And, uh, we have, for the past seven years, we have been working on multiple AI product lines here within Versa.

00:11

So I wanted to take you through couple of, uh, real world examples, uh, of ai, Uh, for those guys, those of you who don't know about versa,

00:17

we are a world leader in SD van. Uh, we have a SS e product lines. So we have more than 22,000 customers.

00:28

We are there in a hundred plus different countries, and we are trusted by some of the world's largest and more, most complex organizations.

00:34

Uh, so let's, uh, so we have a, a unified SE platform, uh, include that includes SSE, sd van, SD lan, et cetera.

00:43

Now, uh, versas a's vision retrospect to AI has been to use, uh, empower the, our customers to build a self-protecting network by using ai, uh, on, uh,

00:53

your networks, on your sec, using your security and using our co-pilots. And let's take a look at how we are doing that.

01:06

So Versa, AI uses three foundational technologies. Uh, we use time series modeling, uh, and anomaly detection modeling for detecting your consuming raw signals from your network.

01:13

So these can be latency, gtar, SLAs, application link rank performances, uh, mosco, et cetera. And we can find anomalies

01:26

and we can predict them, uh, in some time in the future. So you can use those for capacity planning.

01:35
Uh, for example, you can, uh, you can say that if my bandwidth is going to exceed, um, uh, the predicted rate in the next one week

01:40
or next one month, I need to start my capacity planning, uh, effort now, or, uh, some of our customers, uh, so one

01:47
of our customers, uh, have these big ships and we have, uh, our gateway, the gateways deployed on the ship

01:55
and the van interfaces are satellite 5G, 4G, and they use SLA to figure out which part to take. But the ships are always moving.

02:01
So if you make the product reactive that when the SLA goes bad, then I'm going to ship, uh, then I'm going to move the, uh, van path

02:10
that the time has already, uh, uh, happened. So we use predictions of SLA to do the switch before the SLA gets bad

02:18
because now these, uh, network follow a predictive path so we know when to ship, uh, when to, uh, switch the, uh,

02:25
network to a better path, right? So the second thing is we use generative AI in our co-pilots to answer question.

02:32
Um, all the product gets PR pretty complex. There are a lot of features, security, networking. Nobody can answer everything.

02:39
So we have our own, uh, uh, chat bot built on the knowledge base, and it's a functional chat bot.

02:45
And this has been there, uh, previous to the LLMR. So it does all the function calling. It is integrated with all the APIs

02:51
and it's able to convert all your NLP to an actual work that you can do. And the third thing is for AI for security,

02:58
where we have our own custom models, some of them transformer based, some of them, uh, uh, uh, some of them signal based to detect malicious behavior.

03:05
Either it can be somebody's, uh, trying to exfiltrate data from your network, or it can be straight up, uh, malicious JavaScript, uh,

03:14
that is trying to be downloaded there. So using various, uh, versa technologies like API, dp, CASB, uh, remote browser isolation,

03:23

we run ML AI at different junctures all over the place to figure out what's happening. Right Now, we have different product lines where all these product,

03:30
all these models go into. So we have AI security, which goes into our advanced security clouds, data loss loss prevention.

03:42
We have a firewall that detects data exfiltration to Gen ai. So, uh, which, which is, uh, one of our latest products

03:49
where people, when they try to put source code your company source code on chat GP, and try to figure out how to figure, uh, how, uh, like try

03:56
to get some, uh, output out of it. We can print, we can figure out whether they're exfiltrating that, um, we have user and NDT behavioral analysis

04:03
and analytics built around it. And for AI, for networking, we have a product called ani, which is Versa, advanced Network Insights,

04:11
which does all the anomaly detection predictions, um, so forth. And our co-pilot is called vrbo. This is built into, uh, Ani as well as, uh, all our devices.

04:17
It's the brains of troubleshooting. So when one e predicts if something is going to go bad, VRBO is able to troubleshoot how

04:28

to fix when something is going to go back, and then you can, um, use our policy engine to automatically remediate the issue before it happens.

04:37
So we have a full, uh, fully integrated ecosystem which not only predicts, it also knows how to troubleshoot and also knows how to fix it.

04:46
So, uh, for, for all of these foundational models, you have to think, think about multiple things. First thing is data collection. How are we getting the data?

04:59
So for networking, you have to collect, uh, you can't use, uh, train a generic model and deploy into all of these devices.

05:08
So we create models that are very purpose built. So these models would be per customer, per appliance, per network, devices per signal.

05:16
So even for a smaller customer with 300 branches, there will be 5,000 or 6,000 models for them. And we are using the, uh, power of cloud to kind

05:24
of spawn these resources, train the model, and shut those things down, and we pass the savings to our customer.

05:33
So it, it makes it much easier to deploy these, uh, huge models. Then, uh, we have copilots and all the security.

05:38

That's a typical rack pipeline that we have. Um, for, uh, security. We are getting data from, uh, our honeypots

05:45
that are deployed elsewhere. We are getting data from our data brokers that we buy from. We are getting data, uh, data from vendors like virus total,

05:53
where we, uh, we, uh, we pay a lot of money to fix, get the data, then we train the models on top of that.

06:01
So that's the part of data collection and training. And then the deployment is a completely separate, uh, subset, which depends, and it, uh,

06:07
and it depends on what is the use case. So if the data is traversing to the cloud, you can deploy bigger models there.

06:14
If it has to go through the edge, then you have to run smaller models on the edge, and that involves a completely different data set.

06:21
So we work closely with Intel and a MD and Nvidia to try to figure out how to compress the model to smaller size.

06:27
So if you look at a smaller model, let's say, uh, you take a small re-ran model, which is just two gigs in size.

06:35

Now, when you convert it to Onyx, uh, you can load it in. But if you build a straight up Python package with it,

06:41
it's still a 10 gig both of image by going with go, uh, quantizing, an Onyx model, and then reloading it, you can bring the whole size

06:48
of the model as well as the processor down to 500 megs so that it can just, you can ship a local AI

06:57
with your applications as well. Uh, for bigger models where a constant, uh, model training pipeline is installed, uh, we can, uh,

07:03
we can, uh, burst up the cloud and do a lot more processing there. So at, we have to make the decision as to the accuracy

07:11
and performance of each of these models and where, when and where are they deployed and the deployment and, uh, the,

07:19
and the deployment part of it is always going to be, um, uh, you have to have, uh, a pipeline to figure out what, uh,

07:25
to qa the models, how you're testing them when you're deploying them. And the most important is feedback. So the customer always asks, yes, everybody needs ai,

07:34

but what are you doing with it? It's not just for the pretty graphs. You have to consume those alerts.

07:44
You have to make some actionable decisions based on those things. And versa has Versa messaging server, which consume these alerts back

07:49
and sends it over to its branch devices so that you can have policy saying that if the user is malicious, then, uh,

07:57
uh, then kick him out. Or, uh, stop is upload, stop is downloads. And, and, uh, other examples of the ships, uh,

08:04
in tropic steering that we gave value. So that can be done as well. Now this is an example of, uh, one

08:11
of our onei dashboard, which is, uh, versa. Um, versa anomaly, um, versa Advanced Network Insight. And this shows a snapshot of our, uh,

08:20
network signal prediction for a send traffic. So in the middle of the graph, you can see that it's a predict the band, the orange band

08:29
that you see is the predicted traffic, uh, how it's going to be. And the dark line that you see is the actual traffic.

08:38

So you can see that the traffic is following the predicted interval. During the model training process, you can generate, uh,

08:44
capacity planning, uh, alerts if your, um, bandwidth, the predicted bandwidth has already crossed the threshold, or you can generate an anomaly alarm when your traffic

08:52
crosses the predicted threshold. And in either case, you would get all the network behavior, how the net, how your model is behaving,

09:03
how your traffic is going to behave over the different courses of the week. Uh, now let's take a look at

09:10
how a typical infra looks from a versa point of view. So this is a typical infra for a Versa architecture.

09:17
So you have versa on your branch devices, which is connected to the control plane, uh, via Versa controller, and you have a management plane, uh,

09:26
which is Versa director, where you can push the configuration. Now, uh, versa, uh, uses, uh, IPFIX logs, uh, to which are generated towards your, uh, uh,

09:36

towards Versa analytics. So if two branches are communicating with each other, it'll generate, uh, an SLA log that is going to be sent

09:48
to Versa Analytics via the management plane. Versa Analytics will forward it to one e platform via Kafka. One E will generate the insights, uh, from those alerts

09:56
and logs inferences, and it'll push these insights back to the, uh, branch office so that it can make the decision what needs to be done.

10:07
Now let's look at what it takes to provide such, uh, inference and, uh, constant training process, right? So the blue, uh, box that you see, the blue cloud

10:17
that you see is a typical public cloud. So currently we support GCP and AWS, it could be any cloud.

10:27
Uh, now the logs that are generated by the analytics cluster, which is on-prem, it's sent via Kafka, uh, to onei.

10:34
And Onei will send it, uh, multiplex these logs out to your MLI inference engine, which can be a spark cluster

10:43
or it can be a GPU cluster and it can send, and it'll be sending it to, uh, your UEB cluster,

10:49

which is a graph MLAI, uh, uh, powerhouse, right? And once that is done, uh, those insights will be stored back on your data warehouse

10:56
as well as it'll be sent to the notification services. Now, this can be consumed by our APIs, it can be consumed

11:05
by your CM O dashboards, or you can just look at the ui. Uh, the insights and aggregates that are stored are further used

11:12
to train the model asynchronously or like over the course of, uh, every week or every couple of hours based on the model.

11:20
And once the model is trained, it's deployed back into the inference engine so that your real time inference is always, uh, fresh.

11:27
The model doesn't deviate or there's no model, uh, drift happening there. Right? And finally, uh, this, uh, whenever you have the model, it'll, uh,

11:34
you can still see the data in any of the front ends that we have. That includes Versa director and Versa Concerto, right?

11:44
So that's all I had. If you have any questions, please do stop by our, uh, um, uh, booth. Thanks a lot. Thank you.

11:52