Latest / Reframe First / Yuval Yatskan, CEO of ePlane AI and AeroGenie | Revolutionizing Decision-Making with AI

Transcript

Yuval Yatskan, CEO of ePlane AI and AeroGenie | Revolutionizing Decision-Making with AI · Reframe First

• Oleksij Rak: Hey guys, welcome to another episode of Reframe First. I'm Alexey and let me introduce our today's Yuval Yatskan. Yuval is CEO at eplane.ai and also Yuval has been over 20 years in SaaS industry, worked for some big tech and nowadays eplane.ai. So hi Yuval, nice having you here. Yuval: Thanks so much for having here. Oleksij Rak: Awesome. Thanks for joining. So in the pre-call, we discussed a bit about what you do, what you build, and you build something. it would be really great to discuss and to explore further what you build. So it's something AI related. Could you tell us more about what you do there and how you came to the product? Yuval: Yeah, absolutely. Just maybe a little bit about E Plan AI to give the context. So we started our way way back when in twenty sixteen as a marketplace for the aviation ecosystem. And around twenty twenty-three we decided now with the rise of Genai and the need for more critical accuracy in execution to transition to point AI based solutions because we were sitting on like huge data that was very unique and thought that we can use this much better in this way rather than just being a marketplace. And since then we've been focusing mostly on the aviation ecosystem, so different types of customers, again saying with serving them with different AI-related needs, anywhere from dynamic pricing to certain automations, and different types of intelligence, inventory management, etc. But the most recent development to me is the what's very exciting and I think is is worth an entire conversation is what we're about to launch in a few weeks, a solution that we called Aerogenie. And this is something that far surpasses the the Biation ecosystem, has a myriad of other use cases, but it is our agentic decision engine. And what it does in a few words, and we can talk more about this again. I can talk about this for hours, is that it ingests data from different places. Now, obviously, with MCP, it becomes even more relevant that you can pretty much anything MCP-related or any shareable database or anything that you want to. upload it can ingest, but also automatic triggers, then unlike traditional LLMs and and commercial jobs that you have today That are limited in what they can ingest. I mean, like they have not only limitations, like say ChatGPT can upload up to 512 megs or so, but usually because they have to repeat that to reply with a very quick answer, they have to make some trade-offs, skip lines or skip pages, and approximate the answer. And because we built Aerogene using Mojo and a few other optimizations, this is thousands of times of much faster than the equivalent Python. So we don't hallucinate, we actually read the entire data, join all the tables. And the most interesting part, we run thousands of simulations per second to stress these different scenarios before we provide a very detailed analysis of the decision with a very clear audit trail. hashed by the way, 256 bits hashing so the decision could be protected and be replayed as as needed. And then finally, if people want to further execute and automate this this execution, we can provide that as well. And again, we're very excited about this, we can talk about different use cases, but as you can hear, I'm very passionate about this, but I I'll stop here just to see if if if if this is something that is interesting to talk about. Oleksij Rak: It absolutely is. I got a lot of technical questions, not only technical, but could you tell about some examples of how do you use it? In the pre-call, you told me about some legal case, it worked all over the problem. You did a variation Yuval: Yeah yeah yeah yeah I'm happy to give some some initial Oleksij Rak: and all of sudden it solves legal problems. Could you tell a few examples of what it does? Yuval: Yeah, it was actually very surprising to us because when we built this, we didn't really know what to expect. Just to a bit of more context, a a very large aviation rate company approached us to build the prototype, like a solution like that. And when we built this, and surprisingly, relatively quickly, We didn't know what to expect. We realized very quickly that it could do much more than what they asked. And just to give some examples that we actually stress test this and we've been testing this 24-7 in different ways, just to make sure that it again it's it's stable and does what it is, and we we keep on being surprised. Like we see benchmarks that really blow off the water out of the water, like things that that we we thought were almost impossible. But just to to give a few examples. We were asked to look into a legal case by one of our investors. And they obviously employed a very respectable law firm with they're using their own AIs and power legals and analysts, etc. Been working on on the case for a while, and We were asked to to to put the case, just ingest all the correspondence, all the relevant files, into Erogeny and see what we think. And usually, when you think about like, say, if you go like a deep search or whatnot with any of the chatbots, it would take hours. And even though it gives you a very detailed answer, it's still limited to the trade-offs I mentioned. It doesn't do any simulations. So long story short, Erogeny came back with about after like a minute or so. And in terms of tokens, it cost about $1.86. I still remember that number. And it said, look, you guys, this the team has. Been looking at the case in a the wrong way. It should be treated as a different type of case. We highly advise that you stop what you're doing and file so-and-so motion. Otherwise, if you continue in this way, the probability of you winning the case is this, the payoff is that. Everything with very detailed analysis and reasoning. And therefore, we recommend after testing this that you should actually treat this case as a different type of case. Go with this type of strategy, file these motions. And in if you do this, probability of winning is this type of probability, and the PEF is going to be something else. And we're like, okay, that's just a recommendation. Let's bring it to the legal team. They looked at it and they were like, Yeah, unfortunately, it did find vulnerabilities we we hadn't been aware of, and it makes a lot more sense. That's case number one. Oleksij Rak: Yeah. Sure. Yuval: Maybe just another one to show the versatility of it. another use case that a cosmetics related company came to us and said, Hey, we're about to launch this new product. And we tested this ourselves and also used other LLMs to test this and research. And we came to the conclusion that this product should cost $75 dollars. And again, same same idea. Can you please feed up this all this information into Erogyenie and see what it would recommend? So Eerogeny created its own engines, like one and own agents. Like so, once it understood the mission, said, Okay, I got you. Develop its own skills, checked, for example, in Amazon marketplaces and in other places, and said, Look, we disagree with with your recommendation, the competitive landscape. Does not justify anywhere near $75. It should be around $45. And here's the rationale. And eventually the the the the the firm or the the company decided to follow our recommendation because it was so detailed and the rationale was like step by step how and like the sources and how we it the aerogenic came to this conclusion. But point being it did not miss something that other LLMs and other researchers missed in the get-go and all this and again in like very like a few minutes. So that that also was something very surprising to us. Maybe one one last case is we also Oleksij Rak: Yeah. Yuval: looked into this to stress test some investing strategy, like a portfolio, how like the composition of a portfolio and what we should do. And again, like a two years ago, we predicted accurately a certain for example, a certain movement in Palantir stock and a few other things. It's like, okay, that that's very interesting. With again the the a combination of both machine learning and simulations. Oleksij Rak: Awesome. That sounds like a miracle, you know? But still, what are simulations? You mentioned that it can run thousands of simulations per second. What are simulations? Yuval: It's a great question. Maybe it it's worthwhile talking about simulations versus kind of a pure machine learning approach. So a simulation, and and there are I think the most known one is the Monte Carlo one, but there are other types of simulations. But the idea is that you take a data And you run an experiment multiple times and see where it converges. I'll give an example, like a few, like one, like two, two very known examples, and then other interesting examples I just read recently of surprising industries that are using simulations. So let's say that I have a fair die and I want to know what is the probability of getting any of the the faces, like say two, three, five, whatever that is. And assuming that I don't know much about probability, what I could do is I can like throw the dice once or twice or whatever. And after 10 times, if I say, okay, let's see how many times I got the number two as a percentage or as a proportion of the 10th throws. So it might not be one-sixth of the time because it's only 10. But if I repeat this experiment 1000 or 10,000 times, it's gonna approach one sixth. Another interesting example, if people want to approximate the value of pi, the you know the ratio of circumference to diameter, one thing that you can do is take a square, have usually people have like one quarter of a circle inscribed in, and then you can like throw dots. So imagine that you have uniform distribution. that we just like throw dots and then you count the portion of the dots inside versus to the total square and you can infer pi from this. The strength of it is that you don't have to train a model. So, like the machine learning approach is something like this: whatever model you use, by the way, both for supervised learning and unsupervised learning, but the idea, let's say that you have something, let's take supervised learning for example. Let's say that I want to forecast the price of a house. And one step one, I have to decide what features or what parameters I want to use for building the model. I have to build the model, feed in the data, assuming I have data, which is a huge assumption, by the way. And then I have to train the data again, like the training set, cost validation, keep on refining, and then I have the test set. And then when I have a new house that I haven't seen, and the model would tell me, okay, this the house should cost $1.1 million, like a point estimation. Based on the model I created, and assuming again that the model is not underfitting or underfitting, that's what he does. The vulnerability of this is that one, it presupposes I know all the parameters. Because if I find out that wait a second, the other parameters did not take into consideration, I'd have to rebuild and retrain the model, which is a lot. The second challenge, it gives me just one point estimation. yes, you can run like a chi-square analysis to or other type of things to explain what features contributed to the variance and and what explain explains that the result, but it does. And then in addition to that, a model can usually answer one type of or related types of questions. So if you have like, let's say you and I asking very different questions, we'll probably have to build different models. And that's just not scalable. And simulations, on the other hand, you define a certain function. Like, for example, if I want to forecast profit, so profit is like revenue minus cost, for example. And in this case, by the way, an LLM, if you ask a question, can interpret what type Formula and type of data you need to answer that question. So with simulations, you don't have to build a model, you just run the experiment once and then you repeat it multiple times, again, usually thousands of times, and see where you converge. And the the thing is, it doesn't have to replace necessarily a machine learning model, it can either come before or after or with it. Example. Let's say that I want to build the machine learning model, but I don't know what features to use. So I can probably run a simulation and then as a second step run the machine learning model on top of that based on what the simulation uses its parameters. And vice versa. Let's say that I have a machine learning model, I can run the simulations to see the probabilities that are associated with the results I get. Usually simulations would reply with here's the the lower bound, upper bound. And the more like what you call like the peak value, the most likely result you're gonna get. And maybe one more thing I'll say is that a few interesting examples that I read that actually that actually uses simulations are Formula One races. There's so many Oleksij Rak: Thank you. Yuval: interesting parameters: the tire pressure, the type of the time of the year, the weather, the the engine, the driver, so many other parameters. And they heavily use simulations to tell the driver, for example, when it should what lap they should actually go for a pit stop. taken into consideration different times, like do they have to position or not to position before the race? And something that you think, well, it's just racing, it's extremely accurate, extremely data driven, and they heavily use simulations because of that. Oleksij Rak: Okay, you got me really interested in this. That's awesome. Yuval: Happy to see the article I just read in the F1 magazine, just how they use simulation in that type of use case. Oleksij Rak: Okay. In your legal case, for example, how do you this simulation? So you've got, I don't know how many pages, hundreds of pages of legal documents. You need to load them. And what do you simulate? You simulate the outcome based on some decisions made or... Yuval: Yeah, so you t like oversimplifying a bit, but step one is ingesting all the data. And because we built what we built using Mojo and other optimizations, we can actually ingest millions of pages. Oleksij Rak: Okay. Yuval: and I've seen like millions of lines of databases without skipping anything. And we also have a module that use the the upload documents and translates them into structured data. So we can actually use that also as part of the model. And then run different stress cases, like for example, different parameters and like different legal cases that look that have different similarities or different commonalities with our case and see how they ended. Like who was the judge, like where it was tried, what type of court. Which state, etc. There's so many parameters. And what happens if you change the the type of the case to this versus that? And again, thousands of simulations that run per second. And you run you let it run obviously for multiple seconds, so you have a lot of type of data and and stress testing different scenarios like what if I do this, what if I do that? So that what if is something that LLMs are not designed to do. And maybe one thing, I mean I'm sound kind of like I'm bashing LLMs and chatbots. I think they're amazing. I think that they're definitely a leap, like it's a step function from what we had before. It's not just an evolution of something that just becomes a bit better. And they're amazing. But the way I see them, they're more like a proof of concept, meaning If you think of OpenAI, Anthropic, etc., they they build wonderful things. But in some ways, they build them as more like a proof of concept. They they look for companies and people like us to say, okay, what can you do with this? This is like a beautiful tool. How can you use this? There's some built-in limitations. For example, if I'm a Rolls Royce and I build engines, by definition, by design, I don't, I, I, I don't also build the planes or the cars or where the engines goes into. That's something that we need to do. So we took the and and by the way with Erogen you can use whatever LLM model that you like. Deploy it locally, on the cloud, whatever it is. But the point is we give the right environment for the engine to run in an optimal way. That's why we we use the other technologies that we that we put with it to run with this. So again, LLMs are awesome, chatbots are great for what they they were designed to do. But when they end, this is where we begin. Oleksij Rak: Yeah, awesome. So how do you make this mixture of simulations and usage of LLMs? At which point, which step steps in? Obviously, LLMs give the answer to any human readable format, right? But how do you make it? For instance, legal case, you've got hundreds of documents of the case. But besides that, you mentioned you need to take all the... potential other cases from around the internet, Collect them somehow. And then on top of all of that data corpus, run some simulations. So probably there is some LLM in decision making, what to collect, what to bring, what to get. You mentioned Amazon, so there should be some LLM that has written, I don't know, some piece of Python to get the data from Amazon. And then feed all that information into your simulation engine. which produces some kind of result and then LLM steps in to make the final judgment call. Could you tell us more about how it works underneath? do you combine all Yuval: Yeah, still. Oleksij Rak: this? Where your engine steps in after LLM ends, as you said. Yuval: So the the LLM again, I mean today LLMs obviously can work with agents and Oleksij Rak: Yep. Yuval: you can configure them and and you can build your own. But as I mentioned Oleksij Rak: But it's not one dollar, you know, it's not one dollar cost. Yuval: Yeah, but I I but but that that that's the thing because if you think about LLMs again in chatbots today as more like proof of concept, they show you what's possible. And they say, can you take this like what you see in the the in the storefront and and and build something with it? So this is where we are. So I think a a huge part of it is because we we built a lot of the technology using Mojo, and we also put a lot of time optimizing The the functions that go alongside. And we have a list of all the functions and how they compare to the Python. And again, when I looked at that comparison, it ranges between like 200 times faster to about like over a thousand sometimes faster. So that allows you to see more data. And then run these simulations. Now, when you ask how, so obviously the LLM can tell you here's what the, like, for example, in the simulations, how we should think about the simulation. Here's the the relevant data and the formula and the relation between those components. An interesting point just to pause for a second, an LLM will also will hallucinate because of the trade-offs. They don't have all the data, or they would skip, they wouldn't, instead of joining five tables, they only join three, for example. With us, we don't have to make these comparisons with these trade-offs, and we would actually flag out we can't answer this question because you're missing so-and-so data. So, actually, other than just agreeing with the user and say yes, of course, and approximating, we'll say, here's something we can answer. Here's like data that we're missing, we cannot replicate. So, here's the the vulnerability in the answer. And the the the question that you ask about. Building the agents, again, I think it also comes because of the speed. Because think about this way, I mean, like every agent that that runs today is subjected to the same trade-offs and limitations or budgets that the LM has. Like for example, I have so much time as an agent to return an answer. so much time to compute it and computing resources and so much tokens I can spend on it. So if another agent is dependent on my answer, that limits I have to approximate just by design. So we're for the most part we're not bounded by that by this type of limitation. So that gives you some idea of like the approach we took. And the the big surprise was that we we say okay what happens when you no longer have to to make this type of trade-offs. What would that look like in terms of the quality of the decisions, the answers? Because the way we thought about aerogeny, like ideally, it's not that you actually just put prompt. You can. You like the same way that you do with cursor and others. You can put prompts and you'll get the decisions. But ideally, we want arogy to ingest things automatically, make the decisions, and ideally also execute. Maybe just one example that might be relevant. My the the past few years I spent my my career in the cybersecurity space. So imagine you have a threat detection and and response system that identifies a vulnerability or even like say maybe there's a breach or something. So imagine that Aerogeny ingests this data, can read the logs unlike any other machine today. Again, like they read the like millions of off lines in in in so many logs and and decides exactly what the problem is what to mitigate and how to mitigate and then what actions to take based on on what it found. So this kind of end-to-end that's that's our vision but of course people can use it in like anywhere in between in just the thing and then use the recommendations to be executed by the Secops table the CISO whatever themselves but just to show that the the the system was designed to do everything from start to finish. Oleksij Rak: But we understand it right that it needs a lot of data, Yuval: If Oleksij Rak: Besides the case itself, needs to collect all the related data to be able to make the... to find this like the most possible statistical outcome of the simulation. Yuval: Absolutely. Absolutely. Oleksij Rak: So in case of legal cases, you take your case, but you also grab all the possible cases from the internet, similar cases, then... Yuval: Internet and and and and databases, that's legal database can be connected. Oleksij Rak: In case of logs in cyber security you got your log file but besides you need to have all the similar cases found over the internet. Yuval: It's not just the internet, but we can use your your own database, your own logs to see like say that you're company X and you already have threat detection or other like whatever firewalls, they collect data and you have it, it's just that it's like the data is is so big that people find it very hard to digest and analyze and s and synthesize. And now with with the advent of of strong LLMs and and technologies like ours, you can do things that you couldn't do before. So the day the companies in in many cases have this data. They they sit on trove of extremely valuable data. They just they don't Oleksij Rak: Have Yuval: have the capacity to to make the best of it. Oleksij Rak: Absolutely. Now that I think of it, and I had a lot of discussions with this travel industry and one of the big biggest concerns are like, not the concerns, one of the biggest, biggest, biggest like troubles is customer support itself. But the problem is, the question is that these companies sit on a lot of data, like 10, 20, 30 years of data. And all these cases that are potential cases that are happening nowadays with some certain client, they did happen before, over the last 10 or 30 years. And like the biggest problem for them is how to extract that kind of data, because you cannot just feed everything into LLM. You can use some rec approaches, whatever approaches you can structure the data somehow for LLM to use, but it's slow. And it requires a lot of work. From what I hear from you, you can just feed all the data in, in, in Errógini. and it gives the result in a matter of like seconds or minutes. Yuval: Yeah, and and the thing is Arogen does doesn't have necessarily to replace these systems. It can sit as the brain behind them. So it could Oleksij Rak: Yeah. Yeah. Yeah. Yuval: be, for example, that you have another chat support or whatever it is that you have. Again, there are other companies that build this kind of type of systems. But for the decision, the quality decision, especially in what, let's take your example, in what cases to compensate, what type of compensation based on so many other parameters, a simulation might be faster in many cases and give you more quality decisions alongside a machine learning model. And as you as you said, I mean, companies do have this data, and it's extremely exciting to see what could be done with it. Oleksij Rak: For sure. Maybe last question about the product itself is how did you make it generic? I can understand how to make it for specific case, but how did you make it generic so that you can throw in legal data, throw in some beauty goods, whatever else. Yuval: I I I I I love this question because if you think about this, think about like what w the the the way we thought about this. What what is the DNA of a decision? What does it look like decision look like? So Oleksij Rak: Yeah. Okay. Yuval: sometimes and and and people are looking for answers, not just this decisions. And maybe it it's worthwhile defining what we think of, like what what is a decision. And then I think from that you'll see why it it's so generic. So, an answer, let's say that before the the troubles were available, you would go like to your favorite search engine, ask something, would crawl the the internet like very efficiently, gives you like a bunch of things that you have to choose from. And if you ask a question, I have I see these symptoms, that symptoms now with LLMs. You can add your own context either through RAG or MCP, or otherwise, that you just put in the prompt a bit more that you can do, much more than traditional search engine. But the thing is, if you think about a decision, and this is where I think the shortcomings of chatbots, for example, is that you want to be able to replay. Like, for example, if you made a decision, like say two months ago to shut down this line or increase the the price of that. And you're asked to defend your decision. You want to make sure that the data that was used at the time that might have changed since hasn't been lost that you can still replay it the way it was at the time of the decision to justify it. You want to see a clear re reasoning, and again, some channels will give you like the sources and whatnot, but not as clear reasoning of exactly how what led to the decision to defend it. And when we we you thought about this and you want to be able to make sure that you you looked at all the relevant alternatives. That's kind of with the simulations, like the what if. What if the the fuel prices go here? What if the the yen de varies against the dollar? What if this, what if that? It it obviously would impact my my sales cycle, my inventory, et cetera, et cetera. So when you thought about when when we thought about this generically, we said, okay, so first of all, you need to be able to ingest data because that's That is like the main limitation. I need to decide what data to ingest and to be able to do it without sacrificing or trading off anything important. And the approximation that most LLMs do is great for like an initial direction and for many, many answers or many questions people are asking over the internet. Like show me my favorite restaurants here, and then where should I go? This and give me like a recommendation for a movie I should watch based on my criteria. But if you need to make a complex, critical decision, complex that it has a lot of data, it's a very hard-to-make decision, critical in the sense that you have to defend it and a lot is at stake. I can't just rely on LM to do this itself. And that type of topology is something that is very generic. If you think about it, it could be, you can inject any any context or any data. You don't have to train anything because again, we're using simulations. You don't have to have a pre-trained model for this. We can build skills, we can build agents, but then we thought, okay, what what if Aerogenic can build its own agents? Oleksij Rak: So you kind of make a lot of small hypothesis and just continuously adjust the solution, right? Yuval: Yeah, and and eventually w so what when we built it, we we we you built it in the sense that there was no hypothesis was unique to to the the company who asked us to to initially build this thing. And then it's just a matter of curiosity. Wait a second, if that's the case, let's test it. Let me you know I'll give you another interesting use case that we just tested because Angela and I, Angela are our CTO, we didn't know if it would work, we just tried it. So My brother is is has another company in the defense industry and he has a huge deck, like 120 or so slides, like very complex deck. That he tried to shorten with with different LLMs, different services, it it didn't go anywhere. And it's like, you know, see if you can make out of it, anything out of it. We fed it into Erogenie. First of all, before anything, it it sped out a very interesting chart, a flow chart, like almost like a stream of consciousness of how it understood what was presented to it, the the various personas, the various motions, etc. And eventually said, Okay, how here's how I would recommend breaking it down. And he did break it down into separate slides shortened with the the personas in mind something that again it we didn't tell it it just had to think about the mission and break it down themselves and in in you know interestingly enough we asked erogenie how would you market yourself we tried to take you take you to market and we also can you please also create a deck for our investors and again for a deck for example it looked by itself on our website took the the the the the colors the the the themes the look and feel the fonts and everything Built a deck and for the first time I didn't have to change almost anything in the deck. Like the content was spot on. I mean, like I just I I liked the certain design ID better, so I changed like like 5% of the design ID there, but it was pretty much good as is. The pricing strategy, the go-to-market, with what to focus, etc., Eerogenie ran on itself to get this type of answer. Oleksij Rak: So it's Yuval: It's yeah. Oleksij Rak: amazing. Is there still some LLM behind it? Or there is not? Yeah. Yuval: Yeah, of course. So the LLM is used to understand the intention. So for example, if you put a prompt or something is being adjusted, LLMs are again extremely critical to interpreting the intention, right? Oleksij Rak: to illustrate all of that. Yuval: Like what exactly did you mean? Like for example, maybe you you and I want the same thing or something similar but use different words. So an LLM is really good at understanding your nuance and colours versus what I meant. So that's kinda where we use LLM in the initial understanding. And then from then we say, okay, we got you. Then we we go and again, even how to devise the the the simulations, what to consider, what not to consider. LLMs are extremely useful there. We just again we we the execution part of it, they're limited in certain ways that we're not. Oleksij Rak: Yeah, awesome. So how do you go to market with this tool? What did it suggest to you? Yuval: so first of all, again, the pricing strategy is interesting, the per user per month and and how to go about this. And and again, in a few couple more months, I think it should be available to to the public. Right now we're just in the process of launching it, creating the usual the YouTube, the LinkedIn pages, the splash pages, etc. a few videos, and surprisingly, by the way, we just uploaded one video that already had we we we barely boosted it to be honest, like a few bucks a day or something like that. And and we had of those like about 30 or 40 percent conversion to subscribers, which to me is very interesting for a single video that is so so friendly. It's like six minutes. It's to be honest, again, I I our CTO put it there, and I was like, okay, let's see. and we will focus on at least initially on business users that are no strangers to to chat bots. They need to know how to use charts. They need to be, if for example, you told, hey, use MCP to connect so and so services, we want people that are comfortable with this thing. And therefore we're gonna put it in in at the same time, we'll open into like a focus group and then to others as well. Happy to again happy to Oleksij Rak: I am comfortable with subsidies. Yuval: to do this again. The goal is to put it in the hands of as many people to try it out. I should say obviously it should have also a free tier that people can can try it out, see for themselves. But the idea is is really to Test it against like and we hopefully we'll have some testing again with what will show this how it actually works works compared to to any chatbot that you have today in the market. Not to again, not to make chatbots look bad, but to show this is the extent the chatbots are great and we take it from here and run. So for example You'll see the approximations that they make in different types of questions or decisions that you have to make that relies on data. But maybe one simple case that is easy to try is let's say you have a huge Excel or CSV. Again, most accounting systems, for example, or others can produce huge CSVs. And an LLM would have to skip lines. Even though I think a CSV can have up to, if not mistaken, like two million or so lines, an LLM would not read two million lines. If you upload a PDF that has 200 pages, even though it's well within the like what, like 200 or 300 kilobytes at most, an LLM will skip pages. Oleksij Rak: Yeah, it will not fit the context windows. LLM will write some piece of Python code. Yuval: But but they're not designed to do that. That's the point. I mean, I I I don't fault LLMs for that. They're or chatbots. I like LMs are amazing. They're amazing. And you're now I don't know if you saw Oleksij Rak: Absolutely. LLM? Yuval: the the recent I think NVIDIA announcement that they're gonna have like a trillion or so parameters. And also follow closely Android Inc. that that covers these these kind of things. what we see today that the evolution is is so fast but I still think it's directed in in certain limitations like for example at at least at this point simulations are not part of it because of the the token and and CPU limitations or memory limit limitations and LLMs again that they're good for certain things but they they're they have a specific usage. Oleksij Rak: Yeah, lamps have limited context window and our lamps are statistically pre-trained and whatever they were trained on, they can produce Yuval: Yes, yes. Exactly. Oleksij Rak: something. From what I hear from you, you can run, like you can not retrain, but you can take some corpus of data and try simulation on top of that corpus. Yuval: Yeah, again to just upload like a very heavy XN and see if you can ask questions. Oleksij Rak: Yeah. Awesome. Yuval: Yeah. Oleksij Rak: Are you considering some B2B clients or is it rather B2C? Because it should be quite interesting for B2B users, for business. Yuval: No, absolutely. So I I mentioned the cases of like cybersecurity. We're gonna try. I have some still like Oleksij Rak: Yes. Yuval: strong and and warm ties to this industry. and try put it in in their hands and see what they make of it. And again, it could be used in different ways, anywhere from the actual patching or or taking the actions or telling you based on the logs what could be done, anyway to automating these decisions. Again, working with whatever systems you have today to both ingest but also execute autonomously. We already have a chain of restaurants, like a a large chain of restaurants that that wanted to take a look at this and and use it for different ways. Hotels as well. I see a lot of B2B again. We talk about law type of or or litigation type of usages. I know law firms today use something called co-council, and they obviously have their own regulations, but I'll give you this maybe one interesting use case that we we found out for for law firms. The accepting of new clients. Usually you have to have analysts and paralegos. Let's say that I'm a new customer and I want to submit my case to a certain law company, I'll give my all my correspondence and the files and the data. Someone has to go through this and say, okay, is this a good fit for our firm? Can we take it? Do we have the capacity? Will it make sense for us to take it? Is this in our expertise, etc., etc.? Imagine that instead of paying people to do something like this, you ingest it almost automatically. Erogeny goes for this in a few minutes, it gives you. The right rationale and the right decision what to do with this case. So you don't have to replace anything they're using today, it's just really more the ability to handle more customers at scale, for example. Not to mention, by the way, again, our bread and butter is deviation space. I definitely see a lot of of use cases. Again, this is such a complex and and regular industry that there's so many decisions they need to make, anywhere from like CFOs to to operations, to go to market, and what parts to have and not to have, etc. All these things, all these decisions are imperfect today, and something that can definitely be augmented and and much improved with a technology like aerogeny. Oleksij Rak: for sure aerospace should be interested, especially since I know they are running on very thin margin. any kind of operational efficiency for them is a big boost in their margins. Yuval: A any decision that you want to take, like for example, should I open a new airport in so-and-so location? Should I open a new like say that I'm I'm a developer? Should I buy an office building? How tall? or should I build a hotel or what all these type of decisions think about? They're very abstract, but they do require a lot of data. And they do Oleksij Rak: Yeah. Yuval: need like you need someone smart to much smarter than us to to to consult with and brainstorm ideas. That that's where I think a technology like this comes to life. Oleksij Rak: How big is the team standing behind this product and this development? Yuval: Surprisingly small. Surprisingly small. We're we're we're we're Oleksij Rak: I guess so. Yuval: we're fewer than than than ten people in the development team. And Again, I in my experience, usually when you want to build something like this, and maybe I think it's it's worth saying a few words, but I think this is a very interesting question. In my experience, you'd you'd have to have a product management team, not just one person, a team of engineers, marketers, etc., to build this. And why? Because at the time, way back when engineering resources were very expensive, and it took a lot of time to build and test and refine with. Customer that you actually got what you wanted. But today, with the beauty of AI assistant coding and other ways to accelerate coding, engineering has become, again, if you have the right resources, relatively quick and and relatively cheaper compared to the other options. So it's really more a question of like what should I build? And how do I decide about that? And how do I converge from different hypotheses to the right solution? Oleksij Rak: especially nowadays, you cannot build this kind of system with a big team. So if you want Yuval: No, and and and and Oleksij Rak: to move fast, you need to have a small team and empowered by all these two modern tools. Yuval: It it's almost like like close to zero fat. I mean like I I I've been in different organizations, work with wonderful people, but it's a very Like almost like futuristic type of feeling to work with with people that you can see every person exactly what value they bring and and how they augment themselves with AI today. And by the way, when we are thinking about hiring people, that's exactly how we're thinking about. Can this person create a false multiplier by bringing themselves into the equation? Because just like bringing one more person, however tainted they might be, that's one thing. But if this person has access to knowledge and and and how How to use different AI and other systems, by the way, to really like for example, marketing. Used to have, again, I manage teams of like sometimes 20 plus people. And then a bunch of agencies to execute on different things, anywhere from like videos to campaigns, et cetera. Imagine now that you can need like about one-tenth of the people to do the exact same thing, and probably an even better quality. Oleksij Rak: Awesome. What are typical challenges that you have in your development? What are usually very challenging? Yuval: It's really it's really w where to focus, to be honest. It's like really, given that everything you want to build, if you build too much. Then the goal to market becomes a bottleneck, like selling it, marketing it, supporting customers. So in many cases, we'll have to say no to customers, that they want Oleksij Rak: Okay. Yuval: something that can actually make us like a decent amount of money just because it's going to be like a one-off. We're trying at least today to stay away from being a project-based company and become, because I think like Palantir is doing an amazing job. I don't want to become another Palantir for that matter. We're trying to democratize some of their capabilities, for example, but not at the expense of having. Having to support, like you know, tens of different custom deployments that are not scalable. I think scalability is is the key. And for us, what to say no to is is I think is is still very challenging, and where to put our our bets. And and just to give you some criteria, where I'm evaluating today like new directions to take to market, I have two things: one, the applicability, like Is it really relevant for like a small niche or like a relatively bigger niche of customers? And then is the deployment fast? So if I need to be bogged down like in in months or or a few like weeks of deployment, I'd rather go for something that I can deploy it almost instantly. Again, Aerogeny, once you let's say that you have access to it, can say, Hey, you value got it from here. You can connect to MCP. If you don't know something, you ask Eerogeny itself. Hey, can you tell me exactly how to where you mentioned a shareable data link? How do I get a shareable data link? I have this Oracle 19c. How do I do this? Boom, boom, boom, boom, boom, go do it. So you don't need me and you don't need to like to have me to deploy anything per se. And this is to me like where the the the future is going something that the customer is kind of self-empowered to use it almost from the get-go and and run all the customizations all they need to do themselves. Oleksij Rak: Yeah, sure. That's a great problem to have of saying no to your customers. Yuval: Yeah. It's true because the the the thing is sometimes your capabilities could become your own curse because then you end up with different customers. They ping you and everything, but each has its own need. And when they have a challenge or you need to upgrade the system, you have to do it separately for each customer. And that to me is not scalable. That goes again Oleksij Rak: Yeah. Yuval: to have like a small and and lean team that can execute at scale. Oleksij Rak: And that's great that you have your focus and you execute on it. How do you see the nearest future with this kind of tools? Because these tools are emerging and they're quite powerful. What do you think about it? It's Yuval: You I love telling the story before we Oleksij Rak: a philosophical question. It's not business related or anything, but the future is changing and it's changing rapidly. Yuval: You're absolutely right. You're absolutely right. It's very hard to predict something that it is constantly changing. And again, I'll I'll tell you a quick story. So we had a version of Aerogeny about maybe a year and a half ago, a bit even more, that was very innovative at the time. Like you can have, like, for example, graphs and and charts done in in line within your your your questionnaire, including certain machine learning analysis, like a PCA analysis and and this and correlations inside aerogeny. But We had a tough time explaining to people how to use it. Like it was really relevant for very sophisticated data analysts. And and at some point we kind of close like after like a how like six months or a year, we we lost that advantage. Now most LLMs already have these capabilities, and it's very hard to catch up. And I think you'll see like releases and the staggering rate. I like that means that. you'll see more innovations from the likes of Anthropic and OpenAI and Google, etc. And computation is going to become much more efficient. We're not doing anything that others are probably not doing today. we don't we didn't invent module, others are using module. We just again took different technologies, put piece it together in a certain way. But I do think that LLMs would offer more, they will be able to ingest more in the future. But at the time, at the same time, I think the business model for the time being would either to offer LLMs for specific industries, like for example, what Microsoft does with I think it they call it Dragon, the for the healthcare, or like special types of LLMs that are pre-trained for certain things. And there are other companies that I think that's the specialty. But nonetheless, I still believe that there's gonna be room like like people like us, like companies like us, that would always take LLMs from where they stop and extend it further. Because again, you you can't solve for everybody everybody for for everything. And I definitely also think that the agentic AI usage is gonna be much more prevalent than it is today. People to be like in the earliest time of ALF, people just ask questions, received answers, and then moved on. But people becoming more technically competent and comfortable exploring things, especially that again, the LLM companies are doing such a great job making all these technologies accessible. Again, MCP. Imagine that before you need to have like a very talented data engineer to do certain things. And today with MCP is you can you can do so much more yourself. And again, think of the transition for from RAC to AC to MCP for example. All these things I I expect them to to to continue happening in the future. Oleksij Rak: Yeah, for sure. Yeah, a lot of them will be becoming more capable and quite likely they will become specialized. That's what has Andriy Karpatyn mentioned a lot of times in his interviews. But what he extended them with is this statistical simulations on big corpus of data. This is something where I do not even see them moving. This is something completely different. the train statistical models on big corpus of data and that's it. And while you take your unique corpus and one simulation, that's very powerful simulation, combination for me, it seems. And from what you're saying, it works. It indeed is awesome. Yuval: Most companies that use simulations, they have a well-defined problem. Again, I gave the the F1, the Formula One example before that. So they they already know that their parameters and their focus on that. But if you're thinking of building something that is like multi-purpose, and how can I Oleksij Rak: Yeah, generic. Yuval: democratize the simulations across different use case industries? That's not something that the F1 folks are thinking about. Again, they care about Formula One. And chatbots may look into this in the future. I I definitely think it's a possibility, but they'll have to be specialized in so many other things because again, how do you solve for agents limited budget? You can't approximate you can't run thousands of simulations per second with the the current technology and you can't run thousand simulations with that that eventually gonna cost you hundreds hundreds of dollars per per query. That's that's where where the the option would be today if you were, which you can't run it with LLMs today. Oleksij Rak: Layer 2 generic and they are targeted on mass market. Well, you are targeted on some specific, solving specific problems, deep solving these problems. that's awesome. Yuval: Exactly. Oleksij Rak: That was really like inspirational, great discussion. Yuval: Well, I'm I'm I'm I'm glad you found it useful and again I'm happy in in the next few weeks it's gonna be limited beta I'll I'll give you an access if you're interested. play with it so what do you make of it? Oleksij Rak: Sure, I'm your first beta tester. I'm your first user for sure, as soon as you open it. Awesome. Thanks a lot for this conversation. Is there anything else you would like to add on top of what we discussed before we... Yuval: No, I again th thanks so much for inviting me. It's it's it's been great. I think we're we're we're living in a very interesting time today and Oleksij Rak: That's for sure. Yuval: I I think maybe maybe one positive message is that it it opens creativity for pretty much anyone. So if you thought before, let's say that you had an idea for in a project, a startup, whatever, and like, where do I find my my technical person and and how do I create a team and marketing? Today a team of one can do so many things. And again, not like maybe like even like a small team of like a couple of people. can do what I think in the past you need to have like twenty or thirty people do. And and people once they they realize that, I think the the opportunities are are endless. Oleksij Rak: It empowers you extremely. got fed up with one of the tools I'm using and I just asked Claude two weeks ago, can we reproduce it? Claude was like, yeah, sure, why not? Two weeks later, already, prototype was ready in the first night, but two weeks later, I am close to like put it as an MVP to some users and run some traffic towards that product. Yuval: So exactly. Oleksij Rak: It's so weird. That would take otherwise a big team, engineering team, product team, designers, all these people working together, marketing people. Nowadays it's one person, know, and two people. And Yuval: Yeah, it's my cousin so far. Oleksij Rak: it's between regular work, know, so everyday work. Yuval: Yeah. You know, but but but surprisingly, I still find the the human element still critical. Like for example, take marketing. Yes, our marketing team, small as it is, can do amazing things, but I still like to to talk to a human at the end of the day and and see can we like can people empathize what we're doing? It's like a machine can get you so far, and I would assume can really emulate human emotions really well. But again, psychologically still deep inside, I'd like when when we have marketing, I there's a person I can talk to so and again maybe that's gonna change in the future in never know like 10 years or five years from now may look different but I I think you definitely need people in in critical places just because again we talk to each other right now imagine that it's gonna be two machines talking to each other instead of the podcast there's still a huge element of like two people talking exchanging ideas Oleksij Rak: Thanks a lot for the discussion and good luck with release and let me know when you're in. Yuval: Absolutely. Again, thanks thanks thanks for having me. Oleksij Rak: Yeah, cheers. Yuval: Thank you.