Intel Corporation (INTC) Earnings Call Transcript & Summary

Good morning, good afternoon and good evening and welcome to Tech Decoded, Intel's technical webinar series for software developers. Thanks for tuning into today's episode: Modeling, Audience Insights, How Act-On used AI to Find More Customers. I'd like to introduce today's speakers. Syed Ahmed is a Senior Vice President of AI Engineering and Ops at Act-On Software, a marketing automation provider, helping marketers create dynamic automatic programs and campaigns. Syed leads a team on the tech side to make sure that innovations continue to move the needle for marketers and contribute to the responsible development of AI and data science technologies. Sid Kulkarni is a Vice President and General Manager of Data and AI platforms at Intel. Sid drives the overall AI strategy and leads the development of enterprise AI apps to reduce the time to solution and bridge AI skill gaps for developers. Syed and Sid will be joined today by our Q&A moderator, Kevin Ta, an AI software solutions engineer at Intel, Kevin works with customers to develop and accelerate their AI workloads using performance optimizations provided by Intel's artificial intelligence and data analytics product line. My name is Austin Webb, and I will be your host. A few notes about this platform we are using. You can post questions at any time in the Q&A window. Our presenters will be responding in real time there, and they will also go over some questions at the very end, time permitting. You can access the biographies of our speakers in the top left-hand column. The following section contains downloadable resources, including a copy of today's presentation. This section also contains links to our event calendars and our One API academic program. Across the bottom of the screen, you see a series of icons. This will give you access to the slides, Q&A, survey and closed captioning. You can also resize all of these windows to best fit your view and if you hover over the video box, close captioning will be available there as well. If you have any issues with the presentation, sound, slides or other, please note it in the Q&A and a producer will help you out. This weber is being recorded and a replay will be made available immediately after we conclude. You can access this by using your original log-in link, and we will also e-mail out a link as well. And with that, Syed, Sid. I'll hand it off to you.

Thank you, Austin. Good morning, good afternoon, and good evening, everyone. Thank you for this opportunity to present some of the technology details. First, I'll be talking about overall software that Intel actually creates, particularly in the realm of AI. Then I'll be specifically speaking about the platform on which Act-On application is actually built. I will start reviewing this diagram from the bottom here. You see the set of processors that Intel actually creates. But besides creating the silicon, Intel also develops a large amount of software. Most of this software is actually available in the open source domain available for everyone to use. So as depicted here, these processors actually go in cloud and enterprise as well as in client workstations also are deployed on the edge. A layer on top of that, what you see is called as the oneAPI layer that is the acceleration layer that is silicon specific that Intel develops and is also available in the open source domain. Now the layer on the top is divided in 3 parts, roughly mapping to the AI life cycle milestones. So if you look at the first one, that is the engineering data or the data preparation, some of the software that Intel makes available is Modin, SciPy, NumPy, Numba, Apache Spark as well as Pandas. So some of the software is already available in the open source domain, but Intel particularly accelerates it to run it most efficiently on the Intel hardware. When it comes to creating models, Intel supports popular frameworks such as XGBoost, PyTorch, TensorFlow, Scikit-learn, and so on. And to optimize and deploy for the inference, Intel supports -- Intel created OpenVINO as well as ONNX and Intel Neural Compressor, which we call as INC. Now this software as well as hardware is available on the Intel Developer Cloud, which is on the right-hand side. And there is another Intel operating system, what we call it as an ML operating system, which is in terms of cnvrg.io. And of course, we have a very close collaboration with Hugging Face, which I'm sure you are very familiar with. With that, I will go to the next slide. Now in this slide, I will cover the platform on which the Act-On application was built. Now I will start reviewing this slide from the top. So if you look at the top, the platform is built in a very extensible way. What is meant by an extensible platform is customers who will use this platform can very easily change the data sets, can very easily change the underlying models, and can also very easily change some of the frameworks. For example, say, for the deployment framework to suit their organizational needs. All of this is built as a cloud-native and delivered as cloud native containers, which obviously work on both cloud as well as on-prem. The scalability is one aspect we have particularly focused on, which enables you to scale both scale up as well as scale out. And the most important thing we are focused on is we have made it very easy to use and provide the best possible out-of-box experience. So I'll review some of the advantages of using some of these components. So these advantages are actually divided in 2 parts. The first one is about the developer productivity and second one is about the performance. So in terms of developer productivity, we are focused on reducing the time to solution and also reducing the number of data scientists that are required to create an AI application. In terms of performance, I've touched upon this in the previous slide, which is, in this platform, we have assembled a set of most optimized libraries and the consistent version of frameworks that are guaranteed to work with each other. These set of references are the vertical apps, what we call them sometimes, are enclosed in what we call as open platform for enterprise AI or OPEA for short. Most recently, we have announced it in the Intel Vision Conference, and more details are included in the links that are provided at the end of this seminar. All of these references are providing an easy data UI to very easily do the data prep as well as do some of their deployment work. That's the next layer that you see. It is also integrated with auto feature engineering for the automated feature engineering functions. Now I will review the 4 buckets or 4 vertical AI reference apps. The first one in the customer service or the Gen AI. So in Gen AI, we currently support about 5 apps and the links will be provided at the end of this. The first one is chat Q&A, CodeGen, Document Summarization, Search Q&A as well as Visual Q&A. The second vertical that we are supporting is the anomaly detection. I sometimes call it vertical apps or the vertical buckets. So initially, we had released 5 textures and 10 objects in 1 data set. But now we are extending it to marbling perfection, surface crack, defective solar cells. And like I mentioned, these are built in a very extensible fashion. What this means is, although these come with 4 embedded data sets, for example, for the anomaly detection, the customers can very easily add their data sets, either by using the easy data, which lets you do very easy core prep or the data prep, they can also change the underlying models. Say, for example, if you actually substitute this with your own data set. And then if you see the accuracy to be lower than what you expect, you could very easily change the underlying model to attain the accuracy that you want. Let's go to the next one, which is the Disease & Drug, which is in the Health & Life Sciences or the HLS vertical. So the apps we support here are lung disease prediction; blood-based disease prediction, which is, for example, blood cancer; diabetic retinopathy; sepsis detection; as well as breast cancer. So in the initial version, we had the breast cancer available, best breast cancer prediction. Now we have extended it with our underlying extensible framework to some of these other use cases also. The fourth one is the fraud detection. So in terms of fraud detection, initially, we had released the credit card fraud detection. A link of this is actually available at the end of this webinar. But now we are extending it to Airbnb fraud detection as well as airline loyalty program fraud detection. And like I said, all these are built with very extensible underlying frameworks. So this can be extended to use different data sets or different models or different serving frameworks of your choice. I will quickly review the bottom 2 layers as well, which I covered in the previous slide. The next layer in the software productivity layer, which is the Intel optimized ecosystem of Hugging Face, Lang Chain as well as what we call as ITREX. And the layer below that is the performance layer, which is a Intel-optimized, PyTorch, DeepSpeed, Ray framework as well as IPEX. or Intel TensorFlow and Intel PyTorch. So with this, I will hand over to Syed.

Thanks, Sid, and thanks, Austin. Good morning, good afternoon, and good evening, [indiscernible] great to see you all. In the next few minutes, I would like to share and demonstrate the power of Intel's tools that helped us in solving a key problem our customer faced leveraging AI, right? So before we get there, let me talk about Act-On. So Act-On is a provider of marketing automation software. Marketing automation software essentially is a system of record for marketers like CRM is for the salespeople, right? So marketers use this day in and day out. And our vision for the platform is to make it the most smartest, easiest to use marketing automation platform, and we have been pushing the boundaries. We are a SaaS -- flexible SaaS platform based on 3 foundational elements: easy to use and easy to value. Data and intelligence. This is where we collaborated with Intel to help marketers make data-driven decisions. And the openness, being open platform, where our customers are able to easily move the data in and out of the platform to leverage the data and insight. And as you can see here, we have over 4,000 customers and some of the recognizable brands out there. And we have customers worldwide for -- and we are also ISO and HIPAA compliant as well. So typically, what we see is somebody work like a CMO or some working in CMO's organization, purchases our software and deploys it and uses it. What a marketer is trying to do is they go through what is called a funnel. At the top of the funnel, they're trying to drive awareness. Essentially, what they're trying to do is make the prospects aware that there is a solution to their problem. Right? And in the middle of the funnel, they do a bunch of activities, which is called the consideration phase, where they're trying to realize that their product is the solution to their problem, right? And then the last one is, at the bottom of the funnel, is the conversion where they're helping their prospects make an informed decision and then help close the deal. Once the deal is closed, we also help with nurture onboarding and nurturing as well. If you see how the data flows in and out, the platform at the bottom figure here, we generally connect to a CRM and most popular CRMs out there. We help pull in their contact data. We help them create segments, which is the problem that we're trying to solve here, and then reach to the target audience and using the right channel season and the right content. So segmentation is essentially the process of creating audiences. What that means is it allows a marketer to target the right contact with the right content at the right time, using the right channels. What involves in segmentation is 3 broad categories of data sets. One is -- first one is called Firmographic data set. Essentially, a contact's job title, the company that they work for, the industry that they're in, the revenue, the company size and so on and so forth. And the behaviors are the integral part of this as well. What you have is open clicks, people have submitted forms, attended a webinar like this, downloaded a white paper. And then the third broad category is [ computer board ] scores, which is essentially engagement, right? So what you end up typically is about 20 to 40 different attributes to choose from. And this lends to the marketers tending to do a lot of guess work and coming up with permutations and combinations. The way to do this is if you -- you can drive effectiveness of the campaign by creating better target audiences. So what we see in practice is when we looked across our data sets, what we found of the segments that were created and used, only 27% of the segments were using the behavior radar, the data that is generated based on the activities. Most of them, as you can see in this example here, tend to use very, very simple attributes. In this case, a financial services from targeting CPAs and tax advisers is just using a firm type and salesperson. We would like them to use more behavioral segmentation such as attributes like recency of e-mail, their engagement on the website and their interest. So we turn to Intel to help solve this problem. And we had a few goals in mind and as I said, we want to make -- help marketers make data decisions, help discovered new segments that they wouldn't have thought of and kick start their marketing campaign process and as well as continue to push the envelope as part of our mission and vision. We kicked off the project in March of last year with Intel and Act-On team working together to understand the available data set, right? The goal was to bring to market a solution using Intel's modular AI tools, and within a short span of time. In our case, it was 6 months. As you can see, as you can see here at the top of the -- there are -- this graph here, the activity is done by Act-On's team. And the bottom, the activities done by Intel's team. Let me jump into the architecture side of things. So here's the big picture and what we encountered was a data set that contained both numerical and categorical data. So working with Intel team, we decided to use an unsupervised machine learning algorithm, it's called K-Modes, which is variation of K-Means. And in terms of the architecture, the bottom layer, what you see is Intel's acceleration layer extension of Scikit-learn and in terms of distribution for -- of Modin to help with data import and handling. And in the middle layer, what you see is unsupervised machine learning pipeline from ingest to deployment. And at the top layer, what you see is a customized pipeline for Act-On with visualization and clustering. And we had to use ChatGPT as well on top of this, and I will delve into it a while in the next few slides. And what is -- and this, the figure here shows the output of our screen shot of how this shows up in the platform share that -- and the demo as well. So in terms of the challenges, what we're facing here is we were getting a lot of results out of this unsupervised machine learning algorithm, and to the order of over 50 a day. And some of them were very, very similar in nature. You can -- as you can see here, an example, you have a variation of the same output but with additional parameters added to it. And I said -- and also what we saw, some didn't quite make sense. As you can see in this example, the time range picked here for this recommendation was somewhere way out of range, essentially over 2 years. So to solve this problem, we try to build an algorithm internally using scoring technique using the important behaviors that were available to us and giving prominence to important behaviors. That didn't quite work out. So we said, okay, why don't we turn to -- this to a large language model and see if we can get the results that we want. So what we did was took the output of the clustering algorithm and sent this to ChatGPT. And as you can see on the left here is a prompt that is sent to ChatGPT. What we ask is usually 10% of the output that was generated, we say. In this case, the prompt is 8 -- pick 8 segments because we had 80 results generated. But if we -- if the algorithm -- if the model generated 50, we would say pick 5 of them, right? And we sent it to ChatGPT, and we get the results back. And we also asked it to explain why it picked the segment, and I'll show that to you in the platform. All right. In terms of what is available to the customer, we provide lots of levers within the platform for the customer to pick and choose the attributes that -- features that go into the model. And we show segments that are generated out of this, and we show them in a few categories, I'll show that to you as well. And the user can either accept the segment or reject it and then disclose it into providing feedback to the model. And this model is run on a schedule on a daily basis. And we have had the opportunity to get feedback from the customer. The feedback has been very, very positive. And as you can see here, there are a few responses from our customers. I'll take a moment before I move on to the next slide. All right. Let's jump in to the demo. I'll share my screen. Okay. So what you're seeing here is Act-On's live production environment. And we have multiple parts of the platform. The part of the platform we're interested is the place where our customers store their contact data. And in this case, what we have done is we provide our customers with the valuable information right up here in front. As we -- as you can see here, there are 26 new recommend audiences, and we can jump right in here. Okay. So here are the results of the model. Before I jump into that, let me see what the levers that are available. There are 3 different levers available. They can run this model on all the contacts or segment of those contacts, you can select segments as well. Or -- and then you can select what kind of data goes into the model. And you can pick the title, lead source and stuff like that. These are what we call the profile attributes. And then most interesting for this one is these behavioral attributes. As you can see here, percentage of e-mails sent and thus the contact open and how the rate changed for the opened as well as the clicks. So this is the behavioral aspect of the data that is available. And then you can pick the time line as well. You can go back and look at all the behaviors over all time. In this case, we have about over 2.5 years of data or you can go and pick 30 days of data. And then these are the results of the model. And what you can see here is an engaged segment. What this means is these are the contacts that have been engaging on a consistent basis, and they are in the business service industry. And we also surface what percentiles they are in. In this case, they belong to 99.2% and 99.4%, which is between 38% and 75%. So we wanted to give our customers a percentile so that they can see if they're top of their segment, top of their -- the spectrum or the bottom of the spectrum. So -- and let me see if I can find some more -- so also what you can see here is why this segment was picked, and this is an explanation coming from ChatGPT. I can either create a segment from or I can discard. In this case, I'm choosing to discard. And this is asking the user to provide feedback. Why I am picking -- why am I rejecting this. And I can say I've already have a segment like this, and this goes in -- back -- there you go. Demo, right? So let me create a segment here and then go -- there you go. All right. So for this demo, I had pre-cooked some segments before. And as you can see, the segment that I picked had generated about 687 contacts, and I can go in and see -- use this segment in various aspects of the marketing. And in this case, what we have done is we have used this segment in what is called a nurture campaign, and I will give you an example of how that nurture -- what that nurturing campaign looks like. There you go. So this is what the marketer does typically when they're creating what is called a journey. They create these steps and allow -- this allows them to target and reach the audiences and create journeys that will help them drive more engagement. So that's pretty much it. Over to you, Sid, and Kevin.

Great. Yes. Thank you, Sid and Syed, for a great presentation. Let me check. So we're not seeing any questions. I'll give it a few seconds. If anyone has any questions, feel free to ask it in the chat. If there's no questions, I suppose we can -- we can close out a bit earlier. Yes. Thank you, everyone, for tuning in, and thank you, Sid and Syed, for a great presentation once again. And with that, I'll turn it back to Austin to close us out.

Thank you, Syed, Sid and Kevin, for the webinar today. If you want to watch a replay, you can at any time using your attendee link. We will also e-mail out a replay as well. A quick reminder to please complete and submit the short survey. This will give you a chance to shape the series, tell us what you want to hear about and what we can do better. It will also automatically pop up right after we conclude. Thanks so much for joining us today, and we'll see you next time.