Transcript
Hi Mike Matchett with Small World Big Data and we are here today talking with Kognitos about AI and how to really make good use of it. It's a different way of thinking about AI and a different way of thinking about learning. As we know that our normal AI projects initiatives may have stalled out or even failed due to hallucinations or missed expectations. There are great uses of AI coming out, though, that are much more, uh, I would say focused on getting the right things done. Kognitos is focusing on helping automate business processes, so just hang on a minute and we'll get started. So. Hey, welcome, Binny. Welcome to our show. Thank you, Mike, for having me. Uh, so, you know, going just before we even get started in that, just going back in time a little bit. Uh, you know, we have some crossover points in our past. You know, I covered Hyperconvergence you were involved with Nutanix, I covered storage. You were involved with IBM flash systems. I'm now doing a lot of AI stuff. You're doing AI stuff. So this is just an incredible journey of things. I don't know if there's anything to that. Uh, but, um. What what have you been bringing to this AI market? What is your past, uh, sort of gotten you excited about in doing AI today? Well, I've always worked on, um, things that are solving current pain points in the enterprise market in large customers. Um, so my experience has been around that, uh, I'm an algorithms guy. I'm a computer engineer. I've learned a lot of programing languages in my career. And when I looked at AI, this was like five years ago. Alexa was working in our house and Siri, and I was wondering, can we change the language in which the whole world programs these machines? Can we, for once and for all, say that no longer? You need to understand any other programing language. All of that will become the new assembly and you'll just use natural language just like you're talking to Alexa. And that the more I worked on it, the more I realized the time has come. And that's what essentially we're doing. It's creating a new paradigm of bringing AI to enterprise where it can become deterministic. Just like Python is deterministic, English can become deterministic, and you can get away from a lot of the hallucination issues that are plaguing the industry. A lot of things going on there. I just want to talk about, like, I'm a Ruby person at heart, Ruby on Rails going way back. Python was sad to me. But, um, you know, I'm really kind of surprised you didn't just start out by saying, hey, it's all Python, so we're going to use Python going forward. You're just like, no, no, we have llms today that can understand English and help us make some translations to a more symbolic way of understanding the world. So let's push double down on that. Uh, so when you say, um, deterministic, uh, to avoid hallucinations, maybe you could just explain a little bit more about how English could possibly be deterministic, since it's full of, uh, which we say ambiguity and miss reference and, and, uh, implication. Yeah. So the language itself is not what makes a deterministic or not deterministic. It's the person interpreting the language that can make it. So when a kid is crying for a balloon on the table, a kid is just crying. It's highly non-deterministic language. But the parent says, hey, do you want this balloon? Do you want the teddy bear? Which one do you want? First tell me, then I'm going to do it. So now they are making it more deterministic by asking the right questions. So today, Llms are fully capable of asking all the right questions, but we would like to make sure the LLM is not doing the work. It must be something as deterministic as Python that does it. And the ambiguity that natural language has. You can use reasoning models to tease out all the ambiguity out, and then give enough instructions to an interpreter that we have built for natural language so that it knows that. Yeah. While you said add that to the table, that means that, not the other. That and the table means the invoice table, not the vendor table. So and then we can encode that, uh, behind the scenes. So what you're seeing is an English program, but all the details, all the disambiguation is actually learned behind the scenes. Also in English, uh, as side notes. Yeah. It sort of seems like, uh, you're taking AI llms the way people think of them today as I prompt it. And I get this generated thing back, which is almost half art, and saying, no, we're going to turn this more into a conversation until we nail down exactly what you're talking about and do it, but that won't take very long. And the beauty is, I don't have to now learn Python or another language, right? So the biggest. Mistake to it. Yeah, yeah. Biggest mistake right now that's happening is I write an English prompt and out comes some Python that I don't even have time to look at. Right? But I have to trust it. The only way you figure out is it runs in production. Then you realize, hey, it's doing the wrong thing. Instead, our philosophy is different. You can do a prompt or a chat in English. What it should generate is an English SOP that you can read and it is concise. Now if you say no, no, I want something different, then it changes the SOP. If the system says, hey, this is ambiguous, it does not generate something in Python and without showing you it. Actually, since it's English as code, it will, like Grammarly tell you in your SOP you wrote this. You know what? This is not that clear. I'm going to rewrite it for you. Tell me if you agree. So we are updating it back so it's no longer about English as prompt. That's the old way. It's English as code where the long term Maintainable asset is actually this English program because prompts just go away, you lose them after you write in. Python is the long term thing. So the time is up. It's no longer Python, it's English code, right? And when you say English is the coding language, I mean you really have to sort of reframe my mind to think in English. Conversation is the coding language, not a one time do what I say and I turn around kind of thing. So it's now finally, after all the conversations you do, there's an SOP that is generated. And that's the long term maintained thing. If you look at all the businesses, right, when there are a lot of people, we create process to manage people. And that process is these are the things you need to do when you're onboarding a vendor. These are the things you need to do when you're paying for an invoice or a claim in a healthcare. And that process is the living document. We have just made that English document directly executable. Right? Right. And this is a way, um, you know, to avoid, uh, what happens with generations of of older code or older software or things gets patched and and grown and accreted, right? You get this kind of, uh, growing technical debt of things. And so what you're actually saying is we're going to we can use this to just keep focused on something that's not really going to grow. Uh, it's going to just keep stay on on target. Right? Yeah. There are a couple of things that plague the enterprise today with the current way of doing automations. One is that somebody writes an English PRD, but then what gets implemented initially is aligned with the PRD. But after that it gets updated. And, you know, there are fixes. Nobody updates the PRD. So anyway, there's a disconnect between what the business sees as English and what's implemented in code. We solve that problem saying that the English document, the PRD is the executable code. So it's always correct and up to date. Second thing we do is there are so many edge cases in the in in the real world. There are so many ways you can fail to reconcile your bank records. Now, the learnings for all of that we are tabulating as troubleshooting guide on the side, which is also English. So imagine the separation of happy path and troubleshooting both in English and they are maintained over time. Your AI models can keep on evolving. You know one day you're using Gemini, the other day using OpenAI, whichever one is the best one. But the idea is to manage your business through documentation, not by training AI models and fine tuning them, and then realizing, oh, there's a better AI model, just like we do with humans. Humans come and go process documented stays, and that is the driving wheel for the business. And that gives also deterministic output, because whatever is written is exactly what gets executed. Ai is there to help you write and help you understand how to handle edge cases. Write the troubleshooting guide as well and follow that. All of that is based on human input. Human has to approve it and approve it for that one time, or approve it for all the future instances. So basically, we're trying to get into a place where the human is in control, not necessarily human in the loop all the time, because that doesn't save the human much time. And it sounds like if I'm just putting this together in my head, you're talking to some other people who are doing automation with AI and doing that. You know, they're trying to accelerate the ability to do integrations, or they're trying to jump ahead and apply learnings, things they've learned elsewhere to the problems you might have. So if you say ask, you know, ask Alexa for something, you know, Amazon's model might have learned how to answer that question from somewhere else and do a better job. But here what we're really doing is saying no. Every business has got a unique set of operating procedures. I think you're calling them and you're kind of negotiating with the with the with the, uh, the system here to create a set of operating procedures that you both understand, and it can go implement them reliably and you can maintain them as a, as a, as a set of directives. Yeah. For example. For example, if a business user says, I want to create this purchase order in SAP. Now that's a very high level. So our system will go and say, okay, let me first look at what SAP systems you have given me access. If none. Prompt them. Give me access to a test environment. Give me access to production environment. Then we go do, uh, discovery of what custom objects are there, you know, what customizations the customer has done, and then come up with a suggestion to go and refine what they wrote in English and say, you know what? I'm actually going to do this, uh, in your SAP system. Do you agree? And if they say yes, then that becomes recorded as the standard procedure backed by a buy a deterministic execution engine that will run it, that execution engine, when it runs, it doesn't run on top of an LM. Lm only comes in when something abnormal is going to happen. This is something new. So that's the creative engine. And the runtime is deterministic. That's why we call it neuro and symbolic. Symbolic is always deterministic. I like this because what you're really saying is we've looked at what AI is good for at least today's AI and what automation and robotic process automation, whatever the other words for it is good for. And we're using them each to their strengths, and we've made them in the middle and have this negotiated contract with our system, which is which is portable and can last with the business and evolve. And this is how you can also stay current as AI keeps on evolving because the business process is remains the same. It's the AI's intelligence that keeps on changing. Right. So so we can we can always go back to our AI and say, hey, we want to develop a new process. We want to accelerate this process, optimize it, fix it, increase it, whatever, increase the reliability of it, and change the process in the middle. But we agree on that process. I'm just curious, though, you know, how long does this take to get going? Uh, if someone is just starting out to say, like, all right, I've got some business processes in mind, uh, and I want to get started with Kognitos. Am I having, like, months and months of conversations with this chatbot to get something that actually looks like it's doing the right thing? Or, I mean. It's just like what you see, there are two, two scenarios. One is you could have, uh, a standard operating procedure already written. Then you just cut paste that into the system and you already have something you can start, you know, running it, testing it right away. It'll ask you some clarifying questions and off you go. The other one, if you don't have a standard operating procedure, you're just saying I'll just walk you through one example. So have a chat, walk through, and then after the walk walkthrough, it says, I understand what you're doing. I can make a repeatable process out of it. So that's. It takes 5 or 10 minutes to start and then you can put it in production. Over time it will run all learn all the edge cases, learn all the tribal knowledge that happens over, you know, the number of weeks or months and slowly figuring out exactly how your business works. Fields here on the AI implementation front for lots of companies, by saying you don't have to go and build your own inference engine, you don't have to, or your own training models. You don't have to go build up these huge databases of Rag or a lot of necessarily a lot of MCP things on spec that you don't know what's going to happen. You can just start by automating processes you already have in hand and moving from there to see how how better you can make them over time as involves. Correct, correct. All the, you know the AI complexity about fine tuning rag all these terminology, these things will be behind the scenes. It will be invisible for the business user. They frankly don't care. They just want to get to the business outcomes. Sure, sure. It's understandable. Uh, so there's a lot going on here. Uh, it'd be great sometime to have you come back and give us a demo. Live demo of this, uh, on something. But in the meantime, if someone wants to learn some more information about what's going on with Kognitos, it's Kognitos with a "K" you probably have a website. Is there something particular you'd recommend them as a learning path for this? Yeah, absolutely. We put everything on the main web page. You can go to Kognitos.com, and right there there are some use cases that are the most common people try and they're getting ROI out of those use cases immediately. Because of the benefit of putting deterministic AI into motion, the development lifecycle for changes businesses are evolving. So for example, manufacturing, logistics, healthcare, retail, anything that's document heavy. You can look at those examples on our website and, uh, go from there. All right. So, uh, hallucinations are for chatting with the AI around the water cooler. Uh, but when it comes to your business processes, if I get this right, you need to cross that boundary between the neuro and the symbolic and make it deterministic. Uh, and agree on those standard operating procedures with the computer, with the machine, so that everyone's happy and can move forward. Uh, this is this is this is a mind expanding stuff. Uh, this conversation today. Uh, so, uh, definitely worth worth looking into. Wish you luck on Kognitos, and everyone check it out. Uh, any last recommendations? Did you give everyone Binny? Just start. You know. Just just start. Just just get going. All right. Well, thank you for being here today. Uh, and thank you for watching.
