 Hi, this is your host of Limbhakti and welcome to another episode of TfL. Let's talk and today we have with us Nima Nikban Co-founder and CEO of Karnataka Nima is great to have you on the show. Hey, thanks for having me It's my pleasure to host you today This is the first time she and I are talking on camera and before recording as we're talking You know the company has been around since 2010 as you said Which also means that not only you folks predate a lot of modern shiny technology like Kubernetes You have seen also a lot of evolution and today we are primarily going to talk about like Generative AI and the fact is that a lot of industry they have been using AI already You know security has automotive healthcare, but generally I has kind of opened a new kind of like floodgates So there is so much to talk about today, but before we go there, I would love to know a bit about the company itself What was the idea behind the company because you're a co-founder So why you created the company back in 2010 and then we look at 2023 How do you have seen the evolution of space and the evolution of company itself? We started in 2010 as you mentioned and we actually were part of a DOD research project and that goal of the project was Consume hundreds of different real-time data feeds and then be able to give a query capability To analysts to data scientists to developers to you know quickly be able to deploy stuff into the into the field And at the time, you know You know no sequel was all the rage Hadoop was all the rage There was still the big warehouses that the legacy warehouses like the terror data's They all really had a huge amount of trouble dealing with real-time data and being able to do complex query They could do a few things they could do a lot of pre-planned things where they did a lot of indexes and you know Supplemental data engineering to make it happen But to be able to really find that needle in the haystack to really be able to do whatever you want Across all that data as it continued to flow. There was a really, you know, no good solution So we were there as part of that program and we had this idea, you know, hey The you know, the GPU is something now in 2010 even where the you know, it is a tremendously powerful device So, you know, databases have been designed with one thing in mind for 40 years that compute is very You know scarce resource and you know, you should be able to You know organize your data prior to asking your query questions so that you use as little computer As possible now the GPU computer is an abundant resource. So what if we flip the equation on its head? Let's make a database that's for you know Allowing data to continuously stream in to be able to write any query you want without data engineering and leverage all of this Abundant compute in a distributed way In a way that allows the developers the data scientists They know to ask any question they want and get back responses quickly with up-to-date data And that was the basic premise behind Connecticut And we started building, you know in 2010 and you know, we became the you know analytic engine for You know the speed layer for that program where we were sitting on top of actually, you know, a cumulo And and you know doing all the analytic temporal and spatial work for for that project over time You know, we we went into you know, larger large enterprises like USPS, which is one of our first flagship customers Where you know again, they were Doing something that really required a new type of solution. They said hey, we put sensors on every mail carrier We need to be able to analyze this in real time And it looks like you're the right solution for that. And you know, that was one of our first major wins and from there We really focused on becoming a You know real-time speed layer for the modern enterprise So are you mostly targeting the public sector government entities? No, I mean, we have we do have a big DoD Customer base, but we also have financial, you know large large banks, you know large telcos Anyone who's you know trying to take advantage of real-time data from sensor and machine where they want to be able to do advanced analytics that you know potentially fuse that real-time data against, you know, historical data sets to be able to query it without, you know, any type of Limitation and have it be up-to-date have it be performant. That's really, you know, our sweet spot. If you look at modern word It won't be wrong to say that we live in a data-driven world apps can come and go all the time But data is I mean as we say data is the new world, but you know, that's that real asset How do you've seen the evolution of data and when we talk about real-time data? I mean, of course we can look at EVs, smart cards I mean everything, you know, we are collecting and generating a lot of that and sometimes a lot of things have to be done Real-time. So I just want to understand the evolution of real-time data, but sometimes what happens is that there are a lot of technologies Which continue to co-exist where some technologies evolve, you know, so they transform I just want to understand the role importance of real-time data real-time data has is really starting just starting to kind of Get its foothold in the modern enterprise. I think, you know, really it started as You know, very simple calculations that just looked at, you know, the the latest, you know record that would come through and have a, you know, have a very simple, you know You know threshold-based, you know rule-based, you know analysis And it's evolved to a place where it's really the, you know, the the ability for your You know, your data team your your your enterprise to understand what's going on in the business in real-time So to be able to query it not, you know, not just in a window But to query it, you know across all of the activity that might be happening over, you know Several months or a year and be able to fuse that against other data sets That's really where real-time data is going so it started out as something really simple like hey Let's look at the last five minutes of data and run these rules, right to now where it's gone to a place where this is How we understand what's going on our business in real-time We need to be able to query it not just within the last five minutes But within, you know, the past year or multiple years have it be up to date be able to fuse it against other data sets that may be historical or real-time and really Be able to be creative in what we can produce because that's going to give us that extra advantage and be able to, you know Give us that, you know, that next level of efficiency that everyone's striving for It won't be wrong to say that we live in a data driven world How have you seen the evolution data, especially real-time data with the emergence of new use cases And if you can also talk about What do you think is the rule of real-time data in this modern economy? Yeah, I mean, you know, I would say like, you know As if I'm looking at it by vertical like, you know, healthcare is one place where we're seeing, you know, real explosion of New ideas on how to leverage the data that they have, you know, including things like, you know, claims and things like that where You know, traditionally that may have been just static data that was, you know, analyzed You know, on an ad hoc basis you're getting to a much more advanced place where people want to, you know, create, you know Huge huge amounts of infrastructure that is analyzing that data in real time, you know, extracting value from it Extracting predictive insight from it, you know, I think As far as like, you know That's by vertical but as far as like, you know, how people are leveraging data, you know, it's gone from a place I think in the Hadoop era where everyone was just like Let's just drop it in and collect it, right? And we'll put it all on, you know, back then HDFS and we'll figure out how we use it later You know, I think we've gotten to a place now where, you know There's a lot better understanding of of the data inventory of a given enterprise And things like, you know, Databricks and Snowflake have done a really good job of kind of being that You know, authoritative data store for all of the data and making it easy to do kind of the static or batch like reporting and analysis And, you know, I think there there is now, you know, that next That next iteration where we fit in where it's, hey, we've we kind of understand, you know How our data plays together, right? But now because we've analyzed it on a static basis And now we got to put this into action where, you know, I can give my You know, my C-suite a single pane of glass where they can understand everything that's going on in the business And not just have static KPIs that are that are, you know That are being maintained but allow them to double click, right? And be able to ask the questions that they that we didn't think to ask beforehand, right? You know, that's really where that ad hoc query capability that speed layer Capability to do that on real-time data. That's that's where we fit in And I think that's a lot of, you know What we see is the evolution of the like the database space on that kind of capability level, right? Where, you know, we've done what we've done so far has done a great job at dealing with static problems Or dealing with questions that we know people want to ask and now it's getting to okay I've got something where the the data is constantly changing And I don't necessarily know all the questions that need to be asked in the moment And I need to be able to deliver that capability and that's where I think there's still space for innovation Where else do you see the scope of real-time data beyond some of the use cases that you mentioned Or some of the use cases that is Traditionally associated with yeah, I think there's there's two things there I mean one is the expanded role of vector search and embedding generation, right? And you know a lot of that is best run on the GPU And that's something that we've been focusing on is you know making it really easy to do You know large-scale real-time embedding generation doing doing very powerful brute force GPU vector some vector similarity search And you know that that alone is something that you know the GPU is you know really You know unique in and it's ability to do but then also you know It's ability really to brute force calculate, you know in brute force process data Really fits in well with what what we're what we were just talking about which is with LLMs People are going to be able to ask questions and generate code that that you know sequel or whatever it might be that you know, you know Codifies that question and they need they want to have an answer quickly And you might also even have a world where LLMs are talking to each other And you know, they need to be able to say okay, I need to know, you know, how many packages are You know available for delivery in this area, you know, whatever it might be You're going to see that need to be able to run ad hoc query You're going to see that need explode, right? And again, the GPU is really purpose built for being able to do you know massive processing without having to pre-define and pre-plan queries and indexes You mentioned vector search Can you talk about how our vector databases and knowledge graphs being used for Insights on structure data versus the more common language use cases? Yeah, I mean so like with with knowledge graphs, you have, you know The ability to find more Entity correlations, right where you have in the generative case really in a heightened Entity, you know, an entity detection and classification capability from text now knowledge graphs are then being used to To take whatever the user put in tie it to what entities might match via, you know A vector search and entities detected and then pull out new relationships, right? And then, you know, again find more vectors off those, right? So with with knowledge graph and vector search together, right? There's just a A much more deterministic and A powerful way to deal with, you know, structured data to power, you know, enhanced LLM workflows, right? So, you know, I think with generative AI and thinking about LLMs You're going to get get to a place where it's, you know, much more advanced Fixed workflows like, you know, you can think like auto GPT where There's going to be need to be able to Connect to and query knowledge bases and those knowledge bases are going to be able to, you know Need to have multimodal capability, right to be able to answer things by finding Related entities via knowledge graph or, you know, to be able to pull out related time series data via vector search or, you know You know, there's there's going to be that need to be able to converge all these analytic disciplines And tie it with the ability to do it quickly as the data is updating because the LLM is going to need to, you know Have just like if you were you, you know a user you're going to need to be able to have the latest data and have it be You want it to perform quickly so you can take your next step without any contest Genitive AI is the hottest topic these days Talk about the impact it's going to have on your industry and your products Yeah, I mean there's two parts for us. I mean one is powering vector search, right? And, you know, we can do, you know vector search In ways that are unique as far as giving you added flexibility With the results of results of a vector search or the input to a vector search But also being able to do that in real time, you know, so not indexing and being able to do it at scale And leveraging the GPU That's something that's unique to us and, you know, really You know, that's something that kind of fits into the the generative workflows of today The other, you know, the other big part is having a language to SQL capability baked into the product So yes, there are other products that, you know, try to act as, you know Kind of an intelligence layer for multiple databases But we believe that it's also incredibly powerful to have that capability in your database so that your Language to SQL generates the SQL that takes advantage of all of your unique capabilities around time series spatial and graph So we bake that into our product, right? So you can, you know, literally say like execute question Give it a natural language question and have it return SQL for you and that and that's baked in the product, right? And that goes back to what we see as the future Which has been brought on by by this generative AI revolution, which is a User expectation that they can do ad hoc query just through natural language, right? And be that LLMs themselves are going to be, you know, you know, exchanging Needs and queries through natural language So having a database that has that out of the box that can bring all the different analytic capabilities to bear We think is going to be Any unique offering In general, what kind of trends are you seeing especially When we look at generative AI and what does it mean for kinetics? Yeah, I mean, you know, the the generative space is is evolving, you know, pretty rapidly I think, you know, there's definitely pretty much an understanding that the vector search capability is Now widely available across all databases, right? So, you know, six months ago, there was a handful of databases that had um, you know You know, I would say But you know full full featured vector search capability now, you know, pretty much everyone has it, right? But You know from there, you know, we're also seeing the need to be able to take that next step of okay Okay, you have vector search. Can you scale your vector search? Can you scale it with data latency? So all those are the things that are being, you know, competed for now and and that's, you know Where we see our our advantages being able to do it at scale being able to do it as You know, new vectors are being generated in in real time um, and and then on the actual, you know, large language model side, you know, there's I think a a Explosion of different opportunities going on as far as you know, first it was around. Okay, we want, you know We want to make co-pilots that, you know, help for code generation or Help for support, you know, but I think there's um A lot of innovation yet to be had on how do we use Generative AI and the large language model approach for enterprise and enterprise data, right? And that's where we're still I think in the early days How to separate hype from reality because I mean I cover all these emerging latest technologies Some technologies are just like a blip on a rotor But then many others which transform our industry if I ask you when it comes to generative AI Is it a hype? Or is it the next? transformative Devolutionary technology like docker container Kubernetes or the Linux kernel. Yeah, I mean, there's definitely a mix of both right like, you know I think when like chat gpt first came out like people thought out there's like this All-knowing all powerful lm that can kind of do anything And I think, you know, we're realizing, you know, especially as people understand lm's more and how they work that, you know There's really a limit to to how much they can do and you know the you know How you can use them to accomplish certain things right now that being said They still are, you know a tremendous capability that if you weave in correctly for certain Tasks or you know for certain, you know Really thought out workflows can can really deliver some amazing capability, right? So I think there's a lot of that still to be figured out of like, okay How do we how do we, you know weave this in into delivering a A still, you know, wow factor capability for an enterprise, right? But it's not quite like oh There's this like, you know, how 3000 that, you know, just I stand it up and it's gonna automatically know everything about my You know company and and know, you know exactly what to to do for all these difficult scenarios, right? It's it's got it's it's definitely got it's like, you know Certain uses that you know if used correctly and and orchestrated correctly can deliver still some next level next generation capability But you know, it's it's it's going to take some time to to to cultivate that and figure that out But it's also not Just like stand this thing up and you know, now I've got this super Super intelligent brain that can figure out anything out, right? So, you know, I think I think anyone using chat gbt for for more than an hour or two can kind of see that, right? So, um, you know, I think it's it's definitely something that It is it is a little bit hype, but it's also there. There is there is a there there And then we look at these emerging technologies. It's very important for organizations to while stay on Trusted reliable technologies. They should have their roots firmly planted But they should also dip their toes In new technologies. How do you enable your customers to maintain the balance? Sure. Yeah, I mean, you know, one thing we really make easier is, you know, gpu adoption gpu orchestration and You know being able to take large-scale enterprise data sets And bring them to the gpu in a way that that's, you know, well understood. So, um, you know, besides us also bundling our own large language model like we make it very easy for people to, you know, do their own exploration and data science in in kinetica, um, you know As far as large language model fine-tuning if you really wanted you can even do that in kinetica or we can host large language models, but You know, really, you know, what kinetica is about is being able to take real-time data And feed it to the gpu for your data science teams or your analytics teams in ways that allow them to orchestrate it very easily So, you know, this this allows folks to who want to get their feet wet and try things out You know, they can try it on kinetica at the same time, you know, there's a whole as far as large language models there's a whole ecosystem of Uh, you know cloud stack provider cloud stack providers that are looking to you know, roll out their own So I think there's going to be people trying different things like, you know, obviously, you know, your big enterprise is not going to use you know, the standard open ai large language model, but you know, there's now in the open ai large language model on on azure, right and You know bedrock is coming out for aws. And so, you know, there's going to be um, you know different Different products and different comfort levels of using those products. And so we're really focusing on how do we support that Where, you know, we we can help support, you know You know doing large language model work by yourself or support working with other ones that might be hosted So we're really working, you know, also nvidia is coming out with their nemo cloud llm as well um, and we're focusing on, you know, how we interrupt with that. So um, we think there's going to be a kind of a whole wide variety of of uh, generative llm options and we're looking to really support all like we certainly don't think the answer is hey You know use kinetica l kinetica for You know making your own llms and hosting in there that that's not going to be the answer The answer is going to be, you know, much more varied Neema, thank you so much for taking time out today and of course talk about kinetica But also kind of, you know, broader discussions around generative AI. Thanks for all those insights and I would love to have you back on the show Thank you. Thanks for having me and love to be back on again