#227 Sedarius Tekara Perrotta: The Importance of Data Quality in AI Systems

0:00We did it out of need, out of pressure, out of pain of being able to deliver more value for our customers because out of the box that didn't exist. So we built this observability layer and a lot has happened since then. We were founded in 2017 or we went to market in 2017. This need for transparency into data, especially unstructured data or files, different types of documents. This has been needed whether or not Gen.AI has come to fruition. It was relevant. It has been relevant since digitalization. And really what has happened with Gen.AI is a problem that has been old, that a lot of people are very, very familiar with.

0:36of what is in my file system, what can I trust before I hook it into Gen.AI. This has come to the forefront, and this is where we're in the right place at the right time, solving a really pertinent problem for scaling Gen.AI in businesses globally. So my journey into this space all started at the US Peace Corps. I was in Transylvania, of all places, and I was working at a think tank. So I was responsible for knowledge management at a think tank. And I did that for several years in my service. And then I went out and became a consultant doing knowledge management projects around optimizing the value of intellectual capital and best practices, lessons learned, key strategies, all that fun stuff.

1:21and did projects for Harvard, MIT, Stanford University, Society of Organizational Learning, and a whole bunch of stuff for the World Bank. And that led me to spend time at MIT, honing my craft at the Sloan School, where I really got into, that was when I got into AI and NLP specifically, because when you do knowledge management, it's very manual and tedious, and was looking for ways to optimize and speed up that process of getting quality data, because it's always been the problem. The problem with knowledge management or any, let's call it any file system management, is the quality of what's in it.

2:04People create multiple versions. They create and they put it in different places. Things go out of date very quickly. Some things are the approved version. Some things are not. things are non-compliant in there and there's really no transparency into it. So I started a consulting company. It was called Neuron Global and that's what we did. We did knowledge management consulting. We served hundreds of companies on knowledge management consulting projects and we really got specialized in SharePoint. So got to experience SharePoint firsthand. At the Ignite, the Microsoft keynote, they actually said that 2 million SharePoint sites are created every day and 2 billion files are uploaded to these sites every day.

2:53It's just the scale is... Yeah. And for listeners who don't know SharePoint, just give a little thumbnail. Yeah, it's Microsoft. There's two major file systems in Microsoft. There's SharePoint, which is a little bit more customizable, which enables you to create portals and communities. And then there's OneDrive, which is more of a Google Drive, Dropbox, Box version of that. So we're specializing in that portal version, which is really knowledge management, kind of 101. And we were doing that for a number of years and ran up into this wall of data quality because everybody's very enthusiastic at the beginning of the projects.

3:36You get a lot of buy-in on the executive level all the way down. And then it gets launched. And then months later, stuff starts going out of date. People find things. It's not what they're looking for or it's the wrong thing. And then they stop trying to use it. And then this negative feedback loop continues until adoption becomes very difficult and hard to enforce. And that was really where we needed to found another company that sat as an observability layer on whatever file system or whatever knowledge system was being used, whether it was Dropbox, Drive, SharePoint, Salesforce knowledge, Zendesk knowledge, whatever it is, wherever people were putting data, we needed an observability layer so that you knew you would be able to identify what was accurate, what was up to date, and what was trusted and what isn't.

4:23So that was the founding of it. We did it out of need, out of pressure, out of pain, of being able to deliver more value for our customers because out of the box that didn't exist. So we built this observability layer and a lot has happened since then. We were founded in 2017 or we went to market in 2017. And then, you know, we were this need for transparency into data, especially unstructured data or files, files and different types of documents. This has been needed whether or not Gen AI has come to fruition. It was relevant. It has been relevant since digitalization. And really what has happened with Gen AI is a problem that has been old, that a lot of people are very, very familiar with, of what is in my file system.

5:15What can I trust before I hook it into Gen.ai, this has come to the forefront. And this is where we're in the right place at the right time, solving a really pertinent problem for scaling Gen.ai in businesses globally. We were talking about SharePoint and what SharePoint does. And the problem is when you get your data uploaded into a knowledge base, then things start going haywire because things get out of date. you find that they're duplicates and that sort of thing. So let's go on from there. Yeah, so I think there's just a natural data entropy that happens whenever you are trying to save documents anywhere, really, whether it's in OneDrive, SharePoint, Google Drive, Box, Dropbox.

6:07It's just this natural problem of trying to manage this chaos of everybody creating documents at different times and different versions and things going out of date and out of compliance. Anyway, this is a real problem for Gen AI, especially RAG initiatives, retrieval augmented generation, and the Microsoft Copilot because they need that data in order to retrieve answers. And that is essentially the major difference between web-facing content and internal company content because there's a veil across an organization's documents and files. and it hasn't been pierced because there's no transparency.

6:46There's no analytics. There's no way to know what's there. So this is a major impediment to Gen AI at scale, and this is what we work on solving. And so that's interesting about RAG. I was thinking it was more preparing data for fine-tuning or even basic training of models, But it's data in a knowledge base that an LLM or some other gen AI application is accessing. Is that right? Yeah. Generally speaking, and these are techniques that companies are using to improve the RAG responses as they're fine tuning the models. But they're trying to work around bad data because most organizations have kind of put their hands up and said, I don't know what to do about this.

7:39There's no transparency into this. I don't know how to fix it, but we need to find a way forward because we see the potential. We see the awesome benefits of Gen AI in our organization to automate more, increase productivity, increase efficiency, increase competitiveness, all those things. So they kind of work around it. But that's exactly the problem that we solve is the data quality issue, the garbage in, garbage out issue. And that is monitoring and keeping track of what is in those documents and files and what can you trust and what you need to filter out. And how do you do that? I mean, presumably you're using AI to do that as well.

8:19We're using AI, we're using chains of experts, NLP, ML. We have over 150 autonomous agents behind the scenes working for us. So it surfaces in 22 algorithms that are specifically designed to, in the sections of documents, so not, you can't compare a document to document that's not very helpful, because a question is being asked, and it's in a particular sentence or paragraph on a particular page, and that's the area that you need to dial into. So we're breaking it down into those sections, and then we're looking for, we're comparing those sections to one another, and identifying duplication, identifying conflicting, identifying areas that don't have context, identifying areas that are non-compliant or toxic or bias.

9:09And then we're surfacing that on an interface so that people can actually see for the first time the quality of their data and can be empowered to take action. And when you say it looks inside the document, if I have one document that says, I don't you know, Craig's birthday is in October and another document that says Craig was born in March. It'll find that? Correct. That's exactly what it would find. That would be in our system on conflict. So you would link it up to whatever sources you have, any source. And then it's this quality assurance layer that sits on top. And in addition to that, you are able to, in real time, view your Gen AI answers.

10:02And we're able to know when an answer was derived from a document that has our section in a document that is bad, that we've identified with our 22 different scores as a risk. So we can tell you exactly when an answer has been generated that is not trusted, that is not accurate, that is not compliant. So that in real time, you have the power to fix things and to improve your RAG and your Microsoft Co-Pilot initiatives. Yeah, that's interesting. But can't you put this, you know, let it run on your knowledge base and reconcile all of the problems before connecting it to an LLM? Or is that just too big of an ask and you really do it live?

10:56Yeah, I mean, it's crawl, walk, run. Certainly that's where we're going, where you can automate more and more and more and then just automatically fix things. Right now, what we do is we identify it so humans can determine whether they're going to fix it. But humans could also set up filters and let's just say plays that serve as gates of content going into the LLM. So you can automate the filtering out of content and you can automate the creation of labels and named entities as well. So that the information, the sections have more context for that LLM. So there's things that you do that you can do in real time, and there's things that you do that accumulate over time and improve accuracy day by day, month over month.

11:44And so if someone's building a knowledge base for a RAG system, would they first do a pass with – and is the product Shelf or what is the product name? Shelf. Shelf. Okay, they would first do a pass with shelf. They would correct things that they know how to correct. And then shelf would run kind of in the background while during inference. And then say, you know, give the user an alert that this may have a problem. them be careful or look more deeply into it. Yep, that's correct. There's really several use cases. That's the very proactive one. We see more of a use case where they've already implemented a RAG or a co-pilot that they're having hallucinations and they want to fix those hallucinations because they can't scale it.

12:48So that's the more common use case is people saying, we have an issue, Houston, there's an issue here and we need a solution. So a lot of times we come into projects where they've been trying to do this manually. And then when they see this opportunity to automate a whole bunch, to actually see what's going on, there's a lot of resonance with that. And we have very happy customers. But yeah, you can also do it before you try to even launch a Gen AI initiative. And that's called data readiness, Gen AI readiness, responsible AI initiatives. Those are what those hook into. And a lot of organizations are doing those as well.

13:30Yeah. You guys started this company in 2017. I mean, that's recent history. But in the world of Gen AI, that's ancient history. What were you focused on before Gen.ai came along? And when did you develop this solution for RAG? Interestingly enough, we've been doing the same thing since we've launched, which was data quality, which is transparency into unstructured data. Like, what can you trust? What's accurate, trusted, up to date? So we've been always doing that because of the story I told you of not having any transparency in other projects when we were doing it from a consulting standpoint.

14:17What has changed is that humans are no longer the consumers of the good data, of the algorithms, of the analytics that we're running on top of that data. It's now it's LLMs are the consumption. consumption. So it's just accelerated something that we've been always doing and has always existed. There's always been bad data and people have been searching and finding the wrong answers. That's a known problem. It's just everything accelerated in the age of Gen AI because people realized that this was kind of their goal. This was their foundation for automating more and going faster and it has accelerated our, let's say, adoption and accelerated the market need for something more general.

15:08Like we started in the contact center where there's ROI to be proved in shortening a call, in reducing the amount of escalations there are, in improving CSAT. So this is where we've been because there's ROI attached to it. But every other department has been having the exact same problem. It's just harder to prove the ROI. And now what Gen.ai does is it enables an ROI, a demonstrable ROI out of the gate for any of these because we can see exactly how many bad answers there are and show them fixing them in real time, how much impact that they've had on their hallucination and answer accuracy. You know, I had a conversation a couple of weeks ago with a guy about data ontologies and how important it is to, you know, to build an ontology so that particularly for AI's, you know, gen AI consumption.

16:09So there isn't confusion about what different things mean. And is that part of this? Does this help in building that ontology? Of course. An ontology on a knowledge graph is the architecture and backbone of, I think, any real accuracy initiative, any hallucination reduction initiative. You really need to build it on that because it shows you the relationship between all the parts, all the sections of your data. And then on top of that, the metadata creates relationships that can't exist otherwise. So, yeah, of course, we have a graph, we have named entities, we have ontologies, we have custom taxonomies, topic modeling, all these things that are, it was part of the Web 3.0 back in the day of Tim Berners-Lee, all these really cool technologies like the OWL framework and whatnot and RDF and all these other things that just didn't take root because there was an ROI connected to them.

17:12but they were always known to improve the quality and the retrievability of accurate answers. It's just the investment wasn't there, the appetite to invest because it was very hard to prove that ROI. Now, with the speed of everything and running these models in real time and having all those conversations, you're able to see exactly when something is correct and accurate, something is not that helpful. Maybe it's kind of accurate, but not exactly what you need. And then there's the human in the loop there. And then things that are just flat out either wrong because they're pulled from bad data or wrong because the model hallucinated an aspect of the question and the prompt wasn't clear enough.

17:56And how running this, the shelf system on your knowledge base, how is there some metric about how much you can reduce hallucinations? Yep. There's real time monitoring of answer quality and the improvements that you're making using the platform. So yeah, there's a real-time answer quality dashboard where you can see exactly the number of inaccurate answers that have been produced and the source of those, and that you could either fix, filter, or enrich a bad answer, essentially, the data that's connected to that answer. And you fix and enrich in real time, or does it get shunted off into a queue that somebody else is working through?

18:44Yeah, it just depends on the organization and how they're set up. So someone can fix it right there and then, especially if it's something as simple as there are two documents that are very similar, like two products. Think of health care, like a health care plan. They're very, very similar, except some numbers that kind of mean everything. And the titles are even very similar. So you would want to put a named entity on that, telling the LLM that these are actually two different products. And that if you don't know the answer or you're going to try to pull from either of these, ask a clarifying question or essentially create a rule.

19:25When this particular plan is used, it follow these steps. And if this particular plan is used, follow these. But it's a way to have more control over the outputs of your Gen AI answers. Yeah. How does somebody get started with Shelf.io? It's pretty easy. There's three steps. First, you would connect any of your data sources. So connect your SharePoint, connect your OneDrive, connect your Google Drive. We have dozens of connectors out of the box and a connector framework. And then two, it's going to start running and identifying issues. So you have that dashboard. And then the third step is you get to fix, filter, or enrich and start impacting the answers being generated.

20:09And is this, how do you charge for this? Is this a sort of a pay-as-you-go? Are there different tiers, I imagine? Or do you sell seat licenses? It's a consumption model based on different tiers. So it's the number of pages processed, essentially and uh our starting plan is at like 3 000 a month and then it goes up from there yeah uh you you recently did a survey of uh of companies on on where they are in their gen ai journey uh specifically around the issue of uh data uh can you give us some of the insights Yeah, sure. I mean, one of the things that is very interesting, and we heard this from a major implementer at the Microsoft Ignite event.

21:11And the quote is, six months ago, I wouldn't know I had a problem. Today, this is very relevant. So I think one of the things that the survey indicated was how fast the market is moving. The adoption curve is just, it's moving a standard deviation to the right in six months, which is crazy. See how fast things are going. And that another thing that was really interesting was the number of active Gen AI projects. So over 70 % of the organizations have active GNAI projects, not in planning, but active projects. And then of those, the vast majority are still in the POC stage. And I believe Satya Ndodal said at the Microsoft keynote that 10 % of the market right now is ready to go into production in 2024.

22:12But in 2025, that number is being forecast to be 80%. So the speed of this market is blazing. And the other key takeaway I was surprised was that a lot of organizations do realize that they have unstructured data quality issues. They understand that the fine tuning and trying to chain different technologies together isn't the long term solution. It's a solution to trying to get something work right out of the gate and very quickly, but it doesn't scale. So I thought that was another really interesting takeaway from the market. Another thing I can say was the pure scale of most organizations. So over 80 % of the organizations polled had more than a million documents and files.

23:07Yeah, I remember that. Over 51 % of them had over 10 million. So just the breadth of that problem of how to know what you need to address first and how do you even try to manage such a huge number? You can't do that manually. Yeah. And what was happening before solutions like Shelf? uh i mean was it you know you have a a million documents but maybe you know it's kind of like my my office you know there's stacks and stacks of documents it's really only the top inch that that is active the rest are you know virtual archives uh yeah how yeah how does uh How does that relate to what you're talking about?

24:03Well, I think it's a complex problem, right? Because there are some cases where, let's say, take customer service. Customer service has a manageable amount of documents that are serving inbound questions coming from the customer. So the agents are only using thousands of documents, not millions of documents to answer questions. And when it's thousands, you can actually curate and fix the problem. You can't get 100 % accuracy, but you can get darn well close. So in that particular use case, which that's where we were before Gen AI, you can actually quickly, relatively, get to a trusted knowledge base.

24:50You can get to a trusted, secure body of knowledge, corpus of knowledge to support a use case. Now, then there's technologies that are like federated search or semantic search, which go across an entire organization. And I think the users of these technologies have a sophistication because they know that they're using it. and when it pulls something, the document that they're not looking for, like whenever you're doing, let's say a Google Drive or Dropbox search, you're like, no, that's not it. You're mostly using your human memory to identify whether it's the right version or not, but they're not particularly accurate because they don't know what the right version is.

25:35So the larger the corpus of knowledge, the larger the document base that you're trying to work with, the greater the problem is and the harder it is to solve. And that actually goes into what we advise the market to do. And that's crawl, walk, run. You shouldn't try to consume the entire beast out of the gate. We have companies that are like, oh, we're going to give you 1.5 petabytes of data to quality assure. And it's like, we can do that. But what are you going to do? What are you going to do with the output? We're going to identify tens of millions of issues. Do you have people that are going to be able to fix tens of millions of issues?

26:11A better move is to identify what use case you're trying to address. So, for example, it's an onboarding use case. Okay, great. We have an onboarding use case. We're going to make it easier and faster for people to onboard, to self-serve, to learn all the core company cultural knowledge and processes and procedures. And, you know, it's very similar. Then you identify which documents, which folders, which files that you actually need to support that. You port it over, you put the correct permissioning around that, and then you manage that because that also changes over time. It's not like you make this effort to get clean and accurate data and then you're done.

26:52No, because time passes and new products are launched and new people join and new directors put in new policies and procedures. Things are constantly changing. So for it to be accurate over time, you need to be monitoring it in real time as well. But to start with something small, to make it work, to get that confidence, to get people trusting the system, and then to secure it, and then to scale into more and more use cases. This is what we really see working. And this is what we see a lot of the consulting partners that we work with. This is what they're advising to. Yeah. And beyond RAG, would this work for creating a clean training data set so that you can, because that's part of the problem with hallucinations is there's so much bad data in the training data.

27:49and the models are really looking at probabilities based on what's in the training data. But if you got rid of all of the anomalies in the training data, then presumably, and I've thought about this with the coding assistants, if they were trained, because the first ones were, it was almost an accident. right uh they they train up this massive llm i think it was gpd3 and they realize hey this thing can actually read code uh but yeah can you talk about that about you know sort of training new models from scratch with better curated training data and and how that would have would apply to shelf.

28:44Yeah, there's two other use cases that I think are really interesting for your audience. I mean, there's the actually curating and pruning and caretaking the data that you want to use. So if it was 1.5 petabytes, then that use case of identifying the worst offenders in that obviously will have an impact and will be valuable. There's no doubt about that. So that's certainly an option. Anything that you're doing as far as data quality management before you're trying to do the actual AI task is going to have downstream benefits. And then additionally, the autonomous agents. So if you're creating experts on certain tasks and chaining them together, the data that they're consuming also benefits greatly from being pruned and caretaken before it goes into being automated and expanded very quickly.

29:41So they're really fixing the data and investing in the data. It just has downstream benefits for so many different use cases. It's why the structured data management space is so robust. And there's so many hundreds of companies there and huge businesses like Snowflake and their ecosystem and Databricks and their ecosystem. Because people understand if you invest in that, you're going to be able to use it in whatever use case that you want downstream. So you're kind of fixing things at the very source. And then where you pipe that is wherever you pipe it to, you're going to have more value, you're going to have more accuracy, you're going to have more trustworthiness.

30:18So it's the same model for unstructured data. The more you invest in fixing it upfront before you use it for whatever different use cases, the more valuable those use cases are going to be out of the gate and the less work you're going to need to do later. It's kind of like in the software world, having a really good specification versus not, and then designing something and building something and then finding bugs later, and then having to fix the bugs, it's like 100 times more expensive than just having designed it in the front. In the data management space, it's the same thing. You just fix things upfront, you save a lot of downstream headaches.

31:00And it's easy for me to say, because people have urgency to get something out of the gate and it takes time to do that. And that's why we kind of want to address that at both places, at the responsible AI and the data cleansing. But also, hey, we have a Gen AI project right now. We're having hallucinations. We need to improve the answer accuracy. You need to be able to do both. You need to be able to straddle both worlds because different companies are at different stages. And even different Gen AI initiatives within a company can be at many different stages. And you need to be able to walk with the market and enable them at whatever stage they're in.

31:37Yeah. You know, we were talking, you were saying, you know, a million plus documents and to focus on a particular problem or use case first. Is there, and a lot of that, those documents would be stale at this point of the million or 10 million documents. is there still value to be had in that so-called dead data, data that hasn't been accessed for a long time? That's a great question because we've heard some people say, just throw away all your data. As a knowledge manager, as my background of where I came from, that like sends my, you know, shivers down my spine because there are best practices, lessons learned, and key strategies.