Aviz Network Copilot with Thomas Scheibe

We're gonna kick off as a slight presentation for 10 minutes and then we're gonna go into the demo. Uh, I do this because last time I presented he was like, Thomas, you need to at least talk about what you guys do before you demo. There we go.

We're learning with that. Let's start with that and then we're gonna get in 10 minutes going. Sure.

You, we we're a software company, a software networking company. Uh, we're really going after this interesting space of networking. We believe what happened in the server world a long time ago.

Hardware and software separates will happen in networking. We have a software stack. We're supporting customers like you or companies you're working for, where the software second networking, um, we were founded in 2019.

We're passed round a funding that wasn't 21. Uh, we actually have sales presence and presence in four different location or countries. Us, India, Japan, the latest one in Singapore.

Uh, 90 employees, good products, good customers. I'm not allowed to name but you can ask me afterwards when we're off camera. Um, and we have four core products.

What is the other interesting piece about it? We're a networking software company. We have a lot of our partners that actually hardware partners and that's on purpose because you have to deploy the software with hardware if you're networking, right?

Whether that's a switch, whether that's a server, uh, and you see this and you see you're looking at investors. Some of those actually are investors. So enough about the company.

Uh, four products in this space. And really it's around the networking stack and software. If you think, Sue, what you actually really want and use today, when you buy a piece of equipment that has hardware and what is the software piece?

It's the os. We truly believe that moves to Sonic. And I'm not gonna here and tell you why, but we can have a discussion about it separately.

Uh, we truly believe customers need what they have to do. Some capabilities around uh, managing fabrics, managing deployments, whether it's pushing configuration, whether it's monitoring, uh, fabric manager. We truly believe all of you have somewhere or the other, some kind of observability where you actually tap production traffic.

And then looking not just at the header but looking actually actually payload and into correlation. Whether this is metadata extraction, deep packet inspection, a lot of interesting use cases. P cap lately a lot of arguments about how to use P cap actually more efficiently.

We have a product that space, that's the observability. What we're actually talking about today is number one, the network co-pilot. Um, we do have a trademark.

The other companies I just saw earlier today, they used the term networking co-pilot. I probably need to give them a call. Maybe we talk about 8 million.

Um, that's what we're gonna talk about. What networking co-pilot does, and I will get into it, is really going after the problem. And I have been there because I worked for 20 years for a very large networking company talking to a lot of customers.

Some of you in the room, the problem we're going after is today, most of the time what you have, you have a whole set of data island. Uh, you have what I call a click ops. You click on dashboards that are static most of the time and trying to figure out how to piece together data in different data islands.

It's a problem. A lot of customers tell us saying, Hey, wouldn't it be nice if I can use a native language interface? It's a price surprise to get to that data and correlate faster.

Uh, and that's really what we're going after. Hold the data based on your input. Make the network co-pilot your co-pilot.

You as the user gonna be the pilot. It's not the network. Co-pilot takes over the network.

Co-pilot gathers the data for you and knows where to go based on the questions you ask. Gets you the data, makes recommendation, allows you to make decision faster. Clearly very good for automating repetitive tasks.

And we're gonna go through this in the demo, but I can tell you some of the feedback. I quote a person, I'm not getting a name. I dare the months and actually the week in the quarter was to beside when I have to do a compliance report and get all the data from all my different devices to create a report.

Perfect use case for this. Uh, and then the last one, I think everyone looks and saying, and I'm not here to tell you that's one, it's a co-pilot. We are not for self towering networks.

At least I don't believe it's gonna be anytime soon. What I do believe is, and some of you that probably played around was what you can do online around the large language models. They're pretty good in giving you recommendations.

You still need to make a decision, but they're pretty good to giving you options that you then decide to go on. And this is really where this shines and we're gonna go through some of this. So that's what co-pilot is.

What is it not, it's not another data lake. Very, very important. 'cause the first question I always get, so you guys just wanna replace data lake?

No, the answer is not. It's your data. You have, we're getting Data Island rich together.

It's not a magic tool. Other question I always get when we talk about these question answer tools, how did you train this? The answer is you can't train on operational network data that changes every minute, every second, right?

It's your data. It's not public, there's no training. What you do use the tool for is how to know, based on your questions, how to get to the right data.

We're gonna go through this. It's not a plain canvas. People like talking about canvases and painting.

This is not what it's, it's actually product. We'll go through it and it's also not the way we build and it's very important. It's not another way for us to charge a customer more or ionized.

So the way we build it and um, coming here, it's not yet another assistant that you can easily run linking on the backend to uh, chat GPT, open AI or anything else. It's really a piece of software as a private AI platform that you deploy in your own environment. Very, very important.

Uh, it's not tied to a specific hardware right now because I guarantee you actually I know much as guarantee. You have seen all the announcements everyone will have that is a network space, will have an AI assistant. Um, what we built here is, is hardware and neutral.

It will work with different oss, it will work with different hardware platforms. It will with in different environments, which is pretty standard and pretty much all the networks I have seen in an enterprise. Uh, what's important is we're not just the AI LLM company.

We are using this as a tool to actually address network corporation challenges, right? It's not about us. Our, our gravity point is not, Hey, we know how to use LLMs.

Our gravity point is we know how to build networking software, network software stacks and how to answer questions around network problems. And that's really where copilot shines. 'cause you want the ability to do this.

Uh, it's a supportive product, very, very important. It's not a prototype. And you will see we're still in the work.

You will see there are a lot of things we can do more. We are getting a lot of feedback, but it's a product. It's very important.

It's not something we just tie together as a prototype within, uh, two days. And then you can run, you can do this. Actually we have a lot of people walking up to say, I can do this.

And I say, yeah, you can. But when it comes to deploying an enterprise, you want an enterprise software product to support it, right? Uh, it is a private AI platform.

Very, very important piece. So when we started this, we looked at this and saying, Hey, should we just point to a cloud model and then feed the data Clear feedback from everyone I talked to was, don't do that. I can't, my data has to stay here.

I don't want to go through security audits and explain why the data is good or not good where I send it. Very important the way this works, data stays local. You control, you control what goes into the platform.

You can turn a platform off and data doesn't go anywhere. You don't train anybody else's LLM. It's your instant of an LLM that runs local.

Very important piece. And then we way we build it, it's modular. Uh, and you work at, I will show you why that's important.

So how does it actually work? Now, co-pilot is, as you expect, as the co-pilot says, it's a, it's an chat interface. You can ask questions and you get answers.

But what does it actually do under the hood? The LLM or choice of LLM is used to translate your question into based on what you ask, into figuring out where to get the data from, which will be called data connectors to your existing data sources. And then as a set of agents we built and we're adding over time more.

And actually what exposes you, you can add more is how to figure out what to do with that data and then answer actually the question that you ask. And then the LLM may see is really good in summarizing and presenting the answer, right? And the reason why this is very important is because I guess this question again, how do you train the answers?

You don't train the LLM. There's some tuning you have to do based on the question you asked where to go. But the data is really your data.

You pull it out of your existing data sources and then you use agents to actually stitch it together, figure out what to do with it to answer the question properly. So the last one I kind of made the comment earlier now is actually trademark. Uh, and I always say this because I'm in awe when I joined the company, I would've sought some of the large network company owned us.

Surprise, surprise. I don't, Good question. Yeah.

What's the value proposition for customer? You, you're just providing a chat interface to ask questions, but then all of the heavy lifting of training the model, collecting the data in a data lake has to be done by the customer. No.

How does it help an enterprise customer hold Off till after the demo? That's probably much easier at that point. But the short version is the value proposition is pulling together the data that you have in different tools and answer a question right at that point.

Instead of having to click into different tools and trying to find, oh, where do I see packet drop on this device? Where do I see my ticket summarization of a service request in a different tool. You can literally just ask the question, give me the data for this device based on what I know is in that ticket service ticket, right?

Today, you either have to build scripts to do this or you have to say, Hey, ticket request, tell me what device. Log into my whatever tool you have today with your CMDB, whatever that is, and say, now show me data for the data. And this tool, you literally can just say, Hey, give me the information based on what I see.

Open tickets are follow up question. That's where do I use? It's much more intuitive, it's much faster.

Um, and quite frankly, you don't need data scientists. Very important piece. Okay.

But very good question. I always have said, I always got the question, how do you train? We are not training on the data.

The training is around what the use cases are, what questions you want to answer. You need to train what data sources to connect.