local machine, I've got uh some Python libraries which enable me to ask Parakeet to do things. And one of the things that I can ask it to do is to connect to my local microphone and transcribe everything that I say into that microphone into text.

And so that's exactly what's happening. Now, I could also, if I wanted to, I can stop this. I'm going to go ahead and just control C. Um, and I've got another script in the the command I've got saved here. So I when I was first developing this, not developing, when I was testing out what Brian developed pretty much I one of the first things I did is I recorded a wave file of myself um reading the poem by Lewis Carol Jabberwocky which if you're familiar with that um poem it's it's a lot of gibberish and kind of silliness although it's it's pretty cool because it's um you know it's it it makes sense even though it's gibberish.

So, I recorded that and I said, "Okay, well, what is what is this model going to do with a bunch of gibberish?" And um right now I'm sending that I'm streaming that wave file up to that server and the server is listening to it and it's transcribing what it hears. And this is kind of a cool test, I think, because a lot of this is just gibberish, but it's it's trying to come up with in some cases even like phonetic, you know, just phonetics that will kind of like uh you know, will be similar to what it's hearing. And it does a pretty good job I think.

Yeah, it does a really good job. Uh David, what was I mean so uh it listening to us on the microphone live was that just an endpoint you were connected to. How like I didn't you kind of just brought up the terminal and it was already going like what was the what is it connecting to? Yeah. So I've got another window here in which I am port forwarding um from my local terminal up to this. Okay, great. So, it's just that service. Yeah. And then we could pop into this service if we wanted to and do something like if I can get it and yeah, we can see that it's got a GPU running.

It says it's not doing anything right now, but I'm not sure. So, David, when you were when you were configuring this um in the editor, you're bringing it up, it you know, it asked for a number of cores. Yeah. Which is interesting. um where what h how does it know what that resource pool is, right? Uh do you have to register all those? Like how how do that work? Yeah, that's a that's a great that's a great question. So um so you can ask it for how many? Yeah. Yeah, you can ask it for how many cores you want here, but we actually have um so an administrator is able to set up resource definitions.

Oh, great. And within within um AWS, the way that that looks is is basically you pick which which AMIs you want your users to be able to have access to. And then you set those that list of um AMIs. And then Fuzball resolves that and says, you know, either I can serve that request based on the fact that I've got access to this and that and the other AMIs that I can spin up for you or actually no, I can't do that. And it'll give you an error if it can't do that to, you know, say that's, you know, I can't I can't I can't fulfill that request for you because I don't have anything that's got 500 cores.

Yeah. Right. That's cool. And then yeah, you had GPUs in there as well. So yeah, that that's right. If I that was just I mean yeah, part of this part of the configuration of this particular and in this particular case, I mean this is an inference model, right? And inference, you know, it tends to be typically you don't need too much horsepower to do inference, but if we were training this model, for instance, we might need to allocate a ton of GPUs, right? Yeah, that's cool. That's awesome, David. So, if anybody has any questions, please do ask. I there were a couple that I I I threw into you along the way, David, but um what's the future of this, David?

Where does this go? You know what I mean? Like I you know, like the being able to, you know, tie together things and jobs and services. That's interesting to me. I don't know but like where does this all go do you think? Yeah. So it's really really cool because the way that I described this to begin with is that essentially it's just container orchestration with jobs right and you know that's like okay that's cool whatever but um once you have that once you actually start playing with that and messing around with that you realize you can run a whole HPC cluster with that. That's really all you need to have your users submit pretty much any workflow that they want to run is the ability to complex and diverse too.

Yeah. Yeah. Oh yeah. So, so it it becomes even though that's the the basic idea is it's container orchestration with jobs, it becomes like this this entire way to run your entire cluster and also not only to run your entire cluster but to take your cloud resources and run them as as if that they were an HPC cluster as well. Yeah. So it's look the future of it is to take um you know either you got some some cloud accounts or you got you know a big on-prem cluster sitting there you stick fuzzball on it and suddenly you know you're administering everything that you need to administer and your users are running everything they need to run through this one platform.

Yeah. Yeah. Quite honestly when I first saw it I kind of was like well that's the coolest piece of tech that people should know about. Like because I think I think you know we think about fuzzball in a certain way but and it's so funny you said it's just you know you know it's just the orchestration of some jobs. Oh whatever you said that I go there's a whole group of engineers that said wait a second dude that was a lot. Yeah you know to build that you know and I think it's the utility here uh is is kind of tremendous especially as we start to extend into you know kind of other areas beyond HPC and how do you start to like treat it all together.

I think it gets really really interesting and kind of more like you know not homogeneous to a particular area but more heterogeneous across across different things. Yeah, that's that's the other thing too is that if you if you make HPC more accessible and you kind of bring it out a little bit, right? Like like it's it's it's working in both directions right now. There's a lot of people who are doing stuff that they're kind of starting to realize is HPC, right? And then also if you make HBC is this really useful tool and if you make it more accessible more people are going to just like if you make AI more accessible.

How many people are using AI right now that you never would have thought that it's it's overhyped. Yeah. Totally overhyped David. I don't know who's using it. Right. Yeah. It's everywhere. It's almost like it's if you squint hard enough uh it's all kind of collapsing into one big group of massive huge compute. Like it's it's all pick kind of. It's like I don't know. There was another word like you know performance intense computing, right? It's kind of all that a little bit too, right? Like it's kind of all one world a little bit. I don't know. Yep. Yeah. Everybody needs to use some big chunk of computing resources.

That's right. That's right. So, well, that's great, David. Um, anything else you wanted to show demo wise? I think we I think we covered everything we thought we were going to do, right? Showed you guys a Well, I didn't show you guys a cow. It was just a moose. Um Oh, wow. I mean, I It's the first time I saw a Stegosaurus, so I'm happy. Yeah. Okay. All right. I guess we'll let it slide then this time. Yeah. You know, there were a couple in that list, though. I'm I'm gonna start playing with it. So, yeah. Yeah. Check it out. That's awesome. Well, David, thank you so much.

Like, that was a um that that was solid because it was just like I you know, I like I said, I feel like there's so much cool like fuzzball is just cool. Like, it's really like it is really kind of a different approach. And I think the the more I see it, the more kind of like interesting it gets, I think. And I think that's kind of one of those things with that we see people who adopt it. Uh so um if anybody here really wants to go out and try or they're already using it, they want to talk to David, they want to talk to us about uh what they what they aspire to do um using things, you know, orchestration like this.

It's really interesting. I think another area we didn't talk about, you know, if you're going to orchestrate and move workloads and jobs around to be efficient, uh we can do this. There's a there's a whole feature set within fuzball called federate y'all which allows us to basically like oh this was all single cloud or single environment we can go across multiple environments start to do really interesting we'd be happy to talk to you about that as well uh to actually really kind of start to drive multicloud uh you know implementations of these sort of things underneath the covers and you know you think about David's little flow imagine that going I don't know it does I mean abstract out the underlying compute you know it's amazing so there's lots of as we go.

We we'd love to talk to you more about it. Uh if you're looking for a demo or want to talk about your particular use case, uh please do reach out to us. There's a lot more information about Fuzball solution briefs, webinars, videos. Uh oh gosh, I don't know. You go check out the work workflow catalog. We're happy to talk about that too if there's stuff in there that's interesting you. But uh we'd love to engage with you. So uh David, thank you for doing this. I really really appreciate it. Yeah, thank you very much.

And I think another thing too that we want to make sure that the audience knows is that if there are particular workflows that you are curious about how they were going to run in fuzzball, um, totally reach out to us and we'd be happy to do one of these webinars on it actually and kind of like show you how how it works within fuzzball. Yeah, for everybody. And yeah, I'm sure David will sprinkle in a dad joke or two, but that's cool. Few cows for sure. Few cows. Awesome. All right, buddy. Uh, thank you so much for again for doing this. Uh, always fun talking to you.

Always love doing these with you. And then um everybody, thank you for joining. Uh we hope it was uh valuable to you. Um like I said, if you joined, we'll definitely send a video and a bunch of other stuff to you. But uh thanks to everybody and have a great day. Bye.