How Viasat uses AWS AI to customize algorithms for satellite customers

I didn't know they can make good morning. I should have put notes as well and thank you for joining us this morning for our session today. How Viasat uses AWS AI to customize algorithms for satellite customers. I'm Andre Kearns, Marketing Director for AWS Aerospace and Satellite. We have a lot that we want to share with you today. So let's jump right in.

Here's the agenda that we're gonna cover today:

First, I'm gonna provide a brief overview of the massive satellite data challenge that the global space industry faces. And I want to do that as context for the role of data science across the industry.

I'll then introduce how AWS is helping our customers to meet these big challenges. Uh and we're doing that by providing innovative solutions tailored to the global space industry.

I'll then introduce our terrific speakers for today. And so we have leaders from AWS and Viasat who will come up but introduce themselves, they'll introduce their role and really uh share how they work together to advance Viasat's satellite communications mission to connect the globe um using AWS and to close things out.

We'll open up the floor to answer your questions. Right. That's the, that's the fun part. And just as a final reminder, we are recording this session right now. And so uh this session will be available to you all on YouTube. I would say within a week.

A recent report by space industry analyst NSR projects that 52 exabytes of space data will be transported by satellite networks by 2029 52 exabytes. Now to put that astronomical number that pun was intended to put that astronomical number into context, 52 exabytes is more than 10 times the amount of words that have ever been spoken by every human being that has ever existed. This massive growth in space data really underscores the growing role of data science in driving the shape of the future of the global space industry.

So um our ability to capture to transmit and to analyze exabytes of space data is really going to unlock amazing innovations that will benefit our society for years and years to come. And so this is our challenge and it's also our opportunity to develop and apply good data science methodology. AI/ML advanced data analytics uh to be able to extract valuable insights from this massive space data and to be able to do it reliably and to be able to do it rapidly in a timely manner and to be able to do it at scale solving big challenges like this is why we created AWS Aerospace and Satellite as a business.

And so every day we work with our customers to help them to reimagine their space systems to automate their data space workflows to accelerate innovation that they're driving and to ultimately improve the outcomes from their space missions. And so we do that by delivering innovative solutions that are really designed for the industry and really designed to lower the traditional barriers associated with space, whether it's latency issues or network bandwidth limitations or traditional lack of access to scalable data infrastructure.

And so here are examples of how we're working with our customers every day to apply cloud based AI/ML to massive amounts of space mission data to enable industry success and to enable mission success.

We have extended edge computing into space and that's allowing our customers to analyze data in orbit, to save time and money. We're automating space data workflows, allowing our customers again to deliver faster time to insights and extract and realize much more value from the data that they're collecting.

We're expanding our connected world with customers like Viasat using services like SageMaker, which you'll hear a little bit more about in this session to detect anomalies and improve satellite communication network reliability.

And finally, we're helping our customers to transform those space systems to optimize planning performance and safety.

And so with that as context, I'd now like to introduce our speakers. So first, I'd like to call up Clare Z Price, who's Customer Solutions Manager for Aerospace and Satellite. And then after she talks, she will introduce Sue Singh Chief Development and Data Officer for Viasat.

And with that over to you Claire.

Thanks Andre.

Hi, everyone. I'm Clara O Price and I'm a Customer Solutions Manager supporting the AWS Aerospace and Satellite vertical. Raise your hand if you know what the CSM does. Ok, good. I thought so. So I thought may be helpful to start with an introduction, my role. So you get why I'm here and how I support customer success.

So the CSM in AWS is a um an AWS investor role that works as part of the account team to provide a consistent and unified support to our customers. Um so um although AWS is a technology company, we understand it takes more than technology to solve a business problem, agree, right? Um it involves identifying and understanding the underlying root cause of the issue, selecting the right technology, building readiness around the people and the process implementing the solution and managing change before the problem is solved in the most efficient manner.

The CSM is a certified solutions architect and a certified project manager that takes this problem solving approach to help our customers removing blockers in every step of their cloud journey.

So how does the CSM do this? Um you may have heard of the empty chair concept in AWS, which may sound a bit cheesy but I tell you it works. Um so when we have internal meetings, um you know, we put an empty chair like this in the room uh to remind of, um to remind us of the customer, remind their real needs and to guide us to make the best decision for them. The CSM works very closely with the customer and services, the voice of the customer in these meetings when the customer is not present.

Um so, and you may also have heard that um 90% of AWS products, features and services come from customer request, which is true. The CSM is the enabler of the feedback loop um to ensure um that customer voice and customer requests are captured properly documented. Um prioritized and finally products into our offering road map.

Viasat is a strategic customer of AWS and a very important customer which has support via a um you know, it's a vertically integrated company that operates on on a massive scale from manufacturing their own satellite communications equipment to operating their, you know, geostationary satellite fleets in space and offering satellite communication services such as broadband and mobile directly to customers all over the world.

Viasat is one of the first movers to the cloud in the in the space industry. And they operate mainly on AWS.

The culture of Viasat, it's similar to what we have in AWS. Um it's a company that's made of builders where everyone shares the similar mindset of continuous innovation and improvement. I've had the pleasure to support a few bright minds within Viasat. Um to name a few, the um Director of Machine Learning Sean Austin, the Senior Director of Enterprise Architecture Mayo Sla, right. And uh Senior Engineer Manager, Matt Butler, um Software Engineer, Mark Sel and the list goes on,

I helped Viasat recently to build a real time data and machine learning pipeline that ingests terabytes of telemetry data generated by satellites that are orbiting more than 35,000 kilometers away from earth on a daily basis to perform advanced network analytics um in response to Viasat's cost cutting initiatives.

Um my team uh with the support of our, our, you know, our account executive Hates Magnusson and our ANS leadership. Uh we were able to, you know, engage in and proactively engage in a series of activities to help Viasat to reduce 40% of the overall cloud spend to date.

In addition, and in parallel to um you know, the machine learning pipeline and cost optimization effort, we're also supporting Viasat in designing and building a fault tolerant um cloud architecture. with the goal to improve their service availability and resiliency using services such as AWS Backup and Resilience Hub.

I'm going to um play a small video that will help you to understand the Viasat business better.

So do you guys notice anything in the video. What did you see in the video? You see the customers? It's not about Viasat, it is all about their customers and I can testify that that is something we have in similarity as well is that we're extremely customer obsessed Viasat is no exception and they are passionate and proud of what they're doing, the changes they're making, um how they're unlocking opportunities for every person living on earth no matter where they live.

So, the focus of today's presentation is about Viasat's success story in building a machine ops pipeline using AWS services such as S3, Kinesis, EMR and SageMaker etc.

Um and the empty chair here today represents the Director of Machine Learning Sean Austin. Um from my experience of working with Sean, he really impressed me upon as someone that's an expert, has an expert understanding of technology trends and the appropriate application of building um innovative solutions that drive customer value, that's impactful, that's measurable.

Um so the um Sean's passion is contagious. Actually, he inspired the account team to unite with multiple service teams to uh deliver above and beyond his measuring his benchmarking statistics. Um so we can make this happen.

Um unfortunately, Sean can't make it on stage today. But for a very good reason though his son Lucas Austin um is uh representing the United States of America to compete in the ISA World Junior Surfing Championship. In Brazil. As we're speaking, we're very happy and proud for Lucas and we wish him the best of luck in winning the championship for team USA.

Uh while we um were missing Sean, we had a blessing to invite um a very impressive, another very impressive person in my opinion, uh to share the story with this audience, uh which is Sue Singh.

Sue is the Chief Data and Development Officer of Viasat who is a visionary, a strategist, a certified search or rescue diver and a great storyteller. Please join me to welcome Sue.

Thank you. Thank you, Claire. That, that was awesome.

So as you can see, we work super close with AWS and our partnership spans many years now, just out of curiosity to tune the dialogue that we have this this morning. How many of you represent sort of space or satellite industry? How many of you are familiar with that? Quite, quite a few? Ok. That's good.

Um so again, this is not a deeply technical talk. I think the key takeaways here are the challenges we faced and you're going to face many challenges. Machine learning operations was one of our challenges that we'll talk about here that we partnered with AWS very, very closely on. We didn't have the DNA in the company. We did a lot of machine learning. Don't get me wrong, but I do think that the challenges as you will see them were pretty unique and and you know, we um obviously leaned in on AWS as SageMaker was becoming popular to try and build a platform out of it for the company, you know, and a common blueprint, not everybody can leverage the entire organization.

So a lot of the takeaway is about those challenges and you may relate to some of those in your organizations. Um and then our approach to solving those and what we saw as a result or benefits from those approaches. So I think that's really, I would say is the key takeaway in this presentation.

Um I'll just quickly give you a bit more of who we are, what we do, how we do it. Um and then we'll get into our use of AWS, what does the footprint look like? Um and then we'll get into the machine learning use cases. As I said, we'll talk a lot about the machine learning operational challenges we had um how we sort of built a, a platform around SageMaker to solve those challenges and then where we're headed next. So that's kind of the outline and hopefully we have plenty of time at the end.

Um so I a we're a little over $4 billion in revenue um which recently acquired Inmarsat. So this represents the combined view of the company. Um we have about 19 satellites in operations globally. These are geosynchronous satellites and we have 11 other in construction and we cover all the way from Ka band, Ku band, S band. And then two of these that are under construction, the orbit satellites, highly elliptical orbit satellites. So a pretty decent fleet that covers the entire globe and provides billions of connections all around the world.

We, we, we're in existence for nearly 40 years now, believe me or not. Um and we have evolved as a company um in terms of where we are today, about 7000 employees, 90 offices, globally, 90 plus i would say. So that's really who we are.

And if you look at what we do, um we serve certain key markets, right? And mobility markets is predominantly one that we cover pretty well. So commercial aviation business, aviation, energy services, enterprise, land, mobile, et cetera, and then government is the other substantial business we have where we cover military department of defense, et cetera, tactical operations.

Um and then finally, the fixed broadband market, which is home, consumer internet communities and businesses. So that's what we do.

Now, this is important to understand as we relate to some of the machine learning work we done we did with AWS.

We are a company that thrives on innovation so small, but you know, a pretty innovative company, we constantly try and push the boundaries of the satellite bandwidth that we can obtain from the technology.

We really optimize the use of that bandwidth across the globe. So a really hard problem. And, that end to end performance improvement is what we try and pride ourselves on.

As Claire was mentioning, we're a vertically integrated company from a legacy by that standpoint, we're vertically integrated company, which means we try and optimize every layer that we touch all the way from the satellites we build.

We try and push the bandwidth boundaries of what it can do for us. To a network or service that operates largely on cloud environments.

Our terminals which optimizes performance on the ground through the services that come through satellite and then ground stations as well where cost efficiency matters with the global footprint and then customer operations and touching everything that touches our customers, right?

So we try to optimize each of these layers. As I said, a lot of this infrastructure is in cloud. We've been using AWS for a while and, you know, believe me or not, our recent sort of stat on the number of AWS different distinct AWS services we use was over 100 and 50.

Which means we pretty much use all of them. If you think about the classical services, we use a lot of EC2 consumption, a lot of RDS consumption, a lot of storage consumption, ton of data transfer every month.

And then we are slowly shifting over to other services now. So we're moving a lot of our applications, containerizing them, starting to use Elastic Container Service.

Converting a lot of our workflows with Lambdas. SageMaker usage is picking up. So we're moving to some of the new services as AWS is introducing them.

And speaking of AWS and speaking of data and speaking of some of the optimization I talked about that sort of is at the crux of our, our business.

We have a lot of machine learning use cases, you know, 125 data scientists across the business. And if you look at the sort of the lower two layers, that's pretty common across all businesses, regardless of which specific industry the business operates in.

So these are more internal operations or external operations type use cases, quality of service matters a lot to us. A lot goes on there a bundled it all in one, but it's an extensive area where a lot of different models play.

And I think cyber security is another area given we do a lot of government business that's a really important facet for us. For both our ground networks and space networks.

And that's kind of the the other two layers that I'm talking about here. There's a lot of machine learning work that goes on in, in every aspect of our ground network, in every aspect of our space based operations, we're operating a hybrid network. A global network and try to optimize a lot across that network.

So you see, Andre talked about anomaly detection as an important use case in the satellite industry. We're trying to push the boundaries there. We're trying to do automated root cause analysis with real time data.

Claire talked about some of the recent efforts that we have and in turning our data platforms to as real time as they possibly can. And one of the drivers is for, for building a platform like that is to do automated remediations where we don't have to intervene as issues happen.

I'm not even talking about the application layer here, such as say, for example, we provide in flight connectivity, you know, it's an entire stack, a ground stack. It has, you know, front end portals and advertising and media, television, et cetera. I'm not talking about any of that.

They all have machine learning use cases. So the data platform that we built on AWS is important and the machine learning we've done I think is good but along the way, I think the teams have diverged a lot.

Historically, if you look at the approaches, the teams have that they were all quite different, right? They were all doing different things. There was no standard tool. I think there were a ton of difficulty. I'll talk about that next, there was no consistent framework, there was no consistent inference.

And every team sort of ran in different directions. And I don't think that was ideal. Right. So that's where we were, we do a lot of machine learning. The key challenge is really were to take those, take those algorithms, take those proof of concepts and production them and production them is a way in a way that, you know, you don't need to touch them.

And so really what we wanted was a standardized environment, a common blueprint for the company that didn't exist. We need a solution that seamlessly scales. And then we wanted these processes to be repeatable.

As you saw in previous slides, you know, what the teams were doing were largely ad hoc, right? They sometimes you wouldn't be able to repeat them, automate them properly. So it created a lot of bottlenecks, particularly, you know, we were really good in building models, still are pretty good in building models.

I think where we weren't as good was to take those models into production in a way that becomes touchless and everything works the way they're supposed to work. So those, those were sort of where we wanted to go those those dimensions that I mentioned, but we had custom approaches.

When I say custom approaches, people were selecting tools differently in different, you know, business units. I think there was a lot of what I call as open source based approaches that were semi optimal etc, our environments particularly when it came to training time or hyper parameter optimization took a long time, right?

So these environments were even if scalable for data were quite scalable for MLOps repeatable MLOps. I'm continually talking about that because that's the key challenge we're trying to highlight here. Not that we didn't do a good machine learning, not that we didn't build good models, it was just that those models weren't giving enough value to the organization.

And then finally, I think a lot of these tools that different business units had then selected often lacked advanced features, right? So there was, we just didn't keep up with the pace of innovation, given the fragmented approaches across the company.

So those were really the key challenges and that's where we weren't getting enough value from all the data science efforts that we had across the company.

So what did we do? Obviously, there was a need to accelerate a lot in the company, right? So we clearly recognize that that's necessary. And we didn't really have, I think we were pretty good as far as data platforms and operating a large scale platform is concerned. I don't think we have the MLOps DNA.

So what we did was to accelerate it, we partnered with AWS, right? And, three specific teams, as you can see here, the Solutions Architecture team. And I think that provided a lot of really good input. The training team to do a really good training on, on SageMaker and then the ML approaches and the Pro Services team.

So we created a joint tiger team which I think was really good because a lot of that knowledge stayed with us. As we approached some of the MVP use cases and I'll talk about that SageMaker best practices, training around the professional services. It all sort of came in as part of the plan we put together for accelerating ML transformation.

We did design. So SageMaker has been evolving as you guys all know, we did design a serverless platform that we can call our ML platform based on SageMaker. And I'll show you glimpses of how we do different things there.

And then as I said, our challenges were training times, you know, time that we were spending in hyper parameter optimization as multiple variants of models came through the shop, which one to pick etc, those were really the key challenges. So our focus was to solve those.

The handoffs essentially not that specific. I think the specific steps were pretty good. It's just that we're breaking down in the handoffs and it happened from data science all the way or data scientists building models, trying things out all the way into actually serving production traffic.

And I think the 0.5 is also important. You know, when you do a transformation like this, when you do a change like this, I think it's important to have a big use case behind it. To drive that use case through the transformation through the process to, to really sort of test it, does it work, does it work for one of my big use cases? Is it going to work for everything else?

And you know, impressively, we did this entire thing in about 14 weeks from start to finish. So let me give you a bit more. So it's a summarized plan, not a detailed plan, but we spent the first few weeks really understanding the technology, the architecture that was necessary for us based on our requirements with the data lab and, and really took this concept of industrializing ML across the organization.

So that was the focus, right? How do we spread this blueprint across the organization? What is it that we need to build? That satisfies everyone. And we eventually finalized a plan, finalized the solution architecture and then, as I said, took that MVP for most of the most of the period.

And, with this MVP, the real challenges were it's a very complicated model. It's, it's in our fixed broadband business. The real challenges there were that, you know, we just weren't updating them fast enough, we weren't updating that that model often we weren't applying all the additional data sources that we had in the company since that thing was introduced.

And so, there were a lot of reasons for why it wasn't being up to date, wasn't being kept up to date, etc. So we refaced the entire code using lambda, as I said, you know, lambda is something we've started to use pretty extensively.

We factored the entire code base using lambda, implemented the central platform for that use case and I'll show you some results at the end in terms of what it looks like now having gone through this process. So that's, that's really what the plan was to accelerate this journey and drive for standardization.

And that blueprint that I'm talking about. As I said, training, people often ignore this the the importance of, you know, infusing new knowledge in the company when that doesn't exist. And so SageMaker, of course, we didn't really have a lot of deep knowledge in the company on it.

And we were, as I said, really good in building models, really poor in production them. And so training was important. As you can see, there was a lot of focus on a first part was an online training. But the the real alongside trainings were all about MLOps and SMOps, which is where we were focused on.

And these trainings did help a lot in that acceleration in, in creating that knowledge body in the company which still exists. And now we have, you know, Claire talk about certification so that, so we have certified engineers now on this, on this.

You know, obviously you've seen a lot of talks here last year on SageMaker and everything that goes in there. Unified web portal, a consolidated environment with all kinds of things. You know, your data wrangler, you have facilities to do training your facilities to understand model drifts your facilities to understand bias.

So it's a comprehensive environment. But there are a few things that not a lot of people know deeply about that it's customizable. And so one of the things we did with SageMaker was created a custom kernel that can bring our Jupyter notebook environments, our sort of R scripts. Our PySpark attachments are sort of APIs not quite APIs these were libraries that can hook up to our data sources in SageMaker.

So this SageMaker now looks like a Viasat SageMaker customized for things that are I would say unique to our environment and our systems and how we've done things historically, right? So that, that, that that aspect of SageMaker I think became important for us.

And I did want to highlight that not a lot of people talk about that. Moving on. So this, this is an example I talked about creating blueprints.

Um so this is one of those blueprints for model training pipeline. And as you can see, I don't know if it's a bit hard to see if you follow the numbers.

Um so this is this is an example where we have one of the Derek data science accounts and you can see the activities that happen within the data science accounts how it connects to our code repository.

Um and then, and so those are some of the customizations in the platform that I'm talking about and kind of goes through the normal cycle of training. Um the sage maker pipeline as you see in the middle from that data science account.

Um and then we make extensive use and I'll explain that in a bit of another small service that doesn't get often talked about, which is eventbridge.

Um I don't know how many of you are familiar with that, but it's a pretty nifty little service to build event driven applications to make extensive use of that.

Um and I'll show you how we orchestrate this platform given the requirements we had in the beginning and that eventually triggers that event, eventbridge eventually triggers the whole life cycle flow that handover that i was talking about. And so hence, it's important, small little service.

I didn't want to highlight that we, we, we had breakages in the process and i think we want to fix those.

Um and so that's a little service that actually comes to rescue along with things we've done with sage maker, of course.

Um so maybe enough on this.

Um let me, let me talk about the philosophy or the approach. I think this is, this is important.

Um and, and this is something we did, which i think is unique.

Um so we divided up the platform in few environments um depending on where the breakages happened, right? There was a data science account, this is where predominantly people test things out, look at data validate it, you know, actually build a model, but they don't do more than that over there.

Um and so it has hooks into few, few of our sort of systems but not the overall system, not the overall application.

Um and you can have a single vpc, you can have multiple, however, you want to set it up and then we have a second sort of core account in the platform which is for dev and q a tasks and then they run in separate bp cs.

And if you think about that, that then starts to integrate with other environments so that you can test the system end to end.

Um and so those integrations come prebuilt from the platform.

Um a lot of that repeatable work was sort of we try to take care of all that repeatable work um in that sage maker based environment through this account, orchestration and eventbridge.

Um and then finally, the pro account, of course, pra and non pro separate cs but the somewhat isolated and really the intention was to get to that carrier grade deployment of the models that we missed before. And this flow of how we, um, you know, brought the models into the operational world helped us.

Um, and, um, if, you know, i talked about the mvp, um i'm, i'm gonna tell you about some of the results we saw as we took that through this process.

Um, it's a very large model.

Um some wi s had people here must be familiar with that and the smile is complicated. it's also old, the code is really old.

Um and so we really wanted that given that it was contentious at that time in the company about its state, its value. we want to take that through the process.

And so obviously, we worked, as i said, part of the plan, the 11 weeks we spent with aws was mostly on refactoring this code, right?

So we moved the environment, first of all to the sage maker based platform i just talked about um and then all the training pipelines were shifted over to the sage maker pipeline and i think our training time was down from like 20 plus hours to minutes.

Obviously, this is a serverless platform and scales pretty well. we had a lot of limitations in our prior deployment of that model.

Um so it was a huge, huge gain just in that, that, that training step and, and hyper parameter optimization step.

And then the multi-account orchestration i just talked about, we actually took this mbp through that setup to test things out and to see if it's onerous, if it's not onerous.

And we have specific ways of provisioning accounts.

Um again, been doing this for a long time. so there are security principles we apply on top of things that company cares about, et cetera, et cetera, right?

And then there was complete observable as to what was going on with this model through cloud watch and, and, and other means.

And then finally, we didn't really have a very good blue, blue green sort of deployment philosophy with this model. And i think we do have now, right?

So now we barely touch this.

Um it updates itself since we have done this work. i think we've updated this 30 plus times already.

Um or we had difficulty updating it once in six months. so, so this is the real result

um on that large model that, you know, most people advise that know about and just not giving you a name for it, but they all know about it.

Um and so the results have been good, the results have been really good.

So we understood where the breakages were in the process. As we went through the workshop, we we found a way to sort of stop those, those breakages

um and, and link them so that the the ma life cycle flows through the process through the orchestration through the platform we design and you can see the results, the results prove themselves that, that we did fix the key issues in the company, right?

It's not that we were bad data scientists, it's just that we were, we were having trouble managing, maintaining our models and then that's just one area, you know, you saw the gamma diffuse cases that we have.

Um so, you know, a bit more sort of generalized results, right?

Um that was the specific mvp, which was important to prove a bit more generalized results.

I think, you know, we now have a standard and scalable environment that, that we can truly sort of say that, you know, we can do a lot of machine learning and a lot of this repeatable tasks around putting these models in production are taken care of by the platform, the integrations, the authentication, et cetera, et cetera, integrating with our data sources, code repositories, model registry.

When when do you register something in in the model registry, et cetera, those are all kind of taken care of.

So a lot of dirty work has been done. it's scalable serverless and then the the templates like one of the template examples i showed you these blueprints have now been created, right?

So we store them.

Um and then these are standard across our all our businesses. you saw the different markets we serve, we can apply these templates across all those markets as far as data science use cases are concerned.

Um what else is worth calling out?

Yeah, i i talked about this already.

So we now have, you know, knowledge in the company about, about sage maker, about ms, you know, dev ops practices extended into machine learning operations, et cetera, right?

So a lot of a lot of best practices emerged as we did this work.

And i think we feel comfortable that we now have the dna to take our miles production quickly.

And then um you know, platform god created, which was important ah this platform.

Um there was 11 other thing i didn't talk about.

We have, we have also very um i would say diverse kind of scientists. some only want to code in our, some only want to touch tensor flow, some only want to work with py torch.

I don't know why that is the case, but that is the case in our company.

Um and i think this platform that caters to all of them.

So that that was another point of contention in, in, in our approach to building a platform was that

um that everybody had these different ideas about what it needs to do and what it needs to look like.

So we think that we have a harmonized solution where you know the framework you work on doesn't matter, the language you code in, doesn't matter, et cetera.

And then the last bit of it is that, you know, given the number of staff, data scientists we have in the company and the operational staff, the the well, they are also i think data engineers or machine learning engineers, but we call them ms or the guys who take care of this platform, there are only four of them for the entire company.

And so i think that ratio is good. i think it's very good.

Um i don't know

elsewhere, you know, what, what your companies look like or, or what your ratios look like in terms of ms to date.

But i think it's a good number that we've achieved.

So we, we're happy about that.

And then finally, you know,

um i think, i think this this process, this transformation has been, has been useful and successful.

Um and we have more work to do, of course,

um as i said, you know, satellite networks, uh the the bandwidth

um that these networks offer now they're all increasing.

I mean, if you think that some of our satellites, you know, in hundreds of gigabit per second ranges, the newer ones support a terabyte per second.

So things are changing really rapidly. we're pushing the boundary on, you know, the bandwidth and the efficiencies, but we need to keep the network

um and this hybrid overlay moving as well and, and react at the same speed.

So i think changing our platforms, you know, batch inferences where they happen.

Um all, all of that to real time is critical.

Um given, given how this industry is evolving.

So, you know, for lack of a better word, i'm just calling it live ml but that's what we need to do, right?

The, the root cause analysis which is automated, the remediations i talked about that is automated.

A lot of that requires a very real time data platform underneath.

Um and so we need to do that and, and the machine learning and our models and our ms practices need to evolve with that.

So we're working on that.

Um i don't think we still have conquered the large scale ml pieces, you know, the the classical sort of machine learning several years ago, if you guys are familiar with, you know, the parameter server based approaches, et cetera where you know, you're doing approximation.

But since the model is trained on such large data set, it's good enough, probably better than other approaches, complete approaches.

Um we haven't cracked that code yet.

So kind of call it big ml et cetera.

I think we need to get better at that.

Um and then of course, you know, no conference, no talk is complete these days without talking about generative a i.

So i think we have, we have done early work from the last year or so on.

Developer productivity type use cases with, with generative a i, i think they've been pretty successful, very well received in the technical communities in the company, but i think we need to take a lot of those and move them into field use cases, production use cases

um which requires us to get good at prompt engineering, which will not delay. right.

So i think that's yet another dna that needs to come into the company.

And so we've started some work on that, some of the announcements you saw yesterday. i think it was exciting.

Um you know, bedrock and q, et cetera.

So we'll start looking into those.

Um and that's the beauty of our aws partnership as we're thinking about a lot of these things have approaches, you know, aws is coming up with these services.

So we'll start looking into those and see the approaches that we have or the approach ws has put together is, is better.

So we'll work on those.

Um with that, i'm going to hand it back to andre.

All right, sid, i want to thank you for sharing your powerful story of how you're using aws to standardize and really scale uh an ml ops platform out to all the data scientists across your company.

And also it's really, it was amazing to hear how you've been able to cut the time of your deployments from hours down to minutes uh through your approach.

So that's really powerful.

And claire, thanks for getting us uh kicked off with this session uh and focused on what we're focused on from an aws perspective every day and that's the customer.

So let's give our speakers another round of applause.

So to close things out, we want to share these follow up resources for you uh these qr codes uh to learn more about the story uh that was shared today uh and discover other use cases that may be of uh may be relevant to you all.

And so with that, like to say thank you.

And now we like to open up the floor for questions.