This Week in Applied AI with @ElectrifAi (Leveraging Existing AI Models)

Hi everyone, so sorry, um, you're here everyone. Uh, yeah, I accidentally, um, did not see this, this button, uh, that was hidden and it hit it, it clicked it and it all went down. So welcome, I'm so excited to be here. It is Thursday, I can't believe another week has gone by. This week was an incredible week for many of us. It is the end of Q4, of Q1. Um, 3 months of this year already gone. Amazing. Um, so I am Noelle Russell. I am your host of This Week in Applied AI with Electrify. I'm so excited to be here with you. Thank you for joining us, um, live. I see a bunch of my friends, so, so good to see you. Remember, the chat is live. This is a live presentation. If you catch this on the replay, definitely take notes, save the date. Every week at 3:00 PM Eastern time, I get online and just for about 20 minutes talk to you about the latest, greatest, and fun things I am learning as I venture into the world of helping more companies implement and leverage artificial intelligence through practical machine learning. And Electrify, I'm gonna actually talk to you a little bit about it, but basically I am the, I'm Vice President of Solutions Architecture here at Electrify, and what that really means is that I'm the one, you know, once we've talked to a customer or talked to a partner and said, hey, here's who we are, and they say, wow, that sounds interesting, I get to step in and start building out what we call machine learning accelerators. So we go in and for really low-cost kind of POC, we put together an acceleration package that allows us within about a month to show a company how to leverage existing machine learning in order to see fast results. And it has been an incredible ride, as some of you saw, especially those of you on LinkedIn. Thank you so much for following. For those of you on LinkedIn, I know that you have seen me announce my joining of Electrify, and it has been a whirlwind quarter since then. So I'm super excited to share with you what what I've learned so far, and hopefully if you have any questions, just jot them in, let me know, say hello if you are here. I would love to know that you're watching, and if you have any questions, or more importantly, if you have any comments, just chime in. So today we are going to focus on something that I get asked all the time. So customers are often asking, you know, okay, this sounds great, right? The concept of taking someone else's machine learning, Electrify's existing models, and just Customizing them for our use, it almost sounds too good to be true. So how does that actually work? And so in today's episode, and we are on episode 4, in today's episode we are going to talk through exactly what that looks like. And there are lots of different hybrid ways of leveraging applied AI or practical machine learning. And I'm just gonna talk through one of those scenarios. But the nice thing is, is that this is not something that, you know, you have to, just do with me here live. It's something that I will give you a couple resources and you can go and play with these machine learning models right away. And it'll give you a chance to kind of take a look at the models that are available, the ones that are published, and investigate the inputs and the outputs and how these models might fit into your organization. So great to see you. Hi everyone. Nice to, to have you. I know it always takes a little bit of time for people to see the post and then click on it and join me. So thank you so much. And like I said, this is live. So if you have any questions, comments, just wanna say hello and cheer your friend Noelle on, now is your chance. You can just jot a note in there. So, so great to see you. Okay, so let's get this party started. I'm gonna actually share my screen and go through a small little presentation to show you, to set up the stage. And then I'm gonna take you into one of my favorite places to go GitHub, and it's really where I dove into a lot of the models Electrify offers and started to see how could I use this. So we'll talk a little bit more about that in just a moment, but let me go ahead and share my screen now and I will dive in. Hello, uh, dive— I'm in the inception moment, um, but I'm going to dive into, uh, this model factory Right? Let's take a tour, shall we? We do this every week. We just spend a couple minutes walking through the tour. But then I wanna show you very specifically how to leverage some of these models. And you can do so without even, right, even talking to me if you want. You can actually go and do what I did. Dive into GitHub, dive into AWS SageMaker and start playing around with these. So this is the journey, the customer journey. We always encourage customers to think about when, and I talk about this all the time, some of you who know me, about artificial intelligence and how You know, it's not as easy as it sounds. Or, you know, sometimes we get this idea that it's— and maybe sometimes we see marketing that's like, oh, press this button and poof, you can have AI in your organization. And that's not necessarily true, right? Because oftentimes it's a function of your data and where it is and how good is it and what format is it in and does it match the model and does the model actually, you know, have what it needs to make a good prediction. So one thing that we've done over the last 15 years is focused on building models and something that I heard, you know, I say every week, Andy Jassy taught me during my time at AWS, there's no compression algorithm for experience. There's no way to compress, right, the building of a model from scratch. For example, at AWS we were talking about the cloud But really here it, it works very well for machine learning as well. You can't speed up progress if you're building models from scratch, right? There's just a certain amount of time and effort and learning that goes into building models from scratch, which is why we're really big fans here at Electrify.ai of explaining and showing to you models that are already built. We actually have a few different types of models. We always want to though start, and actually there was a recent article by Andrew Ng, I think he, I can't remember, maybe it was in Harvard Business Review. Anyway, I'll post the link. But he ended up writing an article that said AI is more accessible than ever, but you kind of have to know what you want from it. And I totally agree. I wrote a post earlier this week about this. I think it's really important that what we think about as companies, what is our mission? What are we trying to accomplish? What is our goal? And then identify how AI can accelerate those goals. And that's exactly what we're trying to do here. So every time we get on a phone with a partner or a customer and we talk about what we're gonna do, we often go through this kind of solution framework, right? Where we figure out what, what problem do you wanna solve? Do you wanna have less customer churn? Do you wanna have more customer engagement? Do you wanna know when a customer is going to default on a loan? Do you wanna help them before they get to that point, right? Do you wanna identify, you know, maybe it's predicting errors in claims processing or maybe it's contract analysis and just being able to help the legal team extract information so they're not using their eyes to go through hundreds and hundreds or thousands of pages of PDFs, right? These are classic examples of where AI can help, can augment the ingenuity of our customers. But it also can help do things that we never thought we could do, right? So many times, for example, computer vision. Computer vision is such a huge opportunity because it allows us to do things that we kind of wrote off maybe as like, well, I mean, that's kind of futuristic, it's science fiction. But many times these use cases have become science fact. I'll give you an example of something that we were, you know, we're thinking about is just analyzing video, right? We talked about this last week. What if you could analyze video you're already taking for quality assurance purposes, but now leverage that to predict and identify patterns in performance, behavior, sentiment analysis, right? There's so much you can do to extract value from data you're already collecting for any number of reasons. But now you can use AI to gain even more insight, more value from it. And I always find that I spent the last, you know, 15, 20 years of my career talking to companies about, hey, doesn't matter that you don't know what to do with the app. Just save all your data. Just put your data somewhere. Someone, some magical being will come and make sense of it someday. But it won't happen unless you say, you know, put your data somewhere. So over the last, you know, my time at VMware, my time at AWS, my time at even Alexa, we're like, just save everything so that we can then use it to gain insights eventually. Well, at Electrify, we believe eventually is now. If you have data that you've been saving and hoping you could gain insight from and you don't have those insights yet, now's a chance, now's the time, right? And we've created what's called an ML accelerator or machine learning accelerator to get you Just look at one use case that your data could do and for a super low cost in less than a month, use it to see value and poof, you are on the AI journey and we can help you get there. So pretty exciting stuff. This is typically what one of those Machine Learning Accelerator projects look like. We, our goal is to head in the direction of 4 weeks because I like 4 weeks. It's great to be able to see an AI model turn out a result in 4 weeks. And the way that that often works is that the first week we work on data acquisition, right? Training the model, getting early ideas about how close that data model is. And that's what we're gonna focus on today. We are going to take a look at how is that data structured. We call it a common data model. And in the common data model, we can actually expose to you as the customer or you as the partner and say, Here's what we need for this model to work. And there's some required features, there's some optional features, there's certain things that will cause the model to perform exponentially better than other things. And we work with you on that. But the good news is after about a week, we know whether or not how this model's gonna perform against the data that you have. And so we can deliver to you a quality assessment report and say, here, here's how, here's what it looks like. And from there, we can actually use this time to build train the model and then start inferring results, right? Going through this inference process of asking it questions, getting results, and calibrating whether we're going in the right direction of the KPIs we want. And this can happen in the scope of a month. Most AI projects that I've been on, I mean, some of them, they're still going, right? I started them in, you know, 2014 and they're going on to this day. AI can take a long time. So we wanna try to figure out ways for companies to leverage the power and capability of AI, but not have to actually go through the process of building it from scratch. So I wanted to introduce the concept of our model libraries. We have a couple different types. One is our pre-trained models. These are like computer vision models, natural language models, right, that are trained on like the English language or trained on a common open dataset. These things like we know, for example, when we take, when we use computer vision, there is a baseline, right? ResNet model that we can use that will train a model on, you know, what is it, what's a picture of a car look like or a video of a person, of a hat, of a computer. Like that's known entities. But where there's value is in an organization like ours helping you custom train a model on your specific data to solve your specific business problem. That's where we see the magic happen. We call that pre-structured models instead, right? Because it's not gonna come pre-trained with somebody else's data. It's not gonna come pre-trained with a public dataset. It's gonna be empty, but it's going to be tuned and feature engineered based on our domain experience in banking and finance, in telecommunications, in entertainment, right? In the areas that we, we know work and that we've got domain expertise in. We now allow you to custom train those models for your use. So you get that acceleration capability because we know this work, we know these industries, and we can help you answer questions before you even ask them, which is kind of exciting. So I just wanted to show you a couple, but really the, the question that often comes to my mind, and certainly the mind of organizations I talk to, is, okay, this looks fine and good, right? You can be in these different industries and you can help out with customer training. And customer engagement, lifetime customer value, right? There are certain models that you can help with, but how does it actually work? And that's what I wanna spend the last few minutes here today talking about is how it actually works. So I'm gonna head over another moment where you'll see behind the curtain, go into the live studio and open up a browser and just show you some of our models for executing requests so you can take a look at them and even test some of them. Some of them out. But what I was most interested in, and what my, you know, those I talked to, what they ended up being most interested in, is this concept of a common data model. What does this model need to work? Because, you know, the devil's in those details. Like, the challenge is in the moment where I say, great, I want to use your model, but do I even have the data that you need? And we want to answer that question long before you even think about what you're gonna implement, right? As soon as you identify the business problem you wanna solve, I wanna show you here's what we know in the domain experience and expertise that we have. Here's what we know we need in order to create a good model that'll help you predict the outcome you want. And so looking at these, you know, I always like to look for, I don't know, let's say like a cross-sell model, right? So we have a couple different ones, but if I wanna be able to find out what the likelihood of a cross-sell campaign will be to be successful with some of my, you know, campaign constituents, if you will, right? Or in a marketing campaign, who I'm sending this to. I want to know, does this work? Will this work? And I want AI to give me kind of a leg up there. You'll see inside of this repository is the ability to see the input and output. So you can find out here's what the data needs in order to be successful, right? Here's what— I'm sorry, the model. Here's what the model needs in order to be successful. Here's the data inputs that are required for this model to work based on what we know. Now, one thing you should know about GitHub and one deployment option you have for these models is AWS SageMaker. And I always— let's see if I could just pull it up. I'm just gonna type it in. Oh, AWS SageMaker Electrify. I feel like I've done this before. Electrify. I might be on a new browser. There we are. So we're in the AWS Marketplace, and many of the models that you see published in GitHub are also published as SageMaker models. And what that means is SageMaker's meant to give you the power of automation, right? Like, it's already deployed. It gives you the ability to press a button and have a model that you can feed it data and get an expected outcome. And that description of the input and output is actually found here, right? So you can see, here's the data structure that it needs. Now, because we want to give you the ability to do this test without ever talking to us, you'll notice that pretty much everything listed here is required. But the benefit of working with us directly is that you have the ability to come to us and we can tell you, well, you know what, we actually don't really need that. I mean, it provides some benefit, but we can get 80% there with these 5 things instead of these 10 things, right? So there's an opportunity for us to work based on knowing more about your specific customer or, or data landscape that we can make better decisions about what the model actually needs to be trained on to be successful. And that's again where you get to get this acceleration capability and build on, you know, the experience and skills that we've developed over the last 15 years. So I always, so what you'll see us do over the next few weeks is starting next week, we're gonna start demonstrating some of these models on AWS, on Microsoft Azure, so you can see how this works in process. But in order to do that, the first, First, I had to show you what it looks like. I'm gonna go back to the presentation really quickly and talk about how this works end to end and where this model call will occur. So you can kind of think about it from a business application perspective. Let's say, you know, you're running, you know, any business application in your organization typically has, right, a set of steps. They're run by business rules. Maybe there's a workflow or a process. RPA is often used to automate and accelerate different steps within a business application, maybe the entire your business application. And what we want you to start thinking about, and what our customers are super excited about, is that when they go and they look at a specific step in their business process and they map that to the models we have available, they realize that in some cases it's very simple to see, oh my gosh, right here in step 4 of this business process, I could call out to customer churn and get a probability before I send out the campaign. Or I could send this, before I send it to my actuaries, right, I could send it through this AI model to get a predictive score on its performance, right? So I, I have these moments where I could send it to a human or I could send it to a process that is mostly business, you know, function-based or software engineering-based. But what if, what would happen if you leveraged AI to accelerate that, to bubble up more detail about the results of that problem or give you more ideas about the propensity of, you know, that pattern to show up again, whatever that pattern may be. And of course the best part about that is when you do inject AI, you now have the ability to let your business process start teaching that artificial intelligence model how to evolve, right? We call it reinforcement learning. It allows us to use the successful decisions that are being made within that model to drive evolution in the model itself. And that's something we see oftentimes as being a huge accelerator after we do kind of the smaller project that ends up being a very fast follow for something that we might wanna build later on. You can also see on this screen, you know, so let's say we get to step 4. The nice thing about this is that in a business application, you already know what the inputs are at a certain stage of that business application, right? So what we wanted to do is use this common data model to say, okay, at this stage, here's the data I need. Maybe I need to do a little bit of transformation or a little bit more orchestration to pull data to populate what the model needs to be good. And then, but the nice thing is that I pass in that data, it goes to this model, which is gonna be deployed in one of these Docker containers or deployment options. Also on-premises is of course an option. But I send it to that Docker container,, right? The Docker container does its work. It actually has quite a bit of stuff going on in there, and one of our videos will dive deeper into what's in the container you, you end up hitting when you talk to our models. Um, but you go in there, there's a bunch of, you know, ETL, uh, data quality checks, integrity checks, like there's a bunch of different things that happen. Then it hits a series of algorithms that generate your output, and you see that listed here. The business application just sees the, the percentage, the number, right? The propensity might, you know, propensity is high or claim complexity is high or churn is, you know, 80% versus 40% versus 10%. And this data now can be used by your application, enter web developer, app developer, right? And that developer now chooses, right? And the business of course chooses what are the thresholds for those scores that will determine what we do within that workflow. So maybe we say, you know, if this claims validity check comes back and says this claim is, I'm 80% sure there's no errors in this claim, maybe that's enough for us. And we say, if it's 80%, continue on, if it's less than 80%, bubble it up to someone to take a look at. Maybe it's 75%, right? This is where the human in the loop becomes critical because that number, that threshold is very much dictated by the business, right? And by those using the application. Application. So, so this kind of paints the picture for what we're, you know, the end-to-end usage of these models is it's not a, oh, let's go to Electrify and run our entire business application, right, through multiple models. It's where within our existing functionality can we just bubble up a little bit of AI to generate a big shift of the needle in our results, right? How do we reduce the amount of time we spend on something? How do we increase the effectiveness of who we are sending things to, right? Or who we are processing things with. The AI can be used to help us refine the scope or expand the scope specific to the patterns that are in our datasets. So this is a huge opportunity for a lot of companies to get into AI who may not have, you know, have kind of written it off as like, well, we're just not there yet. But the reality is, is everyone is there. Everyone can be. So how do you, how do you figure out what a good place is to start? So there's lots of different models that we talk about. But one of the things I always like to do is just, you know, spend— and maybe I'll go back and share my screen in just a moment around the GitHub repos. But this is really where I find the biggest value from my perspective as a technologist, as a futurist. I'm always like, how does this model work? So I am drawn to those models that expose to me, here's the data I need to execute and here's some sample output so you can see what it looks like. One of the nice things about working with us directly is that we have Swagger, right? So we use the Swagger API in order to test it out. As a matter of fact, next week I'm gonna be demonstrating a model for you leveraging Swagger so you can see what that looks like. But it's nice because you can vet out, does this work or doesn't it? It gives you the ability to see what the realistic output will be of some of these models and how you might use that data to generate insights for the organization based on that use case. Really exciting stuff. Great, Flavia, so nice to see you. Hi, Warren. Benito, good to see you. I'm not in I don't get to see my chat all the time when I'm sharing my screen, so, so great to see you all. Thank you for joining and commenting. There's quite a few number of comments here, but thank you so much for chiming in and letting me know you're watching live. This is really exciting stuff. It's AI for, you know, I always, I always say this, I'm actually in the process of writing a book about this topic, but right, this is where we, we have an opportunity to get AI for everyone else, right? For the masses. How do we make the world of artificial intelligence, the world of accelerating capability available to those who are not gonna actually go and build a machine learning team and hire 10 data scientists and BI analysts, right? Some organizations can only afford 1 or 2 of them. And even then that 1 or 2 may not have a clear guidance of how they can help the company. So one of the things we do at Electrify is really just increase the ingenuity of those resources, right? Increase the capability of those resources. So I've been with you for about 25 minutes. Thank you for all of you joining. This is so great. Hey, Paul, nice to see you. Great to see some friends. Thank you so much. But yeah, like I always say to my friends who are not necessarily technologists, But like, AI should not be a secret. Like, this stuff is pretty amazing and it can change your business. Like, let's get the word out. So that's one of the reasons why I'm super excited about the work we're doing over here at Electrify. If you are a partner, we would love to partner with you. Please reach out. We have a way to contact us on our website, but I am active, as most of my friends here know, active on LinkedIn, active on Twitter. So send me a DM on Instagram, like TikTok, you tell me, I'll be where you are. But it is such a great time and now on the heels of this pandemic, right? Oh my gosh, starting to see the light in some parts of the world. We are going to start moving away from being worried and into a world of prosperity and affluence. I know it is hard to think about it, but we are going to be like, wow, I can go outside again and I can go to the movies and I can take a cruise and I can go fly on a plane and see my family. So companies that are involved in this, in, in all of these ecosystems need to start getting ready and AI can help. And, and, and of course I wanna help you get there. So thanks so much for listening this week. This has been episode 4, all about how to use a common data model in order to leverage prebuilt machine learning models. So I showed you a little bit about that in the PowerPoint presentation, but most importantly, head on over to GitHub Take a look at Electrify's models. Check out the specific markdown file on inputs and outputs for each model. Start to get a sense for what this looks like. It's not as hard as it looks. My job is to demystify artificial intelligence, right? Let's make this available to everyone. Everyone can win. One thing I often say, you know, let the rising tide of this technology, let it lift all the boats, not just the yachts, right? Not just the FAANG companies and the big tech companies and the companies that can afford machine learning, let's help everybody. So I'm really excited about this. Thank you so much for watching. Check it out. Hey, Barbara, nice to see you. So good to see you all. I'm looking forward to next week. So if you have ideas on what you wanna see, I'm thinking it'd be cool to see a demo of an actual model live. So that's what I'm gonna do. But which model? Let me know. All right, take care. Have a great weekend. For those of you who celebrate, it's Easter. So happy Easter. Happy Good Friday tomorrow, uh, and of course, happy Q2! Let's do this! All right, talk to you all later. Bye!