The Role of Smaller LLMs and Synthetic Data in the AI Revolution

Hey everyone. Welcome back here to Textron tv. Very happy to introduce my next guest to you.

I'm gonna do my best with his name. His name is Ramesh Para Miss Para. I believe that's correct.

It's the best I'm gonna be able to do. Ramesh, welcome to Text Drunk tv. It's great to have you on here.

Thanks, Anna. So you, you are the CTO and CEO of a, a company called Land Grind. It was formerly called Wind, if anyone has, has heard of it.

But let, let's start with your story, Ramesh. I, I'll be honest, we don't often see the CEO and the CTO title with the same person. Usually, you know, they Right.

They kind of separate it, so I'm sure there's a good story there behind it. Yeah. But let's hear, Okay, let me, uh, let me start.

I'm gonna kill two birds with one stone. I'll give you an introduction to myself and that'll also answer the question about the C-E-O-C-T-O. Great.

So, uh, my first job, uh, out of college, uh, I was one of the first five product managers of what is now known as Windows from Microsoft, really? And, um, yeah, so I was, uh, the only technical project manager. So I was, uh, supposedly in the marketing function and product management, but I originally hired in into Microsoft as an engineer.

And, uh, so, you know, there, you see, there's some clue there as to why I handled both the hats of the, the business as well as the, I get it. Yep. You've been doing this, you've been doing this a Lot.

We're doing That's right. You know, you, you said what eventually became known as Windows. What was it?

It wasn't OS two, or No, It wasn't there. There might be people who might, uh, take exception when you say something like that. No, I know, I know, I know.

I'm old enough to remember those days. Plus we're in Boca Raton. Yeah, they actually in the, in the old IBM campus.

That's right. Which is now called the Boca Raton Innovation Campus. Yeah.

They have the room where Bill Gates Yeah. Signed the DOS license with IBM. Wow.

That's Pretty, You could still have your press conference in there, but that must have been quite heady times, though. It was heady times. It, it was called Windows nt.

So they were at that Time Oh, sure. Four five point Windows. The Windows Five one was when, right?

Yes, I remember. Yeah. So, uh, my, so one of the things I did was, uh, I was the only guy who knew much about Unix and Microsoft, and, uh mm-hmm.

My first job in the product management role was to, uh, I did a keynote support for my boss who was doing a keynote for 2000 people. And I demonstrated how you could move a file from Windows to Unix, and I got a lot of applause for that. So imagine, Well, back then, that was like a Houdini trick.

That was a Houdini trick at that point. Yeah. And, uh, yeah.

So, yeah, so, you know, so that's where I started. That was my career. My, that was my first revolution, the networking revolution.

Mm-hmm. And, uh, uh, you know, and, uh, there were a lot of things I did for at that point, you know, and an interesting story was, uh, again, uh, tying back to the business and technical side, in order to get Windows Andt adopted in the government, uh, we had to actually label it Unix. So it was very strange.

I got a guy to start a company and, uh, who took Windows Source code and built the right layer on it, and we were able to certify Windows as Unix, so technically Windows. Oh my God. Was Unix.

So, um, isn't that funny? I mean, it's, uh, it's crazy. Well, I, I, look, I started a hosting company 1996, right?

I think it was 1997 is what you're talking about. Maybe. Yeah, I'm talking about 95.

Actually 95 was the that, Right. So don't hate me, but we were running Sun Solaris on our service. That, and Microsoft made a reception.

We were in downtown New York City. They called it Silicon Alley, right. Not Valley.

And we, and Microsoft made a reception and we had the Microsoft Windows and T server team there. Yeah. And he, they put the bit, you know, back then Microsoft.

Yeah. They put the big push on you, right? Yeah, yeah.

They put the big Yeah, they put the big push on us to switch over. Yeah. I said, I, I said, can you really tell me that NT Today is better than Soliris?

He, he said, it may not be better today, but I promise you this, within three releases, it'll be better and more popular than Solaris. That's right. Yeah.

And he was right. He was a hundred percent right. By the time that three versions of NT later came out, I think Solaris was no more.

Right. But Yeah, so yeah, It's an interesting time. I was in, yeah.

I was in the middle of all that. And so that was, uh, my, my first sort of revolution with networking interoperability. Mm-hmm.

com revolution. I started my first startup. com, and you could ask, uh, or answer questions on anything.

I remember this still. Okay. It was a top 100 website.

So we had, uh, yes, It was Top hundred, uh, traffic. And, uh, and you know, uh, I, I read an interesting article from you while I was looking you up, right? There's this, uh, gravity defying article that you just put out a couple Yes.

Days ago you could have taken your article and just, you know, I could have run it apply to com. Well, that's what I said at the end of the article. How do I know this?

'cause this was my life. I lived that life. I did the startup.

We did an IPO, we went bankrupt. We, I, that's right. I did the whole, I, you know, look, I thought Alta was gonna be the end all for That's right, Isn't it?

There was Lycos, I dunno if companies like Lycos site it excited home. Yeah. Infoseek ask me, Me.

That's right. There was a lot of them ask me. Oh, there was a lot of them.

And then, you know, boom, Google came out and Google, Hey, it's funny. You know what, Ramesh, I too bad we're doing this over video. We should be drinking beers and talking about this.

I think so, yeah. Next time you are here, I'm there. I'll, I'll give you, I, I promise you, you.

But so what happened after Ask me, So ask me, we actually transformed it in an enterprise software company and sold it. So we were one of the few dot coms that survived and transformed and in some sense succeeded in the end. Uh, this is my, uh, third startup, uh, with, uh, what was called Wind until a couple of days ago and now called Lang.

And, uh, so this is, uh, not my first rodeo on, on a bubble. Yeah. My second rodeo on a, what people call a bubble.

And, uh, so in the last few years, wind Doc has been a company that was focused on database modernization. So what we offered was a pure technology, pure product, no services, uh, other than support. And what the product did was to create, uh, environments, database environments, modern ones for DevOps teams and testing teams and dev teams to use for DevOps for testing, uh, continuous integration and deployment.

Uh, we first came, uh, to FAME as the first company that ported and delivered Docker on Windows. So Docker, uh, Docker, port of Docker source code on Windows. Sure.

And we did that. What was unique was we did that before Microsoft and Docker did it. And that earned us kudos and some not, not kudos from certain players that I shall not name.

Yeah, No, I, I remember that. I, I've, I was here already doing Textron. Yeah.

I, I think You're familiar with that. And then, uh, so that got us some. And then that is the core of our business is containerized SQL Server and Oracle, um, uh, so far until a couple of years back.

And then we were the first to also deliver containerized clones in Oracle databases as well a few years back. So we were trying to bring like modern DevOps technologies and so on to, uh, the old sort of world of old classical SQL databases. And then we kept innovating with that.

And then we then, uh, um, uh, started to do things with synthetic data. We got recognized by Gartner for synthetic data as well as CICD for databases. And so we kept modernizing and kept delving into the database semantics and syntax and all that stuff, schema and so on.

And then now what we're doing is we're essentially, we've launched a beta that attempts to remove the friction between enterprise databases and large language models. So things like, you know, cloud sonnet, GPT and things like that. And that's very recent.

And, uh, we believe that we are the first to market in a product that actually allows the LLM to completely understand the databases and allow the users to ask or get answers from the databases on anything. com, which was you can ask or answer questions from people other people. And now, um, we've got an offering there where you can ask our answers for anything from enterprise database, from, From, From a full circle.

Absolutely. Well, it's not the people anymore. Yeah.

But, you know, let me, I, I just want to, so one of the issues we're seeing pop up with this ai, you know, a lot of people are saying LLMs have reached their, or will reach their capacity soon. We've scraped all the data there is to scrape publicly anyway. And that, and, and, and actually for most tasks that people are going to use the AI for, we don't need everything under the sun models.

Right? We don't need these giant Frontier LLMs. Right.

What we need, you know, people talking about RAG and Vector database and, and, and graph and, you know, small language modules if you will, s SLMs. Right? And that, that's kinda where the, where the future is that and synthetic data, because we're gonna need that synthetic data to con if we're going to continue training these LLMs.

Right? Right. Um, so it, it's very interesting, you're right in the middle of this, it sounds like, where, hey, you don't have to use a, a Vector database.

You don't necessarily even have to do rag. Right. You Right.

You could use your regular good old fashioned SQL Postgres, whatever you are using. That's right. And, and your LLL and your AI can be trained on it.

Can you do inference on it? That's Right. Yeah.

I think, uh, let me address, uh, those points a very interesting points you make because there's a ton of debate out there and there's no clear answers. And, uh, you know, so our, our, our bet is on the fact that, uh, the LLM or any language model will essentially replace RAG and it'll replace data engineering as a function. Mm-hmm.

So it is, it is, it's, it's a very sweeping statement I'm making, but we haven't gotten there yet. Obviously we're just beginning there. The challenges with RAG is that you're essentially taking data and doing the work of the LLMs or the S SLMs, right?

You are actually vectorizing, why do it when there's like a, a much more gigantic operation in, you know, meta or Microsoft, everybody who can do that for you with the LLMs. Now, the interesting thing for us is, which is really separates out, uh, separates us out, is that we allow the LLM to be the captain of the plane. We don't use LLMs as co-pilots.

Our target audience is customers who are saying, look, we might, we can look at LLMs as co-pilots and assistance, but this really is about transferring the sort of the ownership to the LLM and having the og the land grant product to validate the results of the, what the LLM is driving towards. Now, in answer to your question, a specific question on do we need LLMs with like these large understandings of everything in the world, our view is that, uh, the answer to that is actually yes, which may surprise you, but it's because in order for the LLM to comprehend a database, it has to understand the semantics of the world so that it can better, it can figure out what are the right column descriptions based on the column names and things of that sort. That requires an understanding of the world.

So for us, delivering a, we are, we are making a strong statement saying, look, law install Line Grant and your LLM will comprehend the database enough to deliver answers for you from it. The LL M'S gotta be able to understand like the language and the concepts and the semantics of the world at large. So I don't know if you saw it, it was, um, I think it was like one of the Chief AI researchers at Meta recently left, and he said something, he said, we, we've kind of hit the end of the road with LLMs, because ultimately what we need is a world model Yeah.

That understands more than the LLM does. It sounds similar to what you are saying, what I'm saying. Exactly Right.

Yeah, I I would actually agree with that a hundred percent because, uh, the LLMs, they've been, you know, the, the issue is they've been designed for chat, right? So everything is organized around what's the right response to the, you know, what's the right vector that's coming back in response to a chat like interface. The world model actually is going to make products, could e even make products like us redundant too, but not for a long time with a world model there, you could construe a scenario where the LLM simply understands the database just by looking at it.

Right now, there is still friction because ll databases are, haven't been designed for LLMs. So because, you know, you can take a database with the column names and table names and things like that. Land Grants technology translates all of that into a form and delivers it at the right time to the LLM so that the LLM can comprehend the entire database and answer questions.

Now, who knows, right? The, the geniuses at Meta at GPT, they may very well design something in the next 10 years where we don't need a Lang grant. You know, that's possible too.

But then again, right, our goal is to sort of keep current with what's going on so that we become the instruments or the orchestrators of the LLMs or the wlms or the world language models that might come out. I love it. What a great idea, Ramesh.

It's remarkable. Now, what is this, 30 years you are, you're staying cut. I, I thought I was the only one, but, but I guess not.

This is a great story. Now, you guys recently launched an MCP server, you call Ledge, right? Tell us a little about that.

Right, so LEDGE stands for LLM, enterprise Database Orchestration and Governance. And you know, like every person who thinks they're smart, you know, we also said, Hey, let's come up with a nice acronym. Well, ledge is our attempt at that.

Um, now Ledge is essentially sort of the, the technology behind the Ledge MCP series, the technology behind what I've been describing, which is alleged. It, it does two things, right? It orchestrates the LLMs and it governs them so that your enterprise data security policies and token policies are observed as we allow the LLM to understand the database and answer questions that, I'll give you a con concrete example.

Let's take the sensitive data, right? Like the social security numbers lying around in the database. Let's make sure that, oh, when a user asks a query to get an answer from an enterprise database, the user only gets the information on the sensitive information if they're preauthorized for it.

If they're not, ledge makes sure that the LLM does not generate answers that are given to the user that violate that policy. That's a simple example of that. Another great thing about Ledge is like today you're a business person.

You want an answer to something, you want some data, you want, you wanna say, Hey, for example, hey, why did Q4 sales drop and now today, what do you do? You go to your dashboard and you find, oh, I can't get the data there. I go to an analyst, the analyst goes to a data engineer.

Data engineer says, Hey, you know, hang on, I gotta tweak my query. Takes a few days, couple of weeks, gets the query, runs it, and then itrate and takes a few weeks for, for the answer Here it's all very instantaneous because the LLM can now do the job of data engineering for you. And with Lang Grant instrumenting, the LLM, getting it to understand it with the complying with the policies, the business person gets their answer almost immediately.

It. So that's kinda what I've given you two scenarios of what Lech can do. Now Lech can be even more powerful beyond just analytics as well.

You know, you could think of any application you wanna think of, any sort of scenario that you want answers or some solution to. If that requires understanding of the database, Lech can provide that for you. Love it.

Ramesh, people who want to get more information and dig in here, what's the website? It's Lang Grant, so it's Lang as in Lang chain, LANG. And then Grant, we grant you access to databases.

Clever. Very nice. And of course it's in your name.

That'll be, uh, underneath here. Awesome. Well look, this was, I gotta tell you the truth.

I had fun doing this interview. It was great. Kind of reminiscing.

I think this is a great opportunity though too. It's gonna be interesting how this plays out in, in the market here as people understand that they now can not make their LLM smarter, but use the power, you know, they estimate that in fact, we're not anywhere near done exhausting all of the digital information we have. Right.

A lot of it is in databases behind firewalls and so forth. That's right, right. You know, private data.

And this sounds like a great way to, to start integrating that data into, you know, bring some better value to how we're using ai. Absolutely. I think it fits in where it's very timely.

It's an inflection point. It's only the recent models that enable an application like ours a year ago, the models were not good enough for a product like ours. Now.

And that's the other thing, these things, you know, week to week, even day to day, sometimes even day to better. Yeah, it's crazy. A pleasure having you on.

Come back on and keep us posted on what's going on with Land grant. Awesome. Thank you so much.

It was so much fun reminiscing. And I'll take you up on offer and I'll invite you anytime in Seattle or I'll contact you when I'm in Florida. Absolutely.

Next time I'm there. Awesome. Alright, go check out Land Grant.

L-A-N-G-G-R-A-N-T. Interesting stuff going on there. Bye-bye.

Bye. Alright, we're gonna take a break here on Text Drunk tv. We'll be back.