AI Engineering Podcast

Share on social media:

Summary
In this episode of the AI Engineering Podcast Dr. Tara Javidi, CTO of KavAI, talks about developing AI systems for proactive monitoring in heavy industry. Dr. Javidi shares her background in mathematics and information theory, influenced by Claude Shannon's work, and discusses her approach to curiosity-driven AI that mimics human curiosity to improve data collection and predictive analytics. She explains how KavAI's platform uses generative AI models to enhance industrial monitoring by addressing informational blind spots and reducing reliance on human oversight. The conversation covers the architecture of KavAI's systems, integrating AI with existing workflows, building trust with operators, and the societal impact of AI in preventing environmental catastrophes, ultimately highlighting the future potential of information-centric AI models.

Announcements

• Hello and welcome to the AI Engineering Podcast, your guide to the fast-moving world of building scalable and maintainable AI systems.
• Your host is Tobias Macey and today I'm interviewing Dr. Tara Javidi about building AI systems for proactive monitoring of physical environments for heavy industry

Interview

• Introduction
• How did you get involved in machine learning?
• Can you describe what KavAI is and the story behind it?
• What are some of the current state-of-the-art applications of AI/ML for monitoring and accident prevention in industrial environments?
  • What are the shortcomings of those approaches?
• What are some examples of the types of harm that you are focused on preventing or mitigating with your platform?
• On your site it mentions that you have created a foundation model for physical awareness. What are some examples of the types of predictive/generative capabilities that your model provides?
• A perennial challenge when building any digital model of a physical system is the lack of absolute fidelity. What are the key sources of information acquisition that you rely on for your platform?
  • In addition to your foundation model, what are the other systems that you incorporate to perform analysis and catalyze action?
• Can you describe the overall system architecture of your platform?
  • What are some of the ways that you are able to integrate learnings across industries and environments to improve the overall capacity of your models?
• What are the most interesting, innovative, or unexpected ways that you have seen KavAI used?
• What are the most interesting, unexpected, or challenging lessons that you have learned while working on KavAI?
• When is KavAI/Physical AI the wrong choice?
• What do you have planned for the future of KavAI?

Contact Info

• LinkedIn

Parting Question

• From your perspective, what are the biggest gaps in tooling, technology, or training for AI systems today?

Links

• KavAI
• Information Theory
• Claude Shannon

The intro and outro music is from Hitman's Lovesong feat. Paola Graziano by The Freak Fandango Orchestra/CC BY-SA 3.0

[00:00:05] Tobias Macey:

Hello, and welcome to the AI Engineering podcast,   your guide to the fast moving world of building scalable and maintainable   AI systems.   Your host is Tobias Macy, and today, I'm interviewing doctor Tara Gevidi about building AI systems for proactive monitoring of physical environments for heavy industry. Industry. So, Tara, can you start by introducing yourself? Thank you so much, Tobias.  

[00:00:33] Tara Javidi:

As you mentioned, my name is Tara Juvedi.   I am a researcher,   working at the intersection   of,   information theory and, generative AI. And as we will get into,   I'm here to talk about my role as the CTO of Kav AI.  

[00:00:52] Tobias Macey:

And do you remember how he first got started working in the ML and AI space?  

[00:00:57] Tara Javidi:

Yeah. So I did my undergrad at Cheif University   and then did my PhD at University of Michigan for my grad school in in the early two thousands. Now going back all the way to my childhood,   influenced by my own mother,   I had always been very interested intellectually   and fascinated by the way mathematics   and its formalism   allows us to do a like, to take a very first principle approach to solving problems.   At the same time, engineering sound seemed very fascinating   because it aligned with my values in terms of having a positive impact,   you know, socially and and physically in the world. But I have to say, unfortunately,   I thought I had to make a choice   between pursuing a degree in math or doing engineering   until quite late in my, you know, and it wasn't until second year or third year of my graduate studies that I got formally introduced,   to the subfield   of electrical engineering,   as part of,   my PhD studies and after I had already done a master's in mathematics,   that also overlapped with applied math called information theory, where I ended up doing my PhD   work in that space. So I don't know.  

I know you,   you know, the audience is technical, and many people might know. But I don't know if there is interest. I can always, talk a little bit about information theory. And,   if if you think that's of interest.  

[00:02:32] Tobias Macey:

I can add links in the show notes for people who wanna dig a little bit deeper on some of the specifics there.  

[00:02:38] Tara Javidi:

Excellent. Yeah.   So yeah. So so,   and I think the to the audience, this would be interesting because   the sort of at least the the one bit of,   historical information that I would share is, the word   information theory was termed by and and was due to a work of,   Claude Shannon, whose name   is nowadays   very familiar to folks who are following generative models. And so this field basically is focuses on turning information,   you know, from conceptual concept, philosophical   question of information,   and turn it into digital data. And this is really a big part of generative AI models and the way they work, these days. So and that really what transformed,   my trajectory   not to think of mathematics and and engineering in a x or kind of a sense,   but more importantly   brought me to this bigger field of information information within machine learning and data science.  

And so that's   the sort of background. And now as a result of this, I do work in, machine learning and   AI. But from this particular lens, I use this, Shannon's way of thinking about data,   from first principle.   And I always ask myself, is there a way of that my training in information theory   can really suggest a different way of thinking about, digital data information to digital data,   and that others might have missed? So in that sense, that's, you know, the interaction   that I bring, the the sort of, the toolset that I bring to the field of AI   using the sort of view of information.  

Philosophically,   also, I'm really influenced   by Shannon's work in terms of its societal impact, and I'm happy that you will share some some of the background. But, really,   I cannot overstate   the the impact that information theory has has had, societally,   giving us phones and, cell phones and now generative AI. So I'm also   very influenced by thinking about societal impact of the projects that I engage in and also having a lot of fun. So Shannon was also very interested in playing games.   As, again, I'm hoping that your audience are familiar with.  

[00:05:02] Tobias Macey:

And so digging now into   what you're building at Kava AI,   I'm curious if you can describe a bit of what it is that you're trying to achieve and some of the story behind how you decided to spend your time and energy in that space.  

[00:05:16] Tara Javidi:

Yeah. So I was a a faculty at UCSD   engaged in the kind of research that I talked about for two decades. But CoveAI is a startup company that I've cofounded with, Sam Bighatelli, my cofounder. And our aim, as you noted at the beginning of the conversation,   is to push the boundary of AI to operate natively within the physical world. So at at CoveAI,   from a first principle approach, we are building a new generation of AI systems, and that mimics really the way curiosity   functions for humans as we explore the physical world. And that's what I mean by natively processing information in the   world. Now I can geek out on this, and I'm sure as we talk deeper, I will talk about this new generation of AI. But here, that we call curiosity driven AI or artificial curiosity.  

But here, I will leave the details for that later part so that I make sure   that I I don't come across that we are only interested in AI for AI's sakes. These models are critical   to addressing most pressing problems at the intersection   of physical intelligence   and industrial monitoring.   And and I can,   depending on what you think is interesting, I could talk about both of these aspects, the whys and and the what,   of of what we're working on.  

[00:06:40] Tobias Macey:

And yeah. So to that context of   industrial monitoring,   some of the types of information generation   and the actions that you'd   want to take based on that information. Obviously, there is a lot of work going into things like monitoring,   preventive maintenance, etcetera. And I'm wondering if you can just start by giving a bit of an overview of the current state of the art for some of that monitoring and alerting and applications   of predictive and prescriptive analytics in that context.  

[00:07:11] Tara Javidi:

Yeah. So,   bigger picture,   we are all very concerned about lack of intelligence when it comes to our industrial operations,   construction sites, shipping lanes, oil rigs. There's just an urgent need to really up our game,   and build physical intelligence that   recognizes   early signs of imminent and often, in fact, very catastrophic   accidents and failures. And and this is the gist of the problem. Most industrial sites today   rely on a mix of human oversight   as well as precision robotics in service of precision sensing, and then some IoT   systems and platforms that collectively   gather information   sort of passively sitting in the background. Right? So this is sort of the data that comes from the site. Now each method of collecting data   has its,   its own strengths.  

Robotics, precision robotics, for example,   we know of robots that can move   sensors very close to the pipes,   let's say, and and sort of crawl the pipes   in search of potential signs of of, you know, cracks and and structural damage, corrosion.   And then IoT sensors,   you know, sort of have this, nice property that they constantly provide you with very green information that that sort of they can stream to you. But the current state of the art is that sort of puts these, data   collection devices in an sort of an arbitrary and rather unintelligent   way. They've been predetermined   that, okay. I'm gonna put some IoT devices here, and they have some arbitrary   schedule of how their robots will scan the site. And as a result,   you get this data that has tons of informational blind spots. It misses important information   when it's needed,   because the scanning schedule was not looking at the right pipe at the right time, or the IoT devices were giving too,   too coarse of information or inaccurate information.  

So what happens is, usually,   the gap is filled at currently   by people, or I should say by a team of very experienced site operators   that   come in   and they recognize what information is missing, and then they deploy further inspection by the right tool, you know, to the right source of data for testing for the right or wrong form of hypothesis. This is the current state of the,   of the monitoring for these sites.  

[00:09:52] Tobias Macey:

And given the   level of technology and investment that is being made, what are some of the shortcomings   in that approach and some of the ways that the   lack of complete awareness or   gaps in information   can lead to things like environmental or economic impact?  

[00:10:15] Tara Javidi:

Yes.   And so as I kind of give you the sort of the the shortcoming already in my, in my framing of the current answers. So, the data when it's collected in this sort of passive way, it's just you or predetermined   schedules has   gaps. And humans,   even though teams of humans, at least, are good at hypothesizing   about missing data or sources of information or sources of informative   sensors, our ability,   is is still rather limited,   especially at larger sites with massive scales. So   no one operator can walk a 3,000 acre facility every day,   especially,   considering the intellectual   and functional costs associated, you know, sorry,   financial costs associated with walking these sites.  

Basically, there are just massive fishing expeditions   that you take a very experienced   intelligent   person   to go after hopefully nothing. Right? So day after day going there and testing and seeing, oh, yeah. Thankfully, I came back empty handed from this expedition   is, as I said, both financially not so, promising,   not a scalable process, but also,   it creates these jobs that are uninteresting and and humans are not very interested in in, you know, performing such jobs. So those are, I think, the the the big challenges of why we are not doing a good job   monitoring these sites. And and and this I can give you further and further details about how this data is, you know, is not sufficient for us and and the cost associated with with not doing a better job at at different sites.  



[00:12:04] Tobias Macey:

And to the point of a lot of that information collection being very   burdensome and a lot of toil requiring   a lot of human capital to be able to actually   complete that. I'm curious how you're thinking about the potential for   reimagining   or   retrofitting   the actual   collection   piece of the problem and some of the ways that the design of the solution informs the ways that you're collecting that information. And maybe there is either additional information or maybe there's information that's currently collected that is not actually useful given a more digital approach to it and just how you're thinking about that overall   holistic approach to being able to improve the effectiveness and reduce the level of toil involved?  



[00:12:54] Tara Javidi:

Yeah. Absolutely.   So Cove AI fills the, this gap in information in general. And and this is sort of   across all physical spaces and all physical sites. As I said, we are building these, new generation of AI models we call curiosity driven AI. And the the platform really provides an evergreen   bird's eye view   of the entire site   and and gives the you know, this bird's eye view gives them subsequently   the ability to the platform   to hypothesize   physically where and at what granularity,   you know, the platform should look for earliest for the earliest,   signs of, you know, problems and and issues. And so, basically, the platform is in is enabled in in a way   to zoom in on more pertinent,   sources of data   and predict problems before they kind of escalate to be   catastrophic. Now at the beginning,   we wanted to be very you know, in in a way, philosophically,   we have been taking this approach that we don't wanna do AI for AI six. We wanna be very mindful and focused   on on the real world impact of what we are doing. So in terms of products and services,   we have started   and, so far have been laser sharp focused on the energy production and processing site where I can give you a bit more sense of of how this this is done. So our artificial   curiosity   platform,   really, what it does, it just wants to   incorporate the data that you're collecting at these sites, and this is where I can be very specific. For example, you might have humidity sensors.  

Right now, they're just sitting somewhere. They reported it. It's and if if you're in Tennessee, none of it is is all particularly   alarming. Temperature,   you know, thermal cameras that the the inspectors bring to a site and do every quarter or so inspection   are gonna show a high temperature again if you're in a place, that is hot. But if you combine these two with the fluid content of a pipe and put all of that information together,   now you can actually build models   that are far more intelligent about   detection of these early signs. Now if you put all of that data   blindly, then you're, like, gonna be completely   overwhelmed in tsunami of, like, redundant useless data.  

So, really, this platform we call artificial curiosity   works the same way curiosity helps human beings   to operate in the physical world. How do we do that when we are going to an airport and it's very complex? What do we do? We zoom out things that we believe are not important   and focus on things that matter. If we are wrong, I arrive at the gate that happens to me a lot, and when I see, oh, this gate is too empty or just something doesn't add up, I sort of hypothesize as what sources of information   is most,   faulty in getting me here. This baby was crying. Maybe I didn't hear a gate change announcement. Right? So that, that would be called a functional curiosity, the way we actually move our bodies. We go to the, you know,   an agent that has that source of information. That is what the platform is really allowing us to do, is to turn on and off the sensors and the noise that they bring.  



[00:16:25] Tobias Macey:

Now in terms of the overall   ecosystem of AI,   large language models have stolen a lot of the attention, but there's still a lot of other work still happening in the context of machine learning, even generative intelligence.   And on your website, it mentions that you have developed a foundation model for physical awareness to help power some of that active curiosity. And I'm wondering if you could talk to some of the ways that   a foundation model for physical awareness   is maybe architecturally similar and maybe more interestingly divergent from a language focused foundation model.  

[00:17:04] Tara Javidi:

Yeah. So I sort of gave, the the the core idea here. But, really, the the concept here is   current models   ingest data that is spoon fed to them. Right? The data is already prepared, and they sort of take it and turn it into intelligent. When we think about an intelligent human being, we don't only think of them as how they infer   or how wise they are when the information is presented to them. We think of people   as intelligent when they go out and seek the right information   to do something. Right? So and this is this active,   this closed loop operation   that we build into   how the information sources are generated   or collected in order to generate data that then will be ingested by the model is the big differentiator   between what we are doing and what the current,   or or maybe the most popular,   large models being language or or vision models work, if that makes sense. But I can I can go in deeper,   deeper details if if you think it's of interest?  



[00:18:15] Tobias Macey:

Yeah. I think that it's maybe also an interesting   corollary   is so in a lot of the industrial monitoring use cases, there is a set of machine learning that has become   fairly well adopted in terms   of supervised learning models where you know, okay. If I have this set of inputs, then this is the type of response. And I know that I need to be able to do some maintenance on this machine because it has   hit either this number of hours of operation or it's starting to exhibit some of these symptoms, whether it's a particular   noise or a change in pitch, etcetera.  

But in terms of   the   generative model   context, a lot of that is self supervised or unsupervised.   And I'm curious how you're thinking about the application of some of that more attention oriented   machine learning architecture   is applicable to this physical   observation   style of interaction versus just being a very digital native,   approach that we're doing with language models and some of the ways that you're able to do some of that translation of analog signals into discrete digital representations to be able to do that more attention focused learning on that data.  

[00:19:29] Tara Javidi:

Yeah. Absolutely.   So so you're you're raising a really important point. So, if you think about the way we think about the attention as,   the architecture   in, you know, behind much of the success of the Gen AI models, the attention is to the particular tokens that are already being ingested. Now these models have a bottleneck that they cannot handle   very large number of tokens. In fact, long context is sort of the frontier of how good a model is is how much token it can actually consume and still produce,   sensible,   you know, outputs.  

So now think about building these systems with data that you're feeding from a completely useless   store you know, not I shouldn't say useless, but redundant source that is repeating that same temperature reading that it has been repeating for the past,   you know, month. So in this setting,   it's very clear that you don't wanna create these what I say instead of long, it's a volumetric   context.   Right? If you just tokenize it in in a brainless   manner. So that's where this closed loop operation comes in as how would you actually turn and say this sensor right now, the product   the the the readings,   I am,   confident I'm giving an output that is consistent with the reading. So in that sense, it's self supervised so I can start   reading at less. Right? So that closed loop feedback is is exactly the differentiator   in in what we are doing. But the architecture,   you know, the architecture components are similar. So you can sort of think about and now this is I'm going into way too much, details, so hold me. But, really, you can think of this as a   physical attention   that goes,   you know, wider than your, your narrow architecture   by kinda going in the world. And we think of this as the spine as well as the sort of the our limbs and arms and hands going and grabbing the data that you want to bring it into your model.  



[00:21:35] Tobias Macey:

And the other interesting aspect of that problem is, as you mentioned, you are developing these additional means of data acquisition to remove the requirement   of toilsome human activity in that process.   And I'm wondering how you're thinking about when you're removing a human from the loop in some capacity,   there's still a need to make sure that the digital system and the other information gathering systems are themselves monitored in some capacity. And I'm curious how you're thinking about that aspect of how to ensure that that digital collection has high reliability and resiliency   and is able to be adaptable, particularly   in the context of a constantly changing physical footprint where maybe you have a factory that is expanding or going through renovations or there are upgrades of equipment and so some of the ways to be able to adapt to that constant change.  



[00:22:30] Tara Javidi:

Yeah. Yeah. So so you're absolutely correct. But we are a little bit lucky in this luckier than than having to solve the problem from the scratch. When you go to these sites, they already have   some sophisticated   autonomous data collection process. Many of them have ground robots.   Many sites have looked into   operating   drones and so on. So we don't get in the weeds of, like, building a full, you know, hardware   platform for them that is doing all of this. They already have this, and their number one problem is they don't know which one of these devices are useful at any given time. And that is this the the the problem that our foundation   model with its spine, which is an OS, really solving. It's really giving them the wisdom   to operate these, these sites. Now the great news is that these components are becoming more and more advanced. We are getting better and better precision robots. We're getting better and better drones that are very stable. They have their own safety constraints. So that's where we are not getting into the competition of building a full autonomous system with all of these devices. We are building on the power of many of these existing   technologies. We're just really solving the main problem, which is now that we have the ability to collect such sophisticated   data, which one of these data sources   are holding the gold for me as predicting my next failure, and and then we we really exploit what is there in these sites. And they have done a good amount of investment in technologies   along the lines that you have in mind. So we often work with their technologies.  



[00:24:15] Tobias Macey:

As far as the system that you're building beyond just the data collection and some of the model specifics, I'm curious if you can talk to some the overall system architecture   of the platform   and how it integrates into the workflows   and existing systems of the companies that you're looking to support?  

[00:24:33] Tara Javidi:

Yeah. So so the architecture is, again, very much inspired by the way   we call it a human intelligent.   So the, basically, the architecture   is aware of the devices and their physical,   sensing modalities.   And we in our models, we basically the trained   the trained model is taking advantage of certain physical laws of sensing. So for example, when you go away almost,   you you know,   uniformly, does not matter what sensor you're you're talking about. When you go away,   physically, you distance yourself from the sensor, the information gets, sort of,   less accurate. So, you know, this is of the the component that the physical intelligence on the sensing side that we are taking advantage. And then there is an operating system, as I was telling you, that acts like a spine that knows where the information is, but decides what to filter   upward towards the model that then, you know, uses the tokens to,   to recognize something has changed. There is a major   reordering of the factory, and, hence,   the model needs to be reupdated   in the spatial   configuration,   sense and so on and so forth. So this is the architecture. The architecture is really profiling the sensors,   creating a, sort of a ecosystem of data, and sort of having an integration OS that that can that time stamps and geostamps these sensitive information   such that if they become useful, we have we have these physical laws of where it was collected. Hence, it's correlated to the information that   and geo and geographical   location.  

And then the the, you know, foundation model that is uses this to build, hypotheses as where to allocate most of its,   attention. Now it becomes a physical attention. This is the the basic architecture.  

[00:26:33] Tobias Macey:

And one of the interesting challenges that always comes up when you're trying to introduce   either   new or different machine learning and AI capabilities   into a particular operating context is the user experience around that as well as the potential for the human operator to have some sort of either reticence   or mistrust of the system. And I'm wondering how you think about the ways that you introduce   the cav.ai   platform into these organizations   and integrate it into the responsibilities of the person who has to either support or take action based on the outputs of your system and some of the trust building exercise that goes into   maybe giving some insight into why a particular   response is created or why a particular recommendation is made.  



[00:27:24] Tara Javidi:

Yeah. So this touches upon so many different,   aspects of both technical, but also,   you know, one of the things is at the beginning, I said taking this view of, like, engineering and sort of impact on the world.   From the get go,   we were very committed to work with,   real   people and solving real problems. I felt like if I wanted to do AI for AI's sake,   Ivory Tower is not such a bad place. Right? So so we were very lucky that that decision, which was more like   our own value system to work with operators   in the energy sector,   gave us this possibility to think about what is their problem and what kind of interfaces   they use   as people who do this job day in and out and and what kind of help we can give them. And so that's why I completely agree with you. The way this data is presented   is by far the biggest hurdle for the operators to make sense of it. It's siloed. You need five different interfaces   to be building. Like, look at your IoT digital device, and this is handled by your IT team. And then there's an inspector who comes in and gives the data on on a piece of paper. So so that's so you nailed it. And without going, you know, or giving away too much of our secret sauce, that has been   very much in in our DNA, working with operators   and really,   making sure that our,   OS provides,   the data in a way that is well understood   by the operators.  

Again and also because we are closing the feedback loop and need to hypothesize about the sources of pertinent information,   that gives us a bit of an advantage in the sense that now that has to be something that the operator is comfortable   with, and that needs   to be, in a way, trustworthy for them to really follow our our suggestion that, hey. This is what you need to send the the next inspection   crew.  

[00:29:32] Tobias Macey:

And then another interesting   aspect of building any type of   platform or product like cav.ai,   where we're introducing machine learning across a   suite of different contexts   and organizations   is that there's the potential for being able to   do some level of transfer of information gathering and insight creation   from one operating context to another one. Obviously, there are issues around regulation, privacy, etcetera, where you can't directly use data from one organization   in another. But I'm curious how you're thinking about the ability for the overall platform to be able to improve and   learn from collected information and collected experience to then be able to improve the functionality   for all of the consumers?  



[00:30:22] Tara Javidi:

Yeah. No. You're you're absolutely right. And this is something that that becomes more of an art than,   science at the level that you are envisioning it. For us so far,   really the name of the game has been understanding   the physical world and under when I say understanding, it's from an from this lens of information   so that we can really ask ourself,   where should I you know, the eyes of our cameras,   literally,   and our sensors   figure figuratively are looking at the world. And and we would like to understand   how to arrange these eyes and,   and sensors, these cameras and sensors   across the space. So, really, the the knowledge that we are learning and our model is being trained on is this knowledge of how the information   is filling   a physical space and what is the physical loss of sensing that that is gonna be embedded in in that process. In that sense, we have not gotten   too deep into   the, you know, site specific sensitive data, then that like, these   older models would train on your data and your pipes and the arrangement of your pipes to then learn secondary and, ternary effects.  

And that's really interesting,   as how to cut that, that pie. But for us,   right now,   the real knowledge that is transferred across sites is this knowledge of how sensing   gets affected by distance, how sensing gets affected by temporal,   distance, and so on.  

[00:32:03] Tobias Macey:

That's another interesting aspect is the question of time in the context of physical systems where it can have different levels of significance depending on the pace at which the different operations   are happening, where if you're worrying about preventive maintenance for a machine to make sure that you do some repairs on it before it breaks down, you're going to have a much longer time horizon than if you're doing some sort of sensing for risk of explosion,   for instance. And I'm wondering how you think about the overall   decay of information utility   across those different temporal boundaries of the data that you're collecting and how you think about how to apply that rate temporality   and the, maybe, dampening of the signal across those different use cases?  



[00:32:50] Tara Javidi:

Yeah. Yeah. So you so you you are raising an issue that often   it's true for space as well, but somehow it's harder to convey it. But in time, we all have this really good understanding   of a lack of an absolute fidelity.   Right? Time can tick at the microsecond level, at millisecond level, at day level, year level. And when you tick,   take physical worlds, there is no absolute   temporal clock that you would be wanting to,   you know, collect the data. If you do it too fast, you get such correlated data that then it will overwhelm you in in form of redundancy,   if you do it so slowly. So that's sort of the digest of,   what we are doing. And then these hypotheses   are sort of, the   larger memories in some sense that that are built into the system to to keep track of these longer context.  



[00:33:48] Tobias Macey:

And as you have been   developing this platform,   working with some of your earlier customers and helping them understand the capabilities   and,   applications of it, what are some of the most interesting or innovative or unexpected ways that you've seen the platform applied?  

[00:34:04] Tara Javidi:

Yeah. So   no. You know, we are we are still very,   a bit too early, so I can't tell you,   reveal too much. But but I will say   that this idea of hypothesis driven,   you know, prediction,   this has been,   you know, this has been really fun. Like, we we,   talked about it. We we were looking at it from our end. But the first time that we did, one of the paid pilot projects, and and I could see, you know, that that it works   and how well it works and how useful the customer I was saying, this will be very useful for me to, think about scheduling my next inspection. That was, like, one of the proudest moment of my career, in, general.   Now we will be making more   more announcements   about the specific   projects and so on on LinkedIn that I would very much think, if your audience want to follow more specifically, they can follow.  



[00:35:04] Tobias Macey:

And as you have been working in the space, developing the technology,   evaluating its effectiveness, what are some of the most interesting or unexpected or challenging lessons that you've learned personally in the process?  

[00:35:16] Tara Javidi:

Yeah. So so the most interesting is something that we haven't talked. I mean, this as you said, this is a technical,   you know, tech technical audience. But I think it I have to say, in general,   the most interesting part has been around growing a strong and exceptional team. You know, this is these are, like, really the most interesting lessons that I when I look back at the at the experience Kaveh either have been fascinated by is, like, how to recruit the best people,   how to retain them, how to grow them both individually, but at the same time, help them build a collective   together. So this has been really the most interesting how how we have been able to cultivate a culture where all of my team members, I can think of them as important intel individual members with their own needs and their own constraints,   but then the team as a whole, you know, vouching for each other and building a collective sense. So that's been the most interesting.  

Now you said also unexpected.   Right? So if I have to talk about unexpected, I would say   the most unexpected   fact is how much of my time I am on Zoom.   And or I should maybe say it, the remote work culture that was created during the pandemic   has helped us in ways that I really didn't   think about. It would be possible   pre pandemic or pre this, this, remote work culture. So we have been able to recruit the best talents from all around the world, and that's really been the most unexpected if I have to to pick. And and as I said, personally, how much unexpected if I have to to pick. And and as I said, personally, how much time I am still   spending on Zoom.  



[00:36:58] Tobias Macey:

And as you continue to   build and evolve and iterate on the overall problem space of AI for physical environments with an industrial focus, what are some of the particular projects or problem areas or new capabilities that you're excited to explore?  

[00:37:15] Tara Javidi:

Yeah. So   from a societal   perspective,   I'm very concerned that we have grown our footprint technology footprints in ways that goes way beyond our ability to monitor it. And in one particular sector, and that's energy sector, we have seen the catastrophic   cost of doing this from an environmental   perspective. And that's where   I am really laser focused on, and I really would like to see that these   environmentally   destructive   incidents,   to become a thing of the past. You know, there are certain ways humans accept that the world, you know, earlier on, like, hundreds of years ago that we think of them as things of the past. And and I would very much like this to become a thing of the past that our children, grandchildren   get used to saying, oh, this happened in 2018,   but this is ridiculous. How could these people let that happen? That's what I, I'm really, really excited,   but also very committed to achieve.  



[00:38:16] Tobias Macey:

Are there any other aspects of the overall application   of AI to physical environments, the work that you're doing on Kav AI specifically,   or the potential for this technology driven improvement to industrial processes that we didn't discuss yet that you'd like to cover before we close out the show? I think this curiosity  

[00:38:36] Tara Javidi:

driven, this thinking of data generation   in an integrated   way   with generative   AI, or I would call it information centric   AI, is really you know, excites me, but it's also a really important   topic that I think more and more will will create more opportunities   beyond what, we are working on immediately now, and that's I believe that that's that.  

[00:39:03] Tobias Macey:

Alright. Well, for anybody who does want to get in touch with you and follow along with the work that you're doing, I'll have you add your preferred contact information to the show notes. And as the final question, I'd like to get your perspective on what you see as being the biggest gaps in the tooling technology or human training that's available for AI systems today.  

[00:39:20] Tara Javidi:

I think the the existing,   you know, AI systems   are,   you know,   are very good at learning and ingesting   digital data,   larger large amounts of it whenever possible.   The race for the next generation of AI, however, especially in the physical,   AI domain, will   be really the the forefront   or the winner of that race   will be architectures   that really build on this information centric AI. This curiosity driven AI will be the first generation of.  

[00:39:56] Tobias Macey:

Alright. Well, thank you very much for taking the time today to join me and share the work that you're doing at Kav AI and some of the interesting challenges   of being able to move AI beyond just the digital realm. So I appreciate all of the time and energy that you and your team are putting into that, and I hope you enjoy the rest of your day.  

[00:40:14] Tara Javidi:

Thank you so much, Tobias. I really enjoyed our conversations.  

[00:40:22] Tobias Macey:

Thank you for listening, and don't forget to check out our other shows, the Data Engineering Podcast, which covers the latest in modern data management,   and podcast.in   it, which covers the Python language, its community, and the innovative ways it is being used. You can visit the site at themachinelearningpodcast.com   to subscribe to the show, sign up for the mailing list, and read the show notes. And if you've learned something or tried out a project from the show, then tell us about it. Email hosts@themachinelearningpodcast.com   with your story. To help other people find the show, please leave a review on Apple Podcasts and tell your friends and coworkers.