Get to Know the Hyperscience Hypercell

Brian Weiss: Good morning, everybody. Welcome to the webinar where we are going to introduce you to the Hyperscience Hypercell.

I am joined by my colleague Chris Bloomfield, Senior Director of Sales Engineering, and I’m Brian Weiss, the CTO. Chris, I realize we’re doing a “follow the sun” model here. Chris is based in London. I’m based in San Francisco, so we’re getting a little global support model going here.

What we’re going to do today is introduce you to the Hypercell and Release 39. But at the same time, look, I realize many of you who are customers maybe are not up to date in the latest two or three. So we want to give you a feel and a sense for who we are now, where we’re going, the types of investments we’re making, where we sit in the industry, and the types of disruption that we’re driving, as well as some of the details of the platform itself.

Let’s get into it just to level set. Hyperscience is a technology company. We were founded roughly eight years ago by machine learning engineers, and they had just sold their first company to SoundCloud. It was around licensing. They determined that if you could bring a machine learning proposition to the hard problems in the enterprise to bring AI into the enterprise, meaning: Can I get what humans have to do? Can I train and build machines to understand that information and take care of those tasks? That is fundamentally an AI proposition.

They went after handwriting and some of the most complex human information that comes into the enterprise as their first endeavor. That has really grown now to a company that brings AI into the enterprise in really rarefied environments with very sensitive information. We operate both as a software platform as well as SaaS, as well as a hybrid cloud. We operate by the end of this year in FedRAMP High in the federal government space. We are well capitalized, funded roughly $300 million since origins by top tier investors. Stripes, Tiger Global and Bessemer are the primary. We are roughly 200 and some odd employees. More than half of those are engineers. And half again of those are machine learning engineers. We’re very technology focused on this problem of understanding human information and making decisions about it. We are running on pure software margins, and we go to market with an extensive ecosystem of partners, the top tier: IBM, Accenture, Deloitte, Capgemini, those folks are our partners in the market as well as others.

The company has grown up in heavily regulated verticals. What that means is we are dealing with very sensitive data. You can see the government industries here, talking not just social security numbers, but also insurance industries, whether it’s claims you’re talking with data that cannot leave the enterprise. And is in some cases highly regulated. Building machine models with those digital workers that do that work has really hardened the platform to enterprise class. Doing that, we have solved for the problems that right now in the AI marketplace everybody is freaked out about, which is: Where’s my data? Who owns it? What’s training these models? Can you jailbreak and get it out? All those things like needing to go on-prem in the first place to do this, we have solved for those problems and really grown up having to solve them from the get go.

What we achieve in the market is hyperautomation, and we define hyperautomation as automating tasks at human level accuracy. We on a regular basis get 99% accuracy and 98% automation. The beautiful thing about the platform is it is designed to create accuracy. It’s evergreen into the system. We take a human-in-the-loop approach. Instead of people at the end of the process to deal with the things that it got wrong, you put a little bit of people in the process to stand next to the machine when it raises its hand and it’s not confident. That means twofold things. One, you get a lot of automation in that process for a little bit of effort, but you’re also training a sharper and better model at the same time. As it gets smarter over time, unlike classical technologies that were just brute force that don’t actually learn, we of course do this at the highest enterprise standards, as I mentioned before, having to use environments that are air gapped, and being able to have tight control over the type of data you use and what’s going to the model, the kinds of uptimes that are required. This is not a casual affair. This is mission critical work in enterprises that operate at rarefied levels.

The other piece is, so many of you probably think these days, if you feel like I do, that AI for all this excitement, there’s also a bit of a solution looking for a problem. It is pretty exciting what we can do and what’s possible. Most of our customers are excited about ROI. How do I take what AI can do for me and deliver immediate results? The way that we do that is by inserting a digital worker effectively to anything that’s a human task. If you do that in the back office, because most legacy technologies will cap out in accuracy at 70 some odd percent, you get eye-watering ROI as a result.

A couple of examples for that. We work with IBM at the VA and have for a number of years now, and we are up to roughly a billion pages a year processed. That’s nearly every claim that comes through the VA. They are looking at $400 million of savings over that time period per year. What’s important about this is not just the savings, but for us and for veterans, it’s now takes three weeks instead of three months in order to process a claim. Similar at the Social Security Administration, we partner with Accenture. In that account, we are producing over a hundred million claims a year, and each of them have many documents associated with them, and over a hundred million dollars in savings captured out of that one. CFOs love Hyperscience because it’s AI with ROI associated with it immediately demonstrable.

A little bit about the industry as we look at the disruption that an AI driven approach takes to the market. I don’t think it’s any secret for any of you that AI is a transformative technology. It’s probably everything that we do and touch. I think somebody made a comment that it’s on the scale of the internet itself. It’s possible that that’s true, but really what enterprises are looking for now is people still bring up a little bit about how enthusiastic it can be. But let’s just get some ROI. Let’s get something automated maybe. Back office automation is a prime target for that work in the technology market itself.

Think about the markets that are out there. The technology have tried to solve the problem of understanding complex information, human generated, whether it’s handwritten or odd structures or whatever, the things that clog up your system. The IDP market, really, OCR is driven, that’s really about getting pictures off of a page and it’s rules driven. Every time something changes, you have to write another rule. You have to build another template. That’s a few billion dollar market. RPA, which wraps around that is about taking rules associated with it and automating something. But also it’s very structured. It’s rigid. They don’t learn. Your bots don’t get smarter and they don’t learn.

The BPO industry, which is people wrapped around what machines fail at, is a testament to just how much disruption is in this market. If you bring a digital worker to this proposition from the get go, understands information and 99% accuracy, and can automate the tasks that we are trying to circulate by running around and grabbing information from various systems, we don’t really need the people out the other side of it. This is the market that we are going after. It is the entire process of where people get in the way of the flow of data in an enterprise, and how can we bring AI to that proposition?

We do it through a model centric architecture. On the left hand side, you’ve got enterprise content and we feed that into the Hyperscience platform, the Hypercell. Underlying it, there are 25 plus core models. This is intellectual property we’ve developed over eight plus years, tens of millions of training runs around things like handwriting, deskewing, denoising, all of that. NLU models are also incorporated in there for full page extraction. But then our customers build very specific models on top of it with their data for classification, identification, extraction. We also enrich that data. NLP and NER are part of the platform for text extraction. So we then, with the accuracy level that you get out of this modeling process, we pump that down into downstream systems.

I would bet that many of you on the phone right now, this is your experience with Hyperscience. This is the first chapter of Hyperscience. The second chapter is we start to put orchestration and decisioning into that moment of understanding. If a digital worker can understand all of the complexity of a box of documents despite the variety and the different categories that are in there automatically, why wouldn’t you ask it to approve or deny the mortgage or flag it for fraud or make decisions which are based on data from enterprise systems? Instead of pushing it downstream, we’re in fact now pushing data into Hyperscience in order to drive a decisioning process. In some cases, we’re tapping out to additional models which help us make that decision. This now really starts to look like a digital worker which lives inside your environment that you train, supervise, QA their work, but you can also task them to make decisions with what they’re learning about what’s on the page.

There is another chapter that I’ll talk about briefly here, which is once you have all of that enterprise truth in high fidelity, folks from the ML Ops department are getting pretty excited about it. Because if you’re working with maybe training your own language models, that can be conversation bots internal to the organization, or answer the questions, or using a RAG use case to be able to ground those… this data is very, very valuable. People are starting to dump data of Hyperscience into a RAG or an LLM that they’re training. Now you get an ensemble model, which is the Hyperscience models, the custom models from the Hyperscience data, and of course, any third party or external feeding back into it. This now becomes a whole quantum leap forward, not only in automation, but also any kind of other AI driven, task oriented work.

Now, let me talk a little about the Hypercell. The Hypercell is the encapsulation of our platform, the tools associated with it delivered turnkey. If you start at the bottom of this, what you’ll see is you start with enterprise content which is structured, semi-structured and human friendly. Anything that’s noisy that machines don’t naturally understand, we pull that through various connectors, whether it’s APIs or S3. Of course, we’ve got all of our long history of very deep pre-trained models for handwriting languages, rotation, deskewing, all of that. NLU models are also incorporated in there for full page extraction. But then we’ve got the ready-to-train models that are built on your data and working in combination with those. That ensemble is what gets you the accuracy.

The piece in the middle, these copilot applications, the most you’re familiar with, I hope, is where you define and you train, supervise QA, and you measure. This is your sort of how you would think of a worker. We’ve developed those applications to make it simple and easy to get high performing models running on the platform. We also, that can plug in other types of models to ask ’em questions, whether you wanna reach out to a summarization engine or Llama, or you’ve got a model that you’ve built you wanna plug into. This whole thing lives on a pipeline that at the top of it represented with custom blocks. We can move things through a pipeline, make decisions, add more data, subflows making decisions, and then of course, at the end of it, you’re kicking into some… either it’s the decision’s been made and you’re moving it to a downstream system like one of records—so, “yep, go ahead, pay the invoice, I’ve done all my checks”—increasingly sometimes we’re pumping this data into the ML Ops team’s work in training their own models in-house as well. Lots of great, but again, the Hypercell is about making that turnkey and deploying it out.

It’s proven. You’ve seen from our customer base and you’ve seen the types of thing, a billion documents to the VA… that’s not for the faint of heart. It’s secure. We started answering those… there’s no way to start the company and grow to where we have without first answering the question about data security. We can run public cloud, private cloud, on premise and soon to be FedRAMP High in the federal government. It’s scalable. Scales out, and it’s cost effective. You’re getting the right tool for the job. You wouldn’t want to pay for a 60 billion parameter model when for a fraction of that price, you can run on CPUs and build one that’s even more performant on your own data. So that’s the Hypercell.

I’m now gonna turn it over to my lovely accented colleague. There’s no webinar good without a little British accent, Chris? Absolutely. Gonna turn it over to talk about Hyperscience and some of the specifics of R39.

Chris Bloomfield: Great, thanks Brian, and pleasure to be here and speaking to you all today. As we are building on this Hypercell, what have we delivered in our latest release, R39? From a headline perspective, this is driving this fundamental shift in AI driven automation, focusing in on three core areas: building new, the most world class extraction and classification models, and the ability to be able to train, QA, and supervise those in real world applications with our no code Hypercell approach. Coupled that with our flexible orchestration layer to enable end-to-end automation and extensibility.

With that said, you’re gonna just dive in a little bit around the model lifecycle management, and as we continue to iterate over our vision to democratize machine learning, and in particular, streamline this model lifecycle management to make it as simple and robust for any business user to deliver AI led value to their business. We are focusing here on things like data management, and that’s ensuring that only the best quality and most representative data is actually fed into your models. After all, bad data in models equate to poor model performance and a low ROI.

Then we look at model development. This is ensuring that the models that are built are robust to unseen data and are resilient to changes in real world scenarios. Then when we’re in a position to deploy and maintain them, there needs to work seamlessly alongside your existing systems and workflows within your ecosystem, and provide the ability to continue to monitor that model’s performance and adapt it to your day-to-day operations. Finally, doing it all with the utmost security in mind, whether that is compliance regulatory, or whether it’s things like AI ethics involved as well.

Moving on, and let’s double click into this model life cycle management at scale theme. Now, what we see on the right hand side here are five key components that make up R39, the top two I’ll demo for you shortly. The first of which is Incremental Training. Effectively that this enables you to take an existing model, add data as this representative data set changes on your day-to-day operations. Instead of starting from scratch, we can start with that existing model. This will reduce the cost associated with training those models whilst delivering those models quicker to you and reducing your time to value.

The next one is Training Data Management. For those of you that aren’t aware, we’ve invested heavily over the course of the last few releases in being able to work out what actually makes a representative sample set. We’ve taken all the good work that we’ve done around field identification and table identification and applied that to classification.

The next is around Automated Upgrade Management. Ensuring that your infrastructure is resilient to upgrades, that you are able to seamlessly move workflows and models and keep up to pace with the Hyperscience platform.

The next one, Trainer Resiliency. We’ve all been there, network outages, things that are outside of your control that caused unfortunately a trainer to stop. Prior to R39, the trainer would have to start again. That’s expensive. It’s time consuming. We are now able to pick up where the trainer left off, and again reducing that cost associated with training models, but also doing that in a timely fashion.

The last one, we all want to know who’s touched what, where, and why. We have continued to improve on our Audit Log capability, tracking user events and processes metadata. As I mentioned there, the top two, I’m gonna pivot now. I’m going to share my screen and we’re gonna have a look at Training Data Manager with Incremental Training, and then move over to how we apply that to classification models. Bear with me, I’m just going to share my screen. Brian, any commentary whilst I’m doing that?

Brian Weiss: No, remember, the great thing about what we’re doing here is using AI to make the AI better. As Chris shows you Training Data Management, that’s really about casting forward into your data and helping you understand and use the items with the most impact. It’s a really powerful feature in the platform, and we’re thrilled to extend it now to classification.

Chris Bloomfield: Great. Thanks, Brian. So what actually does constitute a representative sample set? That’s a really difficult question to answer. With previous versions of Hyperscience, the way we kind of approached it was that yourselves as an enterprise will go out and try to solve for a particular use case. You’ll go and gather together what you think is a representative sample set. You effectively upload or dump that into Hyperscience. You annotate it, you train the model, you test it and realize that actually you’ve done a horrible job with your annotations. There’s poor data in there and it’s leading to poor performance, and you consistently have to iterate across that.

So with Training Data Manager, we’ve taken this almost like AI copilot approach to take the guesswork out sort of understanding what makes your data set representative and how we can make it clean and free of the anomalies that we’ve seen in previous iterations and training. The most performant model is based really on variation and not volume. What do we mean by that? We don’t need 400 documents for each individual element that makes up your data set. Let’s use the AI capability, this copilot, to be able to group documents together into like groups and then say, “Hey, we need more documents for this particular group to be able to solve for this particular subset within your own data set. But hey, we’ve got too many here and we might over bloat the model.” The idea here is that you need much smaller tailored data set to be able to drive a greater performance. A smaller model equals quicker training equals lower cost for you as well.

I mentioned earlier about humans and being horrible at things like annotations and it’s one thing that pains my team in sales engineering when we work with our customers around technical deep dives and delivering proof of concepts. We are now able to guide the user, what we call a Guided Data Labeling, into where the interesting data resides within each of these groups that have been identified. We take that human input, and then we learn from the human say, “Hey, on this particular document, this is where you were annotating before, we suggest that this is where that interesting data is located.” Humans can choose to follow that kind of copilot approach. But sometimes they might disagree and they disagree for various different reasons.

So with that, we’ve actually introduced Anomaly Detection. We’re actually able to surface up again before you train these models, before it starts gonna cost you as an organization and say, “Hey, we think there’s some annotation errors that exist in here, go and fix that.” We’ve done this across field identification and table identification, and I mentioned earlier as well about the ability to start or retrain a model from an existing model. So we’ve now introduced that capability here, and we can see what in the past would’ve just been “I have to train it from scratch”—lengthy, costly, time consuming process. Now we can take that existing model and we can just incrementally add that new data set as it adapts to your day-to-day operations, and give you that model faster.

Now, we’ve taken that approach and we’ve applied that to Classification. For those of you that are used to training classification models on Hyperscience, this is R38 and prior kind of interface. What you do here is you’ll be creating your semi-structured layouts, and you’ll see each individual layout appear here. Think of that as a classification bucket. You’d simply upload all the documents that you think that constitute each of these individual buckets. And then also you need to provide some sort to that model. You need to be able to say, “Hey, this is not what an insurance invoice looks like here.”

Again, if you haven’t gone through that data set properly with a fine tooth comb, we end up with situations where the model is misclassifying things, or it’s very, very low confident, which therefore means a higher human in loop then in the old days. So what you’d have to do then is actually click on this and download that data set, go back through it, and be able to remove where those anomalies are. And that again is an extremely time consuming process. So we’ve taken all that goodness and all that enrichment that we’ve offered around field identification, table identification, and apply that to our classification.

So this is the R39 version of that same classification model. We can see all the goodness and enrichment of this AI copilot approach. We were able to show you the training data health. “I need more documents, I need less documents.” “Hey, you’ve misclassified this particular document.” Going in and reclassifying this as simple as pulling up the page directly within Hyperscience and clicking on the specific layout. So here it’s being tagged as policy document. Potentially these are addendums to an insurance invoice for argument’s sake. So I’m going to stop sharing my screen. Brian, anything that you wanted to add there?

Brian Weiss: What’s really interesting there is a lot of times we get asked like, “Oh, do I need a data science to do this work?” Right? You’re talking about model training, you’re talking about labeling. We have spent an amazing, an awful lot of money in the technology to ensure that it’s simple and easy. What you just saw there is, now I don’t have to have a data science degree to be able to go in and train a model. I can quickly understand, the machine will tell me what makes it more accurate over time. If I make a mistake and annotate something incorrectly or put in a document that really is an outlier, it’ll prevent me from contaminating the model. All of this is part of how we just make it simple and easy to get to a place where you’ve got really high performance models operating on your behalf. And it’s very simple and easy to keep them evergreen, to manage drift and things like that. Now, those all sound very technical, but again, you just saw it, you can tap into the platform, it’ll tell you how to… it’ll curate what you’re doing. So you get there quickly and it, you end up cheaper, faster, better with folks who can be easily trained to develop out all kinds of different modeling opportunities.

Training Data Management. Great. Chris, you’re gonna talk a little bit now about some of the automation pieces that are also coming in 39.

Chris Bloomfield: That’s correct. The second half of what I was going to discuss in this session is our approach to delivering World Class Enterprise Hyperautomation. It’s four elements here. Similar to that previous slide, we’ve got a couple here around Freeform Text Fields and Multiple Tables Automation that I’ll be demoing. With that freeform text field, this basically allows our flow developers now to add freeform data through the Custom Supervision interface. Being able to pull in information from third party systems, allowing your end users to be able to directly integrate with those sources and change them on the fly directly within the same interface that we are extracting information from as well.

With **Multiple Tables Automation**, we’ve focused in the past around simple gridable tables, there’s non-gridable tables, there’s also nested tables. But what about a combination of all three? And let’s say if I’ve got all three multiple times throughout my document, how am I gonna tackle that with Hyperscience? We’ll see our approach to multiple tables automation.

Then we’ve done a lot of work around language improvements in particular around Chinese. Now you can process any document from structured unstructured human friendly content in Chinese, which is extremely valuable for our international organizations who are requiring that understanding and processing of Chinese content.

The last thing on here before I pivot back into a demo is the out of the box connector now with Amazon S3. Our customers now can integrate and configure the right storage directly from the Hyperscience platform user interface. I’m just going to pivot again into a screen share for a demo. And if you bear with me screen two, and we will start with complex tables.

I spoke just briefly then around simple gridable, non-gridable, nested… well is a fantastic example of all three. It happens to be a loss run. Now, nice simple one pager, but this will go over tens of pages or even hundreds of pages complex. So what we have here is effectively sort of tables within tables. So class as nested tables here, we’ve got a non-gridable table. It’s not like a chessboard, but here we’ve got directly within that table a nice gridable version. So with our multi table approach, now we can very simply annotate this non-gridable table, and then combine that with the gridable table functionality.

So I can go here very simply. Go and tag just one line. We spoke about the AI copilot and I will click Continue to Review, and it’s been able to ascertain, so I’ve only had to annotate that one line, and the AI has been able to kick in and be able to pull in all of that information. This results again, in us being able to process what were outliers around multiple tables. And here we can see an example of one that went a hundred percent straight through processing. We can view that table as it loads. We can see all of the information has been extracted. We can double click into each of these individual line items, whether it’s that inner table or whether it’s that parent table there as well.

I’m gonna move over now to the enhancements we’ve made around our Custom Supervision capability. I appreciate we’ve got a lot of existing customers that are on. So I’m just going to talk quickly about the custom supervision approach. Where we’ve gone beyond just classification, identification, and transcription and human in the loop, we are now able to validate and enrich from multiple different disparate sources and pull together a 360 degree view of a particular business process. This happens to be a hospital insurance claim. And we can see on the left hand side, we’ve got all the different documents that make up this claim pack, and we can see the actual documents in the middle. And on the right hand side, there’s all the information that’s been surfaced up by our digital worker. This is now neatly formatted for a human to be able to review and drive hyper decisioning approaches. There’s not only information that’s been extracted, there’s recommendations based upon information that’s been pulled in from external systems, the based information and the reason why we think it should be rejected. But also we’re able to bring in here Gen AI summarized information about supporting documentation.

Here I spoke about Hyper Decisioning. We can now provide in that interface an ability to not only accept or reject the claim or whatever kind of levels of acceptance or rejection or whatever decision you require, but we can add additional information in here as well. On a per document basis, there might be some superfluous information in here. Somebody might have written something that could be relevant to the case. I can capture that here. And I can also provide holistic information about the entire case. Why am I maybe actually accepting this claim when the digital worker has recommended that I should reject that? So we can capture all of that information to help this hyper decisioning process. Brian, anything in addition that you’d like to add?

Brian Weiss: I think that that piece around Custom Supervision and becoming a decisioning headquarters is something that might be novel for some of our customers. In that, I mentioned bringing data in from outside systems in order to enable that process. Maybe where you thought about human in the loop is a place where new supervision, what we do now is a lot of customers are bringing the decision point for paying a claim, flagging fraud, right? Those sort of things. Deciding whether the discount the person has got on this particular payment is accurate or not, because they’re not paying it. Like those sorts of decisions that ordinarily lived downstream maybe now can be fully incorporated into Hyper. And the advantage is the documents are all right there. If you need to do a review of what you’re seeing and looking at, it’s all right in front of you. If you know you want to be inserted a decision, maybe you can just automate it.

I, Chris, what I wanted to do as we as we close out here, is I wanted to to broach the topic for a discussion here around other LLMs. People often ask like, “Well, how do you compare to a Chat GPT or et cetera?” There is a lot of not only confusion, but also sort of buzziness about that. I wanna dig in a little bit on that. What in the time we have left here, and I’ll start with this example, and I love this one, this the CEO of Canada Life, who’s a customer in lots of subsidiaries asked by German insurance trade magazine, “Like, are you using AI?” And they say, “Yeah, not Chat GPT, that generates experiences. I use Hyperscience to understand data, and to understand information like prescriptions with doctor’s handwriting.” And the interviewers hear says, “Nah, you can’t understand… no, but no, that it’s too messy, doctor tan notoriously bad.” And there’s no, no, it’s… it gets it right. It’s a hundred percent right. We get accuracy. That’s AI, that’s AI going to work to do human things, understand what normally needed humans to be in the process. And very savvy answer, like, of course I’m using AI, but I’m using Hyperscience. ChatGPT actually get it wrong badly. We have a couple examples where that’s exactly the case.

I wanted to touch on this one. I know Chris, you worked closely with this client. This is FTI. And FTI is a customer. They manage their travel order processing in Europe, and so they get things from EasyJet and RyanAir and all of that. They’re trying to manage the changes and the cancellations and the processing of travel orders. And it is time sensitive. These things are happening. They’re also very, very stylized graphic emails that come in with attachments that have itineraries on them and things like that. We at Hyperscience pull all that in, chew through all of that data. We have models built around understanding it and very accurate, except there’s this problem that some of the accuracy is degraded by the question of “Is it a cancellation or is it a change?” Now that’s actually, from a text analytics standpoint, if you only have those two words, that’s actually a hard problem because “I changed my flight,” “changed my plans,” “I wanna cancel my flight,” “I cancel my plans, I wanna change my flight.” Like, think of the permutations that the confusion matrix very high there. So we’re struggling with that.

We said, look, let’s just pop this out to an LLM. Let’s go. ‘Cause those things, you know, 60 billion parameters, not quite that, but they’re really good at that nuance. Let’s see how it understands what we’re struggling to get right here. And we found some really interesting things. Number one, we do all the extraction very well. We’re high accuracy in that if you hand all that over to an LLM, even with the bajillions of parameters, it gets it wrong a lot of the time. It gets really important things wrong, like the dates and the locations. Like switching the “to” and the “from” around. It’s just not understanding what’s on the page. What it gets right, and it gets right nearly every time, is whether it’s a change or cancellation. So we’re able to develop an ensemble where we do the hard work, and then we ask the LLM a simple question: “Is it a change or cancellation?” And then we reconcile both of those in Hyperscience, and if they disagree, it gets flagged up and somebody has to say, “Okay, now I know which one’s right.” And I think, Chris, we’ve like, it’s gotten down to a hundred percent pass. I mean, it’s just screaming for accuracy right now.

Chris Bloomfield: Absolutely.

Brian Weiss: We do have a little bit of time, Chris, I would be thrilled if… could you demo some of that right now? Do you think we can do a quick one?

Chris Bloomfield: Yeah, sure. Bear with me.

Brian Weiss: And I have an ulterior motive that many of the customers who are on with us today may not have be using Blocks and Flows and the kinds of decisioning. What you’re also gonna see here is sort of how we orchestrate asking questions of the data, making decisions about it. You’ll see it kicks out to an LLM and those guy comes back in. So that whole orchestration is part of the Blocks and Flow layer.

Chris Bloomfield: So, thanks Brian. So what we’ve got here, we’ve got a submission that’s made it all the way through, and there is a discrepancy that exists between what GPT has found and what Hyperscience has found. Have a little view kind of under the hood of how we got to that point.

We’re gonna here have a look at the Flow Run. What we can see here, each green box represents a function that occurs on that submission. And this is how we start building the input and outputs of the flow. So we have that traditional IDP approach where the combination of machine and human in the loop as well. And here we can see with the output of the metadata that’s been passed in. This is the data that’s been extracted from the large language model. We can see here that it’s tagged this as it’s a flight change and it belongs to a supply number U2, which converts back to EasyJet in this particular instance. And it’s pulled through all other information, booking reference number, flight numbers, departure airports. And then we do a number of different conversions. EasyJet tend to send their emails without properly formatted dates. They tend to do different languages like different European languages for cities and airport locations. We need to convert them into three character IATA codes as well. And then we end up comparing the two. What has come out of the large language model and how does that compare with Hyperscience’s highly accurate models as well.

What we can see here, and Brian alluded to this earlier, is that the models themselves, large language models can really hallucinate. And what we’ve got here are example a highly stylized email body from EasyJet. This happens to be a flight change, and we’ve got things like planes and different kind of logos in the way, and they kind of… instead of reading left to right, it’s top to bottom. And this is confused GPT. Basically they’ve swapped around the arrival and departure locations, whereas Hyperscience is being able to pull that in because we’ve trained this highly performed model. Now this is tie break decision for a human to come in, “Hey, do I agree with what Hyperscience says or actually do I agree with what a large language model is telling me?” And we can see here that indeed Hyperscience is correct and I can simply click on Complete Task and continue. We have continued to iterate on this particular balance between LLMs and Hyperscience approach. We are now near 100% automation with 100% accuracy with this particular use case. Brian, back over to you, sir.

Brian Weiss: Great, I’m glad we had time to see that. It’s always really intriguing. I get asked a lot like, “How do you fit into the AI landscape of all those things?” And really it’s a “better together” question. It’s a “what is the right tool for the job?” And in that particular instance, you can see like a language model, whether it’s GPT or Anthropic or Llama or whatever, like that’s trained on a lot of internet conversation and text data. It’s gonna be really good at change versus cancel. It doesn’t hold a candle to a targeted model that’s looking at this data day in and day out and learning on the frame of it. And by the way, when accuracy really, really matters, right? This is, we live in the world where accuracy really, really matters. And it becomes very interesting for training models into a RAG use case or et cetera, because the model is only as clean as the data you give it.

We operate in that rarefied high accuracy world where we have to get it right. You can’t get it wrong. And things… people go into the wrong locations and historically getting it wrong means it’s just really expensive ’cause you have to send it to people in order to do the work of getting it wrong. We’re pushing things like at the VA where you go from three months down to three weeks. Now we could get it down to days. That’s real change. And we do it that… so the answer to the AI question is like, it’s an ensemble. It’s a best tool for the job. And we play very, very well and do some really interesting things with these other language models, but it’s all orchestrated on a platform, right? You’re not having a higher data science and maintain all of that work, et cetera. It all gets orchestrated on the Hypercell in a platform environment. And that’s very, very different than anything you send people of OCR where you just, it’s an API and you send it and you get data back. We’re well beyond the place where we’re just sort of pulling data off a page.

I’ve got some interesting questions in the chat. I wanna make sure we hit one of ’em is about copilot. It’s asking, “Hey, we use the term copilot. Is that Microsoft’s?” No, no, no, no, no. That’s just a concept of copilot. It’s not a Microsoft thing. What we’re really saying is that to do a task like I need to annotate a document, I need to figure out how to train my model and make it better, which documents do I pick? The phrase “copilot” just refers to we’re taking AI model and going and looking at the data ahead of time, bringing you the 10 documents that are gonna matter the most. We cut your work down by a hundred fold at least. So that work of sort of prefiguring and running ahead of you and figuring out how to accelerate that process… which look, if you live in machine learning, that typically is a very complicated process. You’ve got data scientists, you gotta be training, you got your PyTorch and all your notebooks and things like… you don’t have to do any of that. It’s all sort of wrapped up on the platform. Chris, what other kinds of good questions do I see here?

Chris Bloomfield: We’ve got a question around FedRAMP, Brian.

Brian Weiss: I think I mentioned that we will be FedRAMP High by the end of the year. So for any of you folks and government clients who are… we also work with really great SIs in the government, but when you’re ready and you wanna go to SaaS, that’ll be the end of the year. I think somebody asked, “Do we sell software On-Prem?” Yes. We started selling software On-Prem. That’s been a part of our pedigree in the industry.

I’ve got a question for “Can scan PDS expected to have the same item.” Yes. And we’re fundamentally combining both vision and text. And so we’re looking at shapes on a page as well as the text that’s on a page. And so PDFs will follow that category. Someone’s asking about grainy, “is it if something’s super grainy?” Yeah. That’s part of what we do. Like our underlying models will sift all of that out. It’s part of that eight years of generating really high performance vision, what’s called computer vision code to do that work.

Here’s a question. “What OCR do you use?” Okay, I’m glad somebody asked just to make sure we’re clear. This is not an OCR engine. OCR is one character at a time. Is it an A, is it a B, is it a C? We’re doing machine learning, which means the machine is looking at many examples and understanding what it means, whether it’s position on the page, the text that’s on the page, and making a decision about what it’s seeing, including handwriting. So the OCR is a very deterministic rules-based approach, and it’s our proprietary underlying modeling that allows us to make sense of the page in the first place and what’s being written on it. What else have we got here, Chris? A couple others?

Chris Bloomfield: Somebody’s asked about our Arabic handwriting. Brian, I’ll take that. I look after the international business and we absolutely do support Arabic handwriting and machine print across any document set. So please feel free to reach out to us after the webinar. We’d love to have a chat.

Brian Weiss: I think we’re close to time on that. We’ve got a bunch of questions here. We’ll make sure that we come through and answer ’em to you. Thank you again for your time and your interest. We hope to see you on the next one. We will make this a regular affair to keep you up to date on not only the software, but also our direction in the market. And we hope you have a great weekend either day or evening, depending on which time zone you’re. Thanks so much. Thanks all.