Um My name is Melissa and Melissa Tone and I'm a general manager of Document A I at Scale. Um I am from Turkey and I came to the US after high school.Even though I have learned English in high school, I always really wanted to be able to come and find opportunities for myself. My family still lives in a tiny town in Turkey on the western side and they are store owners, they make perfumes and sell perfumes. I picked a very different role um and path uh in my future towards machine learning. So I'm happy to tell you a little bit more about that too throughout the conversation today. And when I graduated uh from Stanford University with Computer Science, I worked at Facebook for multiple years uh on product and machine learning side and then joined Scale A I and what I joined Scale A I for was really about data and machine learning. I have realized that throughout my lifetime, the investment to machine learning is going to increase because what people are really looking for is to be able to use this new technology to bring new benefits and advantages to their lives. And I, when I really saw that along with the amazing group of people, I thought I would really make a shift and start working on machine learning scale was really built upon to be able to provide a the making building A I easy for everybody.

That means how can we build products and solutions to ensure that every company um across different industries can easily use machine learning instead of really investing in a lot of these difficult processes to do it themselves because many people actually can't. And the thing that really separated uh our solutions was that we really focused on data. I'm sure many of you are working on artificial engineer, uh artificial intelligence and machine learning and realize that really what's very important to be able to get the best results is about the quality of data. And many of the different projects hit model building plateau that talk about here where they get to a level with their models, but the results really hit the plateau and can't really increase that further to be production ready and be working uh by itself. So scale from the beginning, really focused on how we can focus on cleaning the data and making it really high quality to be able to be fed to the models so that we can get um very high quality results out of those uh models to be able to apply to different industries. And I found this approach really interesting when I was joining as well because there were already many research that is, that was looking at this, the data driven A I how important data is.

But I haven't seen a lot of companies, even existing companies and new up and coming companies investing in this problem because it was hairy. It was really difficult, it was also a a not a glamorous problem, it's very difficult problem to solve. So I was impressed that um you know, scale was really seeing the actual problem and really starting from there to build solutions that really attracted me that paid off later. And today, um even three years after, since I've been here, when we look across the world and different industries, we are seeing that the pace of technology is creating induced demand. What do I mean with that? Um even us as consumers, we want our packages in one day, right? We want um our loan applications to be a in 24 hours or less. We want, if we want to make a spend with a corporate card, everything to be handled afterwards within seconds. And we want driverless cars, we want um our cars to be produced a lot faster and to be delivered faster. So we want more, we want better, faster, cheaper solutions uh across the world. And the only way to meet this demand by the consumers and the companies is machine learning. It is funny because until now we really only talked about machine learning across different industries, about Moonshot projects, right?

Yes, of course, like driverless cars, we talked about it, but we also talk about it in more sense that research projects, not something that is applied and with us uh every day. So what I want to talk to you today is how it's time that actually machine learning can help us every day and how we can apply it into problems that you may not have thought about before in problems that seem more. Um uh you know, that that may seem to a human easy, but in, in reality, it's very difficult to solve them in real life and only machine learning can solve them. One of them I want to get to is actually document processing. Why? So when I was also at scale and have been working on many different business units, including building our government business unit. Um I spent a lot of time thinking about this problem because I, I would meet with many different industries and leaders across the world and realize that this is a real problem for them every day. There is uh thousands of documents, millions of documents across different industries every day that they're dealing with that with important information trapped in them.

And a lot of these industries are really um staying, you know, trapped, not knowing how to move forward and meet this induced demand because a lot of the information they're still here and cannot be extracted from unstructured documents with the accuracy that they need to serve.

So I started thinking a lot about this problem because scale already had many products that we call in the now platform to support companies uh who want to build machine learning in various problems. But a lot of the companies in different industries, we saw before they even needed a machine learning platform. They had this problem, they really needed more of an end to end solution they used and applied the machine learning to get high results rather than thinking about the ML problem at that moment. And it really hit me, I built a business unit around us to be able to tackle this problem and help a lot of companies who are just thinking about machine learning and who are thinking about which you use cases, they can really apply this to be able to get real results pretty much immediately.

And this problem really accelerated, right? As the race to digitization has been increasing, this is a problem, we are going to spend $5 billion as, as the world, as the industry to solve uh this year, just this year and many of that effort will still will not pan out to great results because there there are so many challenges in the industry in processing these things and getting the value out of it.

The main challenges that we see starts with these four, they are clear key entry problems because many of the people use um uh manual uh work to be able to enter these values um and extract information and those costs really increase. And existing solutions don't scale. We'll, we'll, we'll talk about it a lot more and there's many duplicate digitization efforts in companies, especially enterprises across the world. So we are going to be spending $5 billion but it is still up in the air. How much of that will help companies with a real solution?

A big problem as I said in this area and the investment is not panning out to bring a return on investment is that current approaches are falling short. Mainly there are two approaches companies are using in document processing today, one doing them manually and second using OCR, these are falling short because in the first one, it's really not scalable. If as a company, you, you you're growing, your volumes of documents are increasing across different industries and you cannot use manual work to be able to extract this information. It also causes a lot of different problems. For example, there's still a lot of mistakes and it creates a false sense of confidence. Since manual work is doing this, people are trusting the end result more even though they're equal same mistakes. Uh because uh I always say it's funny but our fat fingers are real uh across globe line. When we're typing, we're touching different keys and we are being inaccurate And at the same time, it really causes retention problems with our teams that we really make them do this kind of work because they want to do work that can't be solved with technology.

And they really want to focus on the work that is not rep repetitive. Uh The second current approach is also failing. And as I said, that was using a technology like OCR that has been created in the past. Um And this is really OCR only gives you a raw text meaning it doesn't really know where words are. It's, it's extracting a raw text uh for, for you. And, and at the same time requiring you to create incredibly hard to maintain templates to be able to make any sense of that information. And it's our engineers, machine learning engineers that are building those that, that these are the people we should be focusing, we should be keeping focus on our domain, competitive edge creation and building models and not extracting information and creating templates. So this is why we really focus on a solution that can bring machine learning uh uh to every company who wants to use it without the challenges of machine learning implementation. And what do I mean with that? It's without the challenges that you know, you make the investment to this uh in house, you do not have to assemble machine learning teams.

And even if you have it, you can keep them focused on the domain problem that you're trying to solve and really get a lot more value out of that. Instead of really focusing it on a problem that can be solved with using a solution that exists like our solution. And really helping these companies move towards data centric A I over model centric A I. So we built a solution that we call document A I. And it's a very interesting solution. It doesn't rely on OCR. So it's like a like multi step that uses OCR and also uses a lot of different models across uh natural language processing models to computer vision models that focus on understanding the words, the meaning of words on a page and the layout of a page. So you can always think about it as like building a small human brain to be able to understand where things are on a document, just like we would if we were reading uh the document and understanding it ourselves as humans. So these models do that as well and extract the needed information out of it and it doesn't have to extract like a specific taxonomy. It can be adjusted later. I'll mention why this is super important and um and extract it at a very high quality.

What we further do is like as as we talk uh to different customers and handle different use cases, we do an additional fine tuning on the model for the customer with a sample of their documents to make sure that accurate is really there, we understand any potential edge cases that might come with their use case and really adjust the model to understand and definitely uh be aware of those edge cases as well.

This really really helps with the accuracy and also guarantees the accuracy. Um Yeah, and and with this solution works with any layout, you need absolutely no template. So it's been a really helpful across different use cases. But let's look at like some of the use cases, what where this might be helpful. One of the people that we have partnered early on has been Flexport and Flexport is doing it noble duty they're building right? The backbone, the infrastructure, the creating system of global trade.

What do I mean with that? Right? Like if, if we are getting packages today in one day, those packages are coming from somewhere and freight, freight forwarding is a key aspect of uh satisfying, not only this consumer demand, but also at big scale, being able to ship products to different from suppliers to the purchasers around the world.

So what they're doing is I said really noble to be able to, for us to get those uh packages and to be able to get uh everything that the world needs across different companies. But in the world of logistics documents mean a lot, they demonstrate um ownership. So if you lose those documents, you lose ownership. So they're very important, still works on pretty much on documents. So Flexport initially came to us that they wanted some efficiency, right? They didn't want to deal with these anymore and they were getting very low qualities from the existing solutions they have had and we worked with them to deploy document A I and the results that we have seen are really incredible in terms of the end results for them. One, they didn't have to do anything. Their operational teams could actually focus on problems that matter to them instead of processing documents. And the deployed models immediately have shown that over 95% of all documents across the board could go through with no human intervention.

And and we process thousands of hundreds of thousands of documents per week. And then later that turned into per day uh and everything processed within seconds that also helped them with cost savings. But we don't only, we didn't only stop there, right? The impact of machine learning that was really clear to them, especially during the pandemic during the pandemic. Uh there was a big supply chain crisis and they were also affected as we all were. Um but especially flexport because they didn't have any idea what was happening on the world. They didn't have visibility in global trade, just like any other uh logistics company. And they came back to us and we started talking about what information exists still in these documents and one critical information that existed. And before they didn't need this and hence we weren't extracting, it was arrival times in North America alone. The arrival time accuracy was about 14%.

This was 80% lower than the prepa level. So before the world had visibility on global trade because they knew what was going where at what time and was that holding true on what is reported on these papers as arrival times. But during that supply chain crisis, it was 80% down accuracy. So we quickly get together and fine tuned like did a quick fine tune on the model to be able able to extract this model, uh extract this field across different documents in seconds so that they could again get global trade visibility. And this really gave them competitive edge in the industry because when nobody else could see right, the visibility for global trade, they could and pass this information to their customers that really benefited from this one with visibility and to understand what's really uh lagging behind what's going to arrive and automatically adjusting their end relationships with the consumers as such and giving the visibility to consumers at the end like us, if we were waiting for something or what was available in the inventory and in the warehouse.

So really everybody in the equation have benefited from this. This is not logistics is not the only area, the other area that is benefiting from a solution as such. And machine learning in this area is financial services. And I talked a little bit about the induced demand. I mean, ask yourselves, right? We want every financial services service that we're using today to be in seconds, you're using your card, you want it in seconds, you want to buy a house, you want it to be across, you want it to be accepted and approved immediately. So such a solution have literally changing the world right now, how companies are defining their relationships with the consumers. And a good example for this is BR so R really came to life as a tech forward, um really financial services platform that assisted and users across.

All right, starting with a corporate card. And as they grew, they added incredible solutions across the board. And what they wanted to do when they came to work with us is how do we ensure that our customers when they're working with us don't have to worry about anything. They don't have to touch paper, they don't have to send anything. They don't have to um look at any type of paper. And we really implemented this across the board to be able to get closer to this vision. And the results were also incredible, especially we went after their, you know, we worked with their bill pay uh uh product supporting different companies and implemented this uh solution to be so that the A P accounts payable departments and companies don't really have to do anything.

Um everything 99% of documents go through a process with no intervention whatsoever. Everything process this in less than five seconds and really enhancing the user trust with the end users that they uh they have. And one thing I love about this is because bre X is actually a solution that have used OCR and other kind of intelligent document solutions before. And they really did not get any of the results that they were hoping for. And they finally have seen what does it mean to really implement machine learning and see the results in action for that? And it's not, it's not only financial services and uh logistics. The story is the same across the board in manufacturing ecommerce and being able to understand, right?

Like customer sentiment across the board, machine learning can really make the difference. Not only these Moonshot applications were talking all together, but also really in day to day use cases uh that we are using every day, companies needed to be used every day and it's really time to be aware of what can machine learning can make a difference too in these use cases and have our companies start implementing some of these to be able to carry the benefits to the end users for us.

So thank you so much, everybody. And uh I really appreciate you listening to this. I'll also look through this um across the chat, but thanks for joining us. Um I'll stay back for a few more. Minutes if there are any questions and look through the chat. But otherwise I really appreciate everybody uh joining. Thank you, everyone. Thanks Krista and Shri Yeah. Um Yes, I think the presentation will afterwards uh will be shared. But if not, you can feel free to reach out to me here. And I'm also Melissa Talk ma at linkedin. Thank you all. Have a wonderful day.