Cloud Crunch
Cloud Crunch

Episode · 2 weeks ago

S4E7: Building a Center of Excellence for Machine Learning and AI: A Journey across HC, Insurance, and PE


As we wrap up Season 4 of Cloud Crunch, we will discuss "Building a Center of Excellence for Machine Learning and AI: A Journey across HC, Insurance, and PE". This powerful and insightful conversation is led by the Director of Marketing, Michael Elliott, and co-host Fred Bliss, CTO of all things data at 2nd Watch. Our honored guest is Jason Mass, Executive Director of Data Insights for 2nd Watch.

Involve, Solve Evolve. Welcome to cloud Crunch, the podcast for any large enterprise planning on moving to or is in the midst of moving to the cloud, hosted by the cloud computing experts from second Watch. Michael Elliott, Executive director of Marketing, and Fred Bliss, CTO of all Things Data at second Watch and now here are your hosts of cloud Crunch. Welcome back to a new season of cloud Crunch, and this season we are going to focus on AWS Reinvent, the biggest cloud conference in the world. Our intent is to enable you, the viewer the o portunity to immerse yourself and how cloud has evolved since last year on topics like preparing and building a center of excellence, extracting data insights, managing a cloud native environment, and data center evacuation. Joining me today is Fred Bliss, CTEO of Data Insights and Jason Maasts, Executive director of the Data Insights Practice here at second Watch. Welcome to cloud Crunch, Fred and Jason, thanks for having us. Thanks good here. So the focus of this video cast is building a center of excellence for machine learning and AI, and I have to admit, you know, I've seen a lot around machine learning and AI. But I'm not very knowledgeable when it comes to this subject, so I'm gonna be asking a lot of questions and try to try to act as the audience as I learned along with you. So how has AI changed? You know? Can you kind of give us a history of how AI has really evolved over the years, and then how that really starts. It's to play into data and analytics. Yeah, I...

...mean it's I think looking at the history is important here and given where it were a podcast of that's uh not going to be several hours. I'll try to cut to the chase. But you know, at the end of the day, Uh, it starts with the stats, right, And we've seen this evolve over the years. There's been a lot of money, of a lot of VC money, a lot of private equity money, a lot of software product marketing going into this over the last couple of years. But I think there's things that are fundamentally different today uh than even uh they were a year ago or even a couple of years ago. So I think if you look at the the big auto AMOULT tools like data robot um, the idea was to empower uh non technical users and um even some data scientists to be able to leverage different models to find the most accurate one. UM. And I think there were a couple of problems right one. Uh, A lot of big enterprise organizations didn't quite know what they wanted to do with it, and today a lot still don't write. They were um, they were looking to some of these auto mL AI solutions to solve problems that in many cases could have been solved through traditional rules based ones, and in other cases, UM, trying to kind of invent problems. And I think what's fundamentally different today is that, uh, the different machine learning models out there, UM, they've evolved and they've matured in a way that ah, it's highly unlikely, unless you're working at something like Google or Facebook, UM, that you're going to be inventing a new model for your use case. With the evolution of large language models over the last couple of months, look at open a I S GPT three, UM, what's happening with uh, you know even just uh uh stable diffusion And in all these um uh these models out there,...

...we're moving from a model centric AI world to a data centric one. And now given the kind of the intersection of compliance and data privacy and AI ethics, bias explainability. These black box tools that produce you know, predictions and results aren't gonna cut it anymore. Now it's all about how did you train this model, what was the data you used? And how can we leverage some of these large language models um that exist today to create new solutions. Yeah. I mean, if you think about the history of a like, AI is really just another word for you know, predictive analytics. You know, it's been around for for a long time. It's really taking data using that data to help predict outcomes. Um. You know, tools today have evolved, you know, thinking back you know in time, like even even the two data robot you know, tools have a valve that they're making them so much easier for even you know, business users to to use. So um the really the key to all of that is really having a data platform and where that data has been curated and it is really ready for for our official intelligence for data science use cases. Is it moving though, beyond predictive and starting to get into prescriptive? Yeah, I mean that's that's the big dream, right as there's a whole field of research around um uh, a lot of money being thrown into uh, using AI to basically do anything for us to automate everything. That's uh, that's a whole, uh, whole kind of worms that we probably won't get into here. But let's think about some of the more interesting enterprise use cases, right, well, yeah, and and and how does that apply to you know, say, health care. How does this apply into the healthcare industry or insurance industry or private...

...equity industry. How are they using these tools machine learning and AI to recognize trends or to be able to solve problems or look look ahead? Yeah, think about healthcare or even a pharma UM. If if you're trying to do clinical trials for a new pharma drug that, let's say targets cancer, UM, the biggest problem isn't really necessarily what models am I going to use to try to predict, uh, you know, whether or not this drug is going to work? Um. The biggest problem is, UH, in the case of like these large language models like open the Eye, what made them successful. It was the immense amount of data that they had to train on. You know, you had the entire internet to train to train this data on. The same thing with some of the AI R you had all these images across the entire Internet to use. Now, if you're thinking about trying to create a new drug that targets cancer, UM, what's your limitation. It's how do you find enough doctors to go look through let's say radiology imaging or UM you know CT scans and say yep, this is cancer right here, and it's this type of cancer. Nope, this one isn't and you're basically just going next, next, next. Imagine like those captions that you got where you're identifying traffic lights, because what you're doing in those things to prove your human is to train Google's self driving car algorithm. And in this case, we don't have enough doctors that have the time to go through in label data. And so I think this is the biggest gap right now. UM that's preventing some of the bigger use cases from moving forward. But with the introduction of some things like synthetic data UM tools like snorkel AI are starting to do some of these things, and the ability to create mL tooling and labeling uh UM intake tools for UM for enterprises, we can start to democratize...

...the labeling process and the intake process and the review process to generate that training data and make it better and better and better. So what about use case for private equity? I mean, you're talking about cancer and private equity seems like it's completely data driven. You can get your hands on the data versus radiology images with hip requirements and everything around. That tell me a use case of how this works for private equity. I think a good one for private equity is actually, um again, inspect to the boring stuff, right, it's if you think about a private data is Come on, if you think about a private equity company though, UM, Let's say let's take a boring one, right, boring sector. Let's say they're acquiring H, I don't know, car washes and every month they are quite they are buying a new car wash in a different part of the country. H. What's gonna be the biggest time stuck on all of that? And UM, uh, you know what they want to know throughout this whole thing, it's all that dot commit review. They're going to want to come through all these contracts. They're going to want to comb through every piece of information that uh is going to be costly from a time standpoint in getting that company acquired. UM. To finding out if this is a company we should acquire, and three all the legal costs that go into that looking through all these pds. So what we can do now with UM some of these large language models and the tools that assist in labeling this data is to extract that data from the contracts UM to start to figure out what we care about in this contract and labeling it so that an entire paragraph could essentially be summarized down to UM, so and so is true, right, is this profitable? Or is this a risk? Yes or no? That's the real power, I think, And you can apply that to almost any document type out there, or any any PDF, any document for insurance, it's claims processing. There's so much untapped information out there and UM in these UH the s big on structured data documents UM that...

...there's there's just so many use cases that have just been unlocked. I mean, even even even for insurance, like being able to you know, take historical quote data and and help help use or use that data to really predict whether or not you should provide a policy to a potential policy holder. Right, So you can use that data. You can really apply a lot of different UM, different AI models and and use different variables to help predict UM whether or not you should you know, increase your your policy because you know the potential policy holder UM had some accidents in in the past, or really using data from all of your policy holders to understand if if this new quote is going to be an issue for your insurance company. And Jason, I think you just hit the nail on the head on what's different this time and UM, you know, if you look to some of the laws in New York put out around UM being able to use AI for screening candidates, it's the same thing you're for insurance. Right. If you don't understand how your algorithm and what data is being used to train whether or not we should accept this applicant or whether this would be a risky person. UM, you're gonna have a hard time with regulators. Right. If you're rejecting a person for the wrong reasons, that's no good, right, But it's also no good to accept people for the wrong reasons. So you've gotta have a lot of faith in what's coming out of this model, which means it's all the way back to how you train and label this data UM and and the ability to explain it in the transparency going into it. That's what's fundamentally different this time around. So then, Jason, how do you start building these type of solutions? Well, I mean I think I think the key here is just making sure your data is ready for it. I think a lot of uh,... know, organizations start with, hey, I want it's the big shiny object in the in the room. I want data science, I want AI UM But a lot of times what we find is that they're not ready for it. They don't have an enterprise data warehouse, they don't have UM their data curated, UM, their data really ready for UM AI and m L. And so, you know, really building these solutions, if if you build it up front, you build that data platform, you can enable a lot of different data science use cases. And so really starting with that platform, getting the data ready, cleaning it, transforming it so that it's ready for AI and mL is how you should start these type of projects. Well, I imagine most enterprises don't have probably that skill set yet in house. So how should they engage with with help making that happen. I think that's that's that's the big difference now, right is uh in organizations? But sure, it's going to be different in every organization. Right. Sometimes they've got data science teams that are part of the business. Sometimes they're separate teams. Sometimes they're part of it, sometimes they're sitting in their own middle world, or sometimes they don't even exist in the organization, right. But I mean we see that as well. Yeah sometimes uh, you've got a data science application out there, but really what they wanted someone to make dashboards, which is one of the most uh painful things for people in the field. But I think where you have to start and where we're seeing a company start now is what's different this time. As I'm hearing businesses, business leaders UM talk about problems that uh in asking how they solve them, and they're not saying, hey, how do we solve this with AI? These just happen to be business problems that hey, they are now solvable by AI. And that's what's different this time around. It's not just whould we use what what use case should we come up with? So I think where we can start now is UM is more run any I strategy, right, like, how... we create that center of excellence in the internal tooling that they need? Because what you don't want to do as an organization unless you're really really small, is you don't want to outsource all your i p to some third party um AI company that's going to go build this model for you. In the case of that P company buying car washes, they do this enough times, they're going to have an awesome model with a lot of really good training data to be able to understand and have a good flow for acquiring car washes. That's a really valuable i P UM paying another company to go build that for you and then they own that. I don't know is it worth It depends the scale, I think, right, But those are some of the questions that I think you need to ask and UM. Building the right teams and structures in place, and having a good strategy UM for what you're going to build, how and how you're going to scale it and what the right team structures are super important, right especially with data privacy laws. Absolutely, I mean I couldn't agree with fred more. I mean, really you need to again have a strategy around AI mL strategy UM. You know, second Watch offers a data readiness, a data science readiness UM offering that that really looks at your data, looks at some of your use cases UM sees really we're exploring whether or not you have enough data to really make it an effective use case UM for data science. So UM, we're we're working with the business and those assessments to to really understand what you're trying to do, what the data environment looks like, UM, and what what different kind of variables that you want to use when you're you're building your models to to help predict and really solve for that use case. So I couldn't agree more that a data strategy, a data science readiness strategy, is really important to engage first on.

So you have that that solid plan. Well, I want to thank you Fred and Jason for joining us today. Very interesting conversation for me around building a center of excellence for machine learning in AI. As you think about the attendees going to reinvent, any final words of advice for them as they they want to you know, most most reinvent they want to learn, So any words of advice for them? Uh, what are some of the things that are not being automated within your company? Think about that right again? What are some of the you know, unstructured data is an easy place to start, UM, don't jump just to the what can we start predicting? Because the first thing you need to get us buying and trust into leveraging mL within your organization. So what's something that you can solve for today and get moving? Excellent? Well, thank you for listening to our show. This video cast is intended to add value to any large enterprise that is planning on moving to or currently focus on leveraging the value of the cloud. Send your comments or suggestions to cloud Crunch at second watch dot com and see it reinvent. You've been listening to cloud Crunch with Michael Elliott and Fred Bliss. For more information, check out the blog second watch dot com, forward Slash Cloud dash blog, or reach out to second watch on Twitter and LinkedIn.

In-Stream Audio Search


Search across all episodes within this podcast

Episodes (43)