🦄 Making great presentations more accessible.
This project aims to enhances multilingual accessibility and discoverability while maintaining the integrity of original content. Detailed transcriptions and keyframes preserve the nuances and technical insights that make each session compelling.
Overview
📖 AWS re:Invent 2025 - Automatic Enterprise ML with Amazon SageMaker and Nextworld (SMB201)
In this video, Valerie Knafo from AWS and Luke Hollenback from Nextworld discuss how Nextworld leveraged Amazon SageMaker Autopilot to build an AutoML platform that dramatically improves enterprise application development. Nextworld's no-code, AI-native platform enables non-technical users to create business applications. By implementing SageMaker Autopilot, they achieved a 95% reduction in model build time (from four weeks to two days) and 33% accuracy improvement. A notable case involved organ transplant viability prediction, where AutoML reached 90% accuracy using only 12 of 120 dimensions in just two days. The solution drove immediate business impact: 100% of renewal contracts upsold to AutoML with 54% average contract value increase. The serverless architecture and automated data science lifecycle eliminated the need for expert data scientists, empowering citizen developers to build production-grade machine learning models through a simple point-and-click interface.
; This article is entirely auto-generated while preserving the original presentation content as much as possible. Please note that there may be typos or inaccuracies.
Main Part
Introduction: Nextworld's AI-Native Platform for Enterprise Applications
Hi everyone, I am very excited to introduce you to a company that is rethinking enterprise software. My name is Valerie Knafo. I lead sales for the commercial segment West for AWS. Nextworld is a business applications platform company that created a no-code platform powered by AI, leveraging SageMaker running on AWS. What struck me when I talked with the Nextworld team to better understand their application was the immediacy and scope of the impact.
When they implemented this for their customers, it meant an augmentation of capabilities. It enabled non-technical people to support and create new applications. It gave them the opportunity to get to market quicker and it lowered their costs. For Nextworld, the impact was equally impressive. It increased stickiness and adoption from their customers. They saw an immediate increase in revenue and an improvement in the quality of the results they were driving.
I am very happy to introduce Luke Hollenback from Nextworld. He is a Senior Director of Engineering and AI. What he and his team did and led was create an incredible platform that builds and deploys AI models that actually work. So please welcome Luke to the stage. Yeah, so like Val introduced, I am Luke and I oversee a lot of the teams that do our backend services and our AI facilities at Nextworld. I want to talk to you a little bit more about what we are so you have a better idea before we get into what we actually did using AWS.
Like Val said, we are a platform for building and running enterprise-grade applications. We are all cloud-based. You interact with us on the browser. We are AI native, which is really important for this story. What I mean by that is you use AI to help you build your applications. We have assistants in the form of agents, and they can take natural language and build apps for you. You iterate with them in that co-pilot experience that we have all been familiar with over the past year or two.
We have a ton of facilities to bake AI into the applications that you build, ranging from traditional machine learning and AutoML, which we will talk about today, up to modern agentic LLMs and vision models. We have a whole multi-agent system. We have Ed, our chatbot, but he is agent number zero, and he can actually let you interact with a team of agents underneath him, all for business applications.
The third tier is that any application you build on Nextworld is automatically exposed via AI. We have MCP automatically exposing anything you build to Ed on platform or to anything off platform. Everyone is trying to come up with an AI strategy for the company right now. People are struggling to figure out what that means, but they know how to build business applications. All you have to do on Nextworld is build that business app, and you will automatically have AI strategy baked in.
We are enterprise grade. Everything is audited. It is all managed infrastructure. You do not need to stand it up or manage it or watch it yourself. The Nextworld platform is a generic powerful platform for business apps of any kind, but it has an ERP heritage. If you know what that means, it is enterprise resource planning—sales, inventory, financials—a lot of really heavy, complex business apps. We were built from that from the ground up, which means our capabilities are pretty robust.
From Traditional Machine Learning to AutoML: Dramatic Improvements in Build Time and Accuracy
Let us talk about machine learning and AutoML. We have actually had a traditional machine learning facility for a while. Everything on the Nextworld platform is no-code. You go in, you drag and drop. It is geared towards the citizen developer. We do not want you to have to be a software engineer to build stuff on the platform, although you can be. We have a lot of advanced facilities too.
We have had traditional machine learning facilities for a while. They let you pick an algorithm, train a model, turn knobs and switches, deploy it, and test it—all the traditional data science workflow. But it is catered towards a persona that is experienced and educated in data science. It also takes a lot of time to do that. The cycle time to actually build a model, train it, get it to an accuracy that you feel good about that is production viable, and then solution it and deploy it into a production pipeline is pretty costly. It can take weeks at the minimum.
We started exploring a concept called AutoML, which automates the entire data science lifecycle for you. It performs feature engineering, picks algorithms, tunes them, and ultimately gives you candidate models that you can deploy, do reinforcement learning on, and use in real business applications.
When we started exploring this, we had our team of data scientists first attempt to build it ourselves. We had a complete design and they were hitting the ground running. We got decently far, but in parallel we started looking at other solutions through a buy versus build analysis. Amazon SageMaker Autopilot came to the top of our evaluation. We quickly realized that SageMaker Autopilot was going to be everything we wanted and more. It had all the AutoML capabilities, and it had a bunch of other non-negotiables that we needed to integrate.
We are a multi-tenant platform with a lot of automated operations that has to happen. We don't want humans having to do infrastructure setup and management. SageMaker Autopilot enabled that, and we'll talk about that in a bit.
We released our AutoML product, and immediately saw value. We have the pleasure of being part of the solution architecture for a number of our customers' products and applications. In doing that, we were able to see real improvements in both build time and accuracy. Across a number of them, on average compared to when our data scientists would help them build models and deploy them by hand, we actually saw a 95% build time improvement. We went often from four weeks of development to two days.
On top of that, we saw average accuracy improvements of 33 percent. We attribute that simply to the fact that AutoML is able to iterate through train, test, tune, train, test, iterate repeatedly. It bumps that accuracy up as high as possible way faster than a human can, and it does it automatically, so you literally don't get tired. It will find something that is honed in exactly to the use case you need.
On top of that, our product owners were able to do this, not our expert data scientists. Domain experts are able to do this using AutoML because they don't need data science expertise. The machine does it for them. Down on the bottom, this was valued in Nextworld. The month that we released AutoML, 10 percent of our contracts were up for renewal. 100 percent of those contracts not only renewed, but they actually upsold into AutoML. On average, that resulted in a 54 percent contract value increase to us.
There was immediate value for this in our customer base, and that's continued. We've still been selling and upselling AutoML. We have people demoing it today. Our sales engineers are out there using it at conferences and so on. It's been one of the most surprising success stories in our product line.
Real-World Success: Organ Transplant Viability Prediction Reaches 90% Accuracy in Two Days
Let's zoom into one of our stories. One of our customers actually came to us and they take data on cadavers of people who have passed away, and they try to figure out what the viability is for transplanting organs from them. They have to figure it out quickly because it's a time-sensitive matter and lives are at stake. They had a solution that was built by hand, completely outside of Nextworld, and they were around 60 percent accuracy. When I say accuracy, I'm referring to an F1 score around 60 to 58 percent. It's actually not production viable, as many of you probably know. It's a prototype at best.
Our data scientists went and used the traditional machine learning facility on Nextworld to build another solution. We were able to achieve around 75 percent accuracy, again referring to F1 scores. It was production viable, though not necessarily super good, especially when you're talking about something life-critical.
Our product owner over at the AI space then took the dataset that was provided by them, ran it through AutoML, and within two days rather than the four weeks it took our data scientist team to build that solution by hand using traditional machine learning, they got a 25 percent more accurate solution. It was tapping 90 percent accuracy. It only used 12 out of the 120 dimensions that were provided. It was able to throw away 108 dimensions during feature engineering, and it only took them two days. It was 93 percent faster.
So it's a crazy success story. We have many more like this actually, but this one's the coolest because you think enterprise, but you don't necessarily think life science. This is life science. Very impressive for us.
Let's talk about why. Many of you probably know the top process, right? This is the traditional data science process. You're going to go and gather your data. You're going to manually explore that data by hand and try to figure it out. Hopefully your dimensions are labeled. If they're not, you'll try and make sense of them, and then you're going to do feature engineering.
Feature engineering is where you decide which of those dimensions matter and which ones don't. For the ones that matter, it's where you figure out how much they matter. When you're doing inference in production, you want to know which fields or which columns in your data should be weighted heavier and have more impact on the output prediction and which shouldn't. Then you're going to pick an algorithm, tune that algorithm, and train your model against it.
You're going to test that output against your test data set and get an accuracy score, however you want to judge accuracy. Then you're going to iterate for weeks and weeks, and eventually you'll have a model that actually does something, hopefully that's valuable to the customer or the business. You'll deploy it, you'll monitor it, and we're not even talking about actually solutioning it into something that can be in a business process or flow. And then you can do reinforcement learning in real time. So there's a lot of work and a lot of high-skill work that a data scientist or engineer has to be involved in.
There's no way that a citizen developer or a business analyst in the traditional sense would be able to do this end to end at the high level of accuracy that a lot of solutions require. So then we have AutoML down at the bottom. All you do is point to data in Nextworld. It can be data that's actually first class in the system or data that's been integrated in. We have a ton of integration facilities. You pick the features that you want to infer, so you're going to pick the columns that you actually want to be able to predict when you push data through it, and then you're going to click a button.
You're going to click run AutoML, and it's going to sit there and cook for a few minutes, literally minutes. It's not hours, days, or weeks, just a couple of minutes. You'll get a few results back. You'll deploy the model that's optimized properly for your use case. That's really it. It's that simple.
What is cool about our AutoML solution is when you click the button, it does what it does and gives you 8 options of models at the end, and each one has a different optimization. Our tool will use an LLM to give you a natural language description of what each one is optimized for. When I talk optimization, I'm really talking about things like false positives and false negatives. If you know a confusion matrix, that's what we're referring to, but we put it in plain language for the business user relevant to their use case.
To bring that home more, I think everyone knows COVID tests. When that was the thing, the efficacy of those tests, when they would say this test has this percentage false positive and this 1% false negative and so on, that's a good parallel to how you want to optimize for different solutions. So really, we're just bringing home this complex process to the citizen developer and the business user. We built it all using SageMaker Autopilot.
Technical Architecture and Business Impact of SageMaker Autopilot Integration
We had a few non-negotiables. It had to be API accessible. Like I said, we're a multi-tenant system. We had to be able to drive it fully in code. We were able to do serverless, which is a big deal for our margins, but also what we pass on to the user. A lot of the time when you're doing AI, if you have to maintain your hardware and you don't have economies of scale on your side, smaller customers financially is just not even viable.
So the fact that we could use serverless here was a big deal. We already talked about the AutoML tools, that's basically what we just went over. Then model analytics was a bonus that we got. Our data scientists loved that. It actually gave us not just the data for things like confusion matrices, it also gave us graphical representations and we didn't have to doctor them up at all. We just pulled those out of S3 where Autopilot dumps them and we'd stick them right in the UI of our system.
So that's catering to that more advanced user. Like I said, we want to build for citizen developers at Nextworld, but our advanced users have a lot to work with as well.
The architecture is dead simple. It really is as simple as this. You have the Nextworld client, a browser-based client. Users can run AutoML, get analytics, and make inferences. Our servers are actually interacting with Aurora RDS for data, and then they're just hitting serverless endpoints in Autopilot to do inference, or they're hitting compute endpoints to actually do that AutoML training job, and they come up and go down as necessary for cost savings like I said with that serverless mechanism.
Overall, we saw a massive improvement. Like I said, the accuracy of models with AutoML was one of the most surprising things to us at first, but it makes sense once you get into it.
AutoML is able to perform more data science, feature engineering, and testing, and it can complete the process faster in a shorter amount of time. This means you're able to reach optimized model accuracy and a honed solution faster and better than a human team can in the same amount of time.
Across all the solutions that we were able to look at and we had the privilege of engaging with our customers on, we saw a 95% improvement in build time from conception all the way to solutioning and deploying to production. You essentially cut out the human labor part of the data science process with AutoML, and SageMaker Autopilot has only gotten better and better since we started with this product.
Wrapping up here, we saw an incredible amount of value both to our customers and to ourselves. Autopilot is really the only way that we do machine learning now, and we have a lot of it. Everyone thinks about modern AI and large language models, but you need more deterministic predictions in business. You need to be able to integrate those into agentic solutions, which you can do on the Nextworld platform. We have a full agent builder, and we have all the LLM vision capabilities, transcription, and everything else you could need. We can do the modern stuff, but the foundational machine learning work actually matters a lot, and it's almost more important in many cases for the business problems that people face.
This added value meant our data scientists could focus on building features for the platform rather than helping people who probably don't have a data scientist on staff actually build solutions that use machine learning. I don't have a ton of time to take questions here, but if anyone would like to chat, I'm going to hang out in the gray corner over there. I could get into more technical details or talk about Nextworld. I know this was fast, but at least walk away and check out Nextworld and SageMaker AutoML Autopilot. It was AutoML when we first used it, but now it's rebranded. Thank you for listening.
; This article is entirely auto-generated using Amazon Bedrock.
Top comments (0)