Investment in AI has become mandatory for keeping up with competitors. Yet when incorporating AI into a business plans there is a disconnect as to what AI is does. Here, DataWalk’s CEO Paweł Wieczynski discusses graph AI, and how it uses graph data structures and machine learning to analyze relationships.
Interview by Sean Reynaud
WBJ: With graph AI, how is it different from the regular black box AI? What comes first and what comes after, the graph or the AI and how do you explain this to someone who is non-technical?
Pawel Wieczynski: Well, it's actually very interconnected. So, graph enables feeding the AI by first and foremost connecting all the data sets and then feeding the AI with the data. Thus, this is a big part of what I would say is the back end of AI. But also, Graph enables different types of operations that AI systems benefit from. A lot of AI techniques are purely graph-based. You cannot perform many algorithms, such as graph algorithms, without graph capabilities. A lot of types of AI are pure graph. Some types of AI benefit from feeding data, then finding contextual relationships; and when you said “black box” I think that’s the biggest difference when dealing with your typical systems. Graphs are a way of dealing with real world data because everything is connected, there are multiple layers between us and different data points. Graphs enable humans to understand different data sets, making data available for the human brain, and this is called human augmentation. The difference between black box and graph-prepared AI is that a human can understand all the operations being done behind the scenes, even if there are trillions of objects, but the human mind simply cannot comprehend all of it. Graph is a way to present it in such a way as to simplify data, so that humans can be active operators and not something like in Minority Report, where we’re told, “Here’s the result. Go kill that guy.”
The current way, which is an early stage of AI, is what I would call a revolution. A lot of people are still not using AI. It’s like in the 80s, when computer drafting came into use. Some people continued to draw on paper, but they’re obsolete now. But with regards to AI, the majority are still using a very narrow function of AI. This is very effective for very small problems, but with bigger problems, you need to be a more active user of AI. Otherwise, you will be replaced.
Well, I don’t know about the speed of AI development. But graphs are not only a way for humans to interact with AI and manage the system, but also a way to tune large language models. You must know that Chat GPT has a lot of hallucinations, and these hallucinations are something our customers are curbing. We are using a context-driven system that takes the LLM results, matches them with factual data, and tries to recognize where the hallucination is. It ,therefore, speeds up the adoption of LLMs. It makes them more efficient. So, I wouldn’t say that graphs hinder the evolution of AI; they speed it up. Also, at the data-feeding layer, one of the problems AI vendors face now is that the data they have to feed into the system is scarce. So, companies spend hundreds of millions of dollars on content. In the enterprise world, where the data is there, it’s siloed. The biggest issue is to put that data in one place and connect it and enable it for different AI models. This is something that, at the foundation layer, graph analytics enables.
Of course, the same goes for enterprise customers. If it's the government, they have limits on what data they can process, but it still has to be in our system. For instance, we must have military-grade security, military-grade permission schemes and all these logs and traceability. This prevents misuse of data that one agency or another doesn’t have authority to access. Or in the case of the corporate world, there are different limitations. Access to data has to be tracked. We need a high level of transparency. This is something that is equivalent to a lantern in a dark alley. It speeds up things. You can walk faster if there is light.
Oh yes, of course.
No, I think it will be eliminated. We already have some tests with our customers, and when applying the proper techniques, you can significantly reduce hallucinations. By the way, LLMs are just one type of AI. But there are around 20 types of AI according to Gartner. Neural networks, and graph algorithms are a few of them, according to Gartner. (Edit: Gartner, Inc., global research and advisory firm)
In the future, if you want to be an effective person in your trade, most of the trades will require the use of AI. You have to learn to use the ones that are applicable to you. But it’s usually not going to be just one. Most of our scoring, for instance, for fraud detection, are based on multiple AI models and rules that operators put down. For instance, the operator puts down that he/she is not interested in this or that, or people from this area. Combine parameters with some mathematical, linear models and then we have a more accurate AI model – according to our customers.
Well, from the very beginning we wanted to create a leading product that gives customers value, that is ten times better than others. We were one of the first in graph analytics, which is a market that is forming. There is a very narrow set of competitors in terms of data platforms that enables graph analytics. There are only two. DataWalk and Palantir. So because of low saturation it was easier for us. An example would be the Department of Justice - they were one of the largest users of Palantir but they were fed up with the limitations of Palantir not only because of cost, but there were other operational and legal issues as well. They said we need more vendors, because having a monopoly is bad for us. As part of a consortium, we won the frame agreement and the first project we won at DOJ was to replace Palantir in one of their sections, which was called the Money Laundering and Asset Recovery section. This was one of our first customers and, frankly speaking, in order to grow you have to push yourself beyond your own boundaries to build a better product and a better solution for those customers. And with each customer we won, from DOJ, to DOD, to Homeland Security, Department of Labor, we sought to grow with government contracts. And once we were mature enough as a product and company, we were able to penetrate into the commercial market, which tends to be more sophisticated.
Well, the truth is that the government sectors may not be as up-to-date as commercial sectors, facing more constraints in resources. For instance, analysts in government roles might find their compensation differs from those in similar positions within the private sector, such as banks, where the financial rewards can be significantly higher. Additionally, the process for acquiring new resources in government settings tends to be more detailed and extended, which means that initiating and completing projects might take a considerable amount of time, sometimes several years.
The reality is we had to start with a daughter company in the US, and the daughter company had an executive team with a sixty year old veteran of this enterprise space. One of these executives created a number of startups that were later unicorns, and even created a category that is called “utility storage.” So, having people like this in your corner was important to even be considered. Palantir has on its side a set of four-star generals and top politicians. So we still sometimes lose contracts to them. E.g. in a specific contract process, which went through several stages of a Request for Proposal (RFP), we reached the concluding phase with great anticipation. When the bids were reviewed, ours was significantly more economical than Palantir's. Despite this, the decision ultimately favored Palantir.
The way we get our foot in the door is finding those customers that are fed up with Palantir, or require a different agility level that we can provide. Or, some customers require self service because Palantir tends to take over and outsource their customer service operations. But we provide a platform, meaning we don’t even want to see your data.
Yes, and because of that, we need clearances for all locations. We have a small team of 20 with the necessary clearances to access facilities and train government employees.
With one of the top 10 Police Departments in North America, we implemented an algorithm called "Find Path." By integrating it with some business rules, we were able to reduce the investigation time for murders from months to hours, and in one case, to minutes. Handling cases involving millions of records was nearly impossible through conventional means until ten years ago. Now, we offer a solution that provides significant value to our customers.
Exactly. But the issue is that organizations have a lot of data, and they don’t know how to break their silos. That’s the issue. It’s not that banks don’t have the data: of course they do. Graph models reduce friction and allow for daily modification of data sources. Short iteration cycles allow for faster processing. Agility is a key driver of competitive advantage, and this is what our customers value the most.
