Articul8 Launches Bivia to Help AI Agents Decide Faster

Hey everyone, welcome back here to Techstrong TV. Let me introduce you to Arun Subramaniyan. Arun is the founder and CEO of Articulate.

And it's kind of a funny spelling. You'll see it on the lower third under Arun there, under his name, Articulate AI. Arun, welcome to Techstrong TV.

Alan, thank you so much for having me. My pleasure. It's our pleasure to have you here.

So Arun, look, I mentioned you're the founder and CEO. Give people a little bit of your journey, your background, what led you to this? Absolutely.

So Alan, I'm an aerospace engineer by training. Spent about half my career doing largely things that fly. So things like jet engines and rocket motors and things of that nature.

I've always been fascinated by things that flew. But also being very closely associated with running very large scale computing. So we were part of the first team that ran large scale GPUs before anybody knew what GPUs were, at least in the general population.

And that journey got us to building larger and larger data centers with a lot more GPUs, which got me eventually to Amazon, where we were trying to get people to use the GPUs. And that was the time when GPUs were not that popular. And then a lot of the gen AI companies started getting funding, which ended up coming to Amazon because that's the only place where there were GPUs.

And we helped a lot of those companies actually build their first models, including the likes of Anthropic. And that journey led me to Intel, which asked me to build a similar business for them at scale, where I built the AI business for Intel across many different industries. Within the process, we actually built a software platform to be able to sell hardware.

" And about a month before ChatGPT was launched, we actually went to production with one of our customers for a search application for a model we had built with their dataset. And when ChatGPT became popular, a lot more customers wanted to have a similar solution inside their enterprises, which was secure, safe, scalable, but at the same time doing things with their own datasets. And it was a software solution and a software product, which really was independent of any hardware provider.

So that was really the genesis of Articulate. But really, if you think about what is it that we do at Articulate, it's really the mission of mixing, combining mission critical workloads with very large scale AI working together. Excellent.

Yes. If someone would've asked you when you were back in school for aerospace if you're going to be the founder and CEO of an AI company working with AI agents, you probably wouldn't have had that on your bingo card, but- Not at all. I would've laughed at that, definitely.

Well, look, if they would've asked me in law school if I was going to wind up running a tech media thing, I would've also laughed. So there you go. I didn't know what...

A, I didn't really know anything much about computers back then. And B, there wasn't a commercial internet, Web, all of those things, so crazy thing. But that's life, right?

That's the beauty of it. Absolutely. I would say to your other question of what is it that we do in Articulate and why did Articulate come to be?

Think about it as a confluence of all of our years of training. So the founding team actually was about 70% PhDs, but only one of them was a computer science PhD. Everybody else was from a different field who had used computing.

And all practitioners who have used it for years, who now have come to a point where the technology's there, our knowhow is there, and most importantly, the industry wanted a solution. So it was really that combination that got Articulate. And we've been hyper-focused on working with industries that require a deep understanding of that particular domain before you actually go solve a problem like that.

And so think of manufacturing, automotive, aerospace, large scale semiconductors. And we work on the, what I would call the core of that particular industry's problem. So whether it is manufacturing or design or operations, how do you make sure that you combine super complex datasets together to get you to your outcomes?

Excellent. That is a great story. And good on you and the whole team there, Arun.

We're going to jump into what Articulate does. Yeah. But before we do, let me just do a little housekeeping.

Website if someone wants to go get more info? ai. And as you said, we are the company with the funny name.

ai. And the name is, the H is not arbitrary, so if you spell out Articulate completely, there are eight characters between A and E. And so that's really where the H came from.

And if you spell it out- Little Easter egg there for you folks at home. Yeah. Nice.

Exactly. Nice. All right.

Let's jump into it. So how does it work? What's it doing?

Okay, so think about any of the industries I named, right? So whether it is manufacturing or automotive, aerospace. You have a lot of structured information, so think information that the company collects and structures it into databases or whatever stores they have.

But in addition to that, they have a lot of unstructured information, things like reports and say even information that is coming at them from measurements, or they have to mix images and tables and text together. They have to do simulation data that is coming in in a very different format. Today, there has to be an engineer, there has to be a technician, there has to be somebody who's a technically knowledgeable person who actually has to assimilate all this before getting it to an outcome, right?

And there's a lot of friction between data hitting them to them making decisions, right? And the computing systems of today require maybe anywhere between a few months to sometimes a few years to build an application that goes from this complex dataset to an outcome. I'll give you an example.

Think of, say, a semiconductor manufacturing plant. And you have these pieces of equipment. Each equipment is a few hundred million dollars.

And each of those are generating information at massive rates, sometimes multiple times a second. The technician has to decide based on information in front of them, should they stop the line? Is there a fault coming?

Is it going to actually affect the particular machine, and is it going to affect the quality, right? And a lot of these kinds of decisions that they're making, the really absolutely bad things the machine will take care of, meaning it'll actually go past a red line and it stops the line. It is really the decisions that require judgment, where the information has to be assimilated by the technician, and the judgment is super high value, meaning if a technician fails to, or an engineer fails to stop the line, that could be a several hundred million dollar impact if something goes wrong, right?

Now, if they stop the line incorrectly, that also is tens of millions of dollars of impact because they've now affected production. So that's what we mean by very high value decisions, right? And they do this day in and day out, so that means it becomes a very stressful job.

There are very few people who are trained to do this. And they end up with easily 70, 80, 90% of false positive hits because these systems are designed not to fail, right? How do you then go in and have a system that comes in, works with the knowledge of the particular industry?

So when you're looking at a dataset that is coming in, it's not just trying to understand the dataset. It knows, as an engineer would know, what they're looking at. And then assisting the engineer or the technician or the domain expert, what are they actually looking at and what decisions they potentially can do with a probability associated with it, right?

And the probability is important because, like pretty much all of us use some form of AI or other today, and the funny thing with these things is they're almost always confidently answering every question you ask them. Whether they know it or not doesn't matter, whether the answer is right or not doesn't matter, the confidence level doesn't waver as much. That's really what we have to work on, where you need to be able to not only give an answer, but be able to tell where did the answer come from.

That's very important. What is the provenance and what is the confidence level of the answer? So that's, I would say, in a nutshell, what we do across many industries.

I gave you a semiconductor example, but the same thing works in an automotive manufacturing line. One of our customers, one of the largest auto manufacturing customers, in one of their lines, they have about 6,000 faults every hour that gets triggered. So now the engineers and the technicians have to decide which of them, how to triage them.

And after triaging, they need to figure out, okay, how are they going to fix it? A lot of these kinds of exactly use cases is what we've been doing now for over two years in production. So the semiconductor example I gave you, we've been deployed in a semiconductor manufacturing plant in Phoenix, which does not even have internet connection.

So their data sets are so secure that you have to run entirely in an isolated air-gapped environment and have the same level of accuracy and the same speed that we can do with models that are deployed outside the box. So that's an added layer of security that we provide where we operate everything from on-prem air-gapped to a SaaS environment. Crazy.

You know what, Arun, I speak to a lot of people working with agents, obviously. Everybody's making agents, using agents. The thing that I keep hearing from people is trust.

Yes. Until I could trust that the agent's going to do what I want it to do, that it's not going to do what I don't want it to do, probably more importantly. That the guardrails are built in, the governance is there, right?

That it's not going to do more harm than good. Yes. If you will.

Yes. And it seems like that's where the whole AI industry is right now, right? That's the, not the bottleneck, but it's the path that we have to get through.

That's right. And that brings me to Vivia. That's the decision model that we've launched.

And the reason why we've decided to launch a model, by the way, if you notice in my company introduction, we are not a model company. Even though we build a lot of models, we don't sell models until now. The reason for that is we believe that models are a means to an end.

They can't get you an outcome by themselves. So that's really why we've shied away from selling models until now. However, decision models are the first class of models, where just using a decision model alone will get you to the micro outcome that you're intending to get to.

Which is exactly what you said about trust, right? Where you're trying to make a decision based on the context you have, and you do not want the model or the system to give you an answer out of context. Right?

And you want it to tell you that, hey, I can answer this, or more importantly, you want it to tell you, I can't answer that based on the information that you just provided me. Which by the way, is structurally impossible or difficult for existing class of models to do, because they've been trained with everything on the internet. They're trained to actually give you an answer no matter what, right?

That's what they're trained to do. The decision models, the class of models, really what they do is they can only give you a yes or no answer, or they can choose based on a set of choices you give them. And they're designed to only do that.

Now, they also suffer from the same problem that they'll always give you an answer. So they'll give you yes or no, or they'll always choose from the choices you give. What we've been able to do, and the reason why we decided to launch Vivia, is the model is designed to also know when to abstain.

Meaning, I would say the most critical part of Vivia is the fact that it knows when not to answer. So you give it a context, and you ask a question. Say, for example, you give it a context about how beautiful Paris is, how great the streets are and all that.

And then you say, now tell me how good is New York? It will tell you that I have to abstain because I don't have any information about New York. And that alone is a massive differentiator because in, let's say, any kind of critical environment, so whether it's a bank or a manufacturing facility or an automotive facility or even a healthcare facility, you need to know when to answer, and you also need to know when to abstain and escalate to a human.

That's really what Vivia does quite well. Of course, the other thing we also looked at was this model was born out of more than three years of decisions that we've been making inside our platform. Real-world decisions that actually went into training this model, right?

Which means most of the other models have only text, and some of them have images as input modality. We realized that if you're making real-world decisions, you need pretty much all the modalities. You need to be able to see, you need to be able to hear, and you need to be able to think and reason.

So that's why our model out of the box is multimodal. It can take in texts of any kind, images, tables, audio, and video. And you may ask, what's so different about having, once you have images and audio, what's the big deal about video?

Well, in video, you also need a space-time connection. So you need to understand not just that what you're looking at the image, but you also need to know is it changing over time? And understanding time becomes important before you answer a question.

All of those modes are built in out of the box as you start making decisions. Right. The other thing is, so there's two equations to trust, right?

One is, of course, knowing when not to answer. Second is, of course, making sure that you're giving the right answer when you're supposed to give the answer. But also, are you breaking the bank or not?

Is it economical to do it when you're actually putting the system in there? And that's another advantage that the decision models bring to the table, where the cost of making individual decisions have now gone maybe about 1,000 times cheaper than using an LLM to do the same thing. So in addition to having all these advantages, it's fast.

It's usually roughly less than 100 milliseconds to make a decision. So you make thousands of them very quickly. And even if you make millions of these decisions, you're not going to be spending much.

Got it. Yeah. Arun, we're about out of time.

Again, Articulate spelled with an A is the way to go. I want to wish you guys luck with and success with Articulate. It sounds great.

And more importantly, it sounds like the team, you and the team, this is your moment, right? This has come together for you nicely, and the currents of time are here. And then lastly, thank you for carrying the conversation because I'm not talking as much as I usually talk, so I appreciate it.

Absolutely. Thank you so much for taking the time. Thank you for giving me the opportunity, and look forward to talking to you again.

Thank you. Absolutely. Arun Subrah-- I said it in the beginning, and now my lips are like...

Arun Subramanian, founder and CEO of Articulate AI, here on Techstrong TV. We're going to take a break. We'll be back.