So, to have Naveen, I'm sorry, forgive me. I Krishnaraj? Krishnaraj, yeah. Awesome. Senior data scientist, born out of the oil and gas industry. So, you know, excited to have have him on. You know, I guess usually we we get started uh You know, let's just start You know, I think we've had a lot better luck lately. Let's just jump into what What are you doing now? And we'll kind of maybe, you know, work backwards or kind of asynchronously, you know, through your career. But, you know, you know, senior data scientist, you know, what does that mean? What are you working on? What even, you know, what industry you're in? And let's start there. Yeah, I'm I'm currently in AgTech, actually. So, I'm working as a data scientist at Sabanto. We are a tractor automation company. Okay. What we do here is we provide autonomy kit to people who already own a tractor. We provide them the kit so that they can run their tractor autonomously. So, that's that's the company business model. And I've been with the firm for a year now. Actually, today is my first year anniversary. So Oh, nice. Congratulations. I think it's to draw parallels between oil and gas and then agriculture is like these are necessary industries that, you know, again, like what the product that each one produces are critical to the livelihoods of, you know, pretty much everyone on Earth. You know, where And then again, like you need energy to generate the um [clears throat] the agriculture and not even some of the fertilizer and stuff, but at the same time, again, just we need a strong, you know, global food supply. So, it's a definitely definitely interesting to see the convergence of the two of those in your career. Yeah, like you said, like both are fundamental to human needs, actually. Without energy energy is actually a measure of, like, you know, how wealthy a society is, right? So, I moved from energy industry and now Uh, agriculture, like, everybody needs to put food on the table, so it's it's also fundamental to human needs. And, in terms of parallels, the technology a lot of there is a lot of overlap in. So, we use a lot of sensors in oil and gas, and there is a lot of sensors in in autonomy where I work on. So, there are various different sensors, uh, people bring in to this autonomy industry. Like, for example, there are lidar sensor, radar sensors, and uh not necessarily we use all of those sensors, but, you know, these are like oil and gas we have I used to do petrophysics, right? Uh, so there are logging logging tools have all sensors, right? So, this is the same parallel. Uh, and uh uh I I can see a lot of different comparison between the those two industries. Um, like, mechanical aspects of it, and also like um for me from the data aspects of it uh I can see a a lot of different cross cross-disciplinary applications here. And uh I was able to transition into this uh quite quite easily, you [snorts] And And so, was was this your first um role outside of the industry, or had you worked outside the industry before? Uh, so once I graduated with my PhD in petroleum engineering, uh I I I did an internship while I was doing my PhD, uh but then never really worked in the industry. After that, I went to do consulting. Uh, so more of a client-facing job. There I worked on multiple different problems for 1 and 1/2 years, close to 2 years. Uh, so there I did like social media analytics. Uh, had worked on um few other, I had worked on like um some, some kind of like, let's say, like um, uh financial kind of product analytics problems, and then I had also worked on uh one of the key problem which I worked on was like connecting our donors to actual people who can donate. So, that that was another problem which I worked on. So, quite a few different domains and a lot of the first one first company was like, you know, learning the fundamentals like data engineering, cloud, and all these things as well. Um yeah, I worked there for 1 1 and 1/2 years and then I moved to Savant in Actimize. So, I think I'm going up on like 3 years outside the industry now. Okay. Yep. All right. So, I mean, so what technologies or what skills like did you bring, you know, cuz obviously you went through a lot of getting a lot of schooling, looks like, and I think it's like you had a computational science, you know, aspect to your masters and obviously you did a lot of kind of data science type work in your PhD even. So, like how did you you know, what skills and what technologies have you carried over that have been kind of crossover between the two industries? Yeah, so while I was doing my masters, I initially started working on this tool called nuclear magnetic resonance like NMR, right? There my master's thesis was primarily focused on the processing behind NMR. So, before even you get your T2 distribution, so that there was a there is a processing that goes behind it which is basically signal processing aspect of it. So, that is a easily transferable skill to data science field. And my PhD thesis, even though it wasn't it wasn't petroleum engineering, but the research was developing unsupervised learning algorithm. So, the one of the cooler algorithm which we developed was say all three different all three of us are talking in the same room and we have a multiple different mics in the room. This algorithm can capture all the voices and separate it out. So, it's an unsupervised learning technique. So, that is what I worked on but applied to NMR, but you know, that technique can be applied anywhere else. So, that that is one of the that is another transfer of skill. And pretty much like the all the six years I was doing research, I was working with some sort of data and and maths and like I I was familiar with all the libraries. Only new things I had to learn was cloud, big data when I went to other industries. So, those are some of the technology which I had to put in effort and pick up on. I've heard this question a number of times. You've probably heard it, too, being in on on your side, but for the for the average petroleum engineer or manager or somebody in that realm that is not day-to-day with with AI or ML, what are some of the things that you see people saying AI and ML can do and what are the things that, like what are your the realities? Where do you see the biggest discrepancy as far as um that goes. And I know that's kind of a loaded question, but I, I ask it because personally, I feel like it's used in a lot of different ways and ways it shouldn't be, and a lot of people still have no concept of like what all is in. It's not just like, "Oh, we click a button and then the model's there and it works perfectly and it was super cheap and it worked the first time and all of these things." But can you kind of elaborate on that? You You just saw a really good point. One key question which I would answer before even doing any data science work is do I need to use data science, right? Means can I solve it with just a just a simple physics-based solution? Not not not a simple solution like can I solve it with some other domain-based solution? Doesn't really need to be data science or or modeling all the time. So, the first question is do I really need a data solution or data science solution here. Uh so that that I think is is also like somewhat like over sold because I feel like, you know, everybody throwing the data and like solve the solve the problem like it's it's not that simple like sometimes you have to honor the physics or other domain. Uh so that is that is one key thing. And another another key solution like another thing which I've heard lot of people really struggle with is um bringing the model to production. You know, not not and not everything uh come comes to production like you people or or like or to say people don't really plan well in terms of overall data workflow. So they they set up like means what what if the model fails when you are about to do like deployment, right? Means so you need to test it the as a system as a whole first. Plan to fail fast, you know, if if you want to take the model to production. I've seen a lot of like I've done this mistake myself when I started out like, you know, waited till the like, you know, the last minute when you're closing on production to see how well the model actually behaves in the data which is going to be applied to. So that is that is one other key pain point, I would say. Yeah, I think it's a huge thing to talk about is just like do you need data science? I mean can it be more heuristics? Can it be, you know, rule based? You know, just um I think there's a lot of things where and and again and even if you are about to use some data science like don't jump to the deep learning model like, you know, you know, start with your, you know, linear models. Start with things that are actually explainable that you can, you know, show a manager and they're like, "Okay, that makes sense to me." cuz otherwise it's just some black box that you're not going to get buy in or it's going to be very difficult to get buy-in with. Yeah, um that is a transition that happens like, you know, when you grow come into the industry and like, you know, uh take take spend some time because initially you want to do apply all the complicated model as possible. Then, as you mature, you want to start with like more simpler model. Like, for the analogy which I have is uh don't if if you want to cut a cake, don't use a sword, like just use a knife, right? So, [laughter] uh just just just a simple model which which is like fits your solution. That's uh that often is more explainable and more easy to deploy and more more computing uh like more compute effective in most of the time. So, that that is that is why we go with a simple model. So, you're saying in like your grad school and everything that um you were yeah, I think it was more probably on-prem or even just local type, you know, data science work. Um but now you've moved into like, you know, with the consulting, some cloud stuff. So, you know, where you're at now, I mean, cuz again, I think obviously when it's a autonomous uh tractor, it's got to be something's running on the edge, right? But, I mean, how do you guys ride that, you know, are you pushing stuff into the cloud and pushing it down to the edge like in the tractors or or how do how does that kind of deployment work and how are those models trained? Yeah, um so like, going back to grad school days, it's the scale wasn't there. Uh like, it was all like, it's not like it's never was never a big data problem. Uh it was all simple data. Uh mostly clean data as well. So, real data is always messy. So, that difference is there. Uh in Sabanto, like, yeah, a lot of a lot of the models has to be deployed on the edge. And uh we have a um we have an edge edge camera. And uh we want the model to be able to work there. Uh so, like you said, uh Bobby, so the whole whole workflow is to build and train solution on the cloud. Some of the tools that we use is SageMaker, right? Like we, we're building like a whole end-to-end pipeline to like take the label data, clean it, you know, like put it in a proper not clean it like put it in a proper format, pre-process it, you know, and then train the data and then process into a format which you can work on edge and, and, and naturally deploy there. So, that is that's kind of a whole workflow which you're working on. And yeah, again, how do you develop it is is actually interesting because if you do all these things and in the end doesn't work on edge, you you spend a lot of time and and money wasting all the resources. So, you start with something which works on the edge and and then go back, you know, then go and label some small data, see if if it works there and then it's it's a very iterative process. Like if you heard about design thinking, right? Like so it's it's very non-linear process. That's what we do here at Sabanto. And I can I can just speak to the data science part of it. Like there are other people who are more on the robotic side. They they do even more cooler stuff, I would say, on the edge. Yeah, but this is what I'm aware of. Nice. So, let's let's dive a little deeper on SageMaker, cuz I think it's come up a little bit on the podcast, but I'd like to dive into that and like, cuz I mean, I've I've used it a little bit, but it's I mean, cuz are you guys working in notebooks? And then cuz you're able to work in notebooks, right? But then you can right from there publish, you know, cuz usually it's say if you've got a Jupyter notebook locally, you maybe got to convert it to a Python script if you really want to like run it on some kind of schedule or whatever. So, what does that deployment look like from a say notebook? You know, I think you can publish API endpoints and just kind of, you know, how does that work across the whole spectrum through AWS and down to the edge? Yeah, I can see in general like SageMaker has this processing jobs, right? Uh, which is good good with like, you know, uh, pre-processing. Like it can be any any sort of data. Um, so that is actually script file you can run. Uh, and and you can orchestrate it from a notebook or from a like pulling the SageMaker into your local and then running giving commands to run it. Uh, so pro-there are processing jobs which is is primarily for, you know, we can you can run Spark jobs there. You can run whatever pre-processing jobs there. There is also training jobs. Again, these are like Docker containers actually pre-built by uh, AWS, right? So, we it's easy just just plug and play and and actually uh, be able to run production load. So, training jobs is where you actually train the model, and then from that you the model you can bring in a facility to actually automate the model saying if it meets at this much amount of accuracy it can get it to you can get it to production; or, you can also have a manual way to say this model meets the requirements and it can be pushed to production. And this whole thing can be orchestrated by SageMaker pipelines. Right? And then um, there are version versioning I think SageMaker projects does that. And if you if you are not happy with SageMaker there are other tools to do ML Ops stuff as well. Uh, and then SageMaker endpoints are like, you know, they they are very flexible and and you can actually uh, scale it as much as you want but they are pretty expensive I would say. Um, Yeah. Well, yeah. I think that's the one rub on SageMaker. I think it's good but I've heard it's it can be expensive and Yeah. Um, I think it's one of those push and pulls I think you see a lot now where it's, you know, if you're going to run these high, you know, even like GPU intensive type things, it starts to make more sense maybe to pull that back down on premise and have your own clusters. Um And yeah, I guess curious on that. I mean, do you do you guys run exclusively that stuff in the cloud or do you have like some on-prem resources to to run certain jobs or how does that work? Yeah, when I develop, I usually like to develop locally. Um when you still get something scrappy, you don't have to burn money on the cloud and you know, just just do whatever you can. Like you don't have to train on 1,000 or 10,000 or images. Like start with like few images, train locally, make the script work, and then go to cloud. Um that's that's that's how we run things. And um this is not the this is not the only problem we work on. There is another whole module which I worked on is is more path planning, the the path the tractor has to take. So, that mostly I I develop locally. Okay. So, it sounds like right now it's there's a lot of computer vision, like you're feeding images and training models on those. Yeah. So. So, what uh what libraries are you using for that? Yeah, um just just OpenCV for like image processing and then uh like TensorFlow is is is what I usually go for deep learning. Uh PyTorch used to used it a little bit in the past for just courses and stuff. Uh I'm not really fluent in that. I usually rely on TensorFlow. Yeah. Yeah. Nice. That's uh No, I think there's a lot of really interesting Yeah. What's your favorite uh like for deep learning jobs? Uh I've I've only used TensorFlow. I I mean, I I haven't used a ton, but like when I when I've messed around with that, I I stick to TensorFlow. I mean, I guess probably six of one, half dozen of another in a lot of ways. Um just uh you know, I think I don't use it enough to you know, be like, "Oh, I need to make this seismic shift." You know, like I've I've got some things that work in TensorFlow and I can usually kind of build off of them, so. Mhm. I think there's a a of overlap in what you're currently working on and kind of where the operations side of the oil field is going. Right? Whether that's frack or drilling, um you know, the idea of doing automation in these remote parts of the world. Um on heavy you know, heavy machinery, heavy equipment, and all of that stuff. Um somebody was telling me uh like um uh uh like oil can drill by itself, like well can drill by itself. So, the the automation technique there is is is actually mathematical optimization is is what I believe. It's It's been a while, uh but the the It's It's It's the same same problem. Uh I couldn't hear a lot of what you said, Bobby. Uh like uh sorry, John. Is Is it Can you repeat it if possible? No, I was just curious what uh like what how do you see where do you see kind of the the the biggest hurdles in trying to do with the push to the edge, right? Like there's a lot of a lot of promise with edge, um you know, being able to run things locally, being able to automate things, control things. Yeah, what are some of the difficulties you're having like with the push to the edge? Or I mean, I think John used to work in that edge space. Yeah. So, like deploying to the edge, what like what are some challenges you see there? Um and again, I think those are things that translate across industries. Yeah, uh my the only problem with edge is like, you know, resource constrained environment with like a lot of these models are, you know, resource hungry. Uh so, uh can you can you slim down the model to be able to deploy on on a like really constrained environment? So, that that is that is the uh only problem I've looked at, but uh there might be other things which which other people in the company might be aware of. Yeah, but I think John and I were talking about that earlier. About I think now they're starting going where some of these uh um AI models are able to be I think you know small enough I think what was Lama 3 John you were saying but they're able to deploy those onto the phones now like Apple and you know Samsung are looking at being able to run those locally it's just I think it's a natural evolution where you can if you can run it on on the local hardware it's it's going to get better but I think it's just going to be dependent on you know processing power of Moore's law and just being again as I said being able to slim down things and say what do I really need from this do I need the full fledged GPT or do I need kind of a more targeted uh model so Yeah, that, that is the same challenge even in computer vision as well, like, oh, means what architecture do I use to actually be able to run in front on edge? That is, that is one of the most constrained which I have, because if you want to throw like all complex architecture, I did, you won't be run you won't be able to run the inference on edge at a particular frame, frames per second, which you need to actually be the model be useful in real life, right? So that is, that's where the primary constraint is. Yeah, and I guess it's something again. I think we we get into this often as a theme in our uh podcast, but you know, even talking about say about frag operations and whatever else, but like it's I mean again in agriculture, like you can't have a tractor just go off the rails and just destroy a whole you know plot of of crops. Like I mean like it's it has to work, it has to stay you know on a good path. Um, so again, ensuring that you have that you're not dependent on cloud connectivity and then that it's able to you know make decisions in real time, it's it has to be hyper critical, I would imagine. Yeah, so we have other like I said we have other object detection capabilities and we have a the... there are...there are... let me size up the things a a bit more clearly so there the automation capability of Sabanto is like, you know, we we are given a field and we we are able to do like a plan the coverage of the field, how we want to do it, and there are, you know, we also need to plan the paths the tractor needs to take. So, there are various modules or methods which we adopt. But, let let me In terms of safety measure, there is object detection capability. We have a um a time of flight sensor which can detect an object and and stop, basically. It won't It won't hit or it won't hit any obstacles or it won't run away from the field because there are there is a there is a fence which we have and we won't be operating outside that fence. And in remote locations, we have the connectivity needed to actually be able to interact with the tractor. So. But, I would say the remote connection part, I'm not the best person to answer it. There are other experts in the company which take care of who take care of it. Yeah, fair enough. And again, maybe like this next question you may not be like the perfect person to talk to, but maybe you you kind of heard But, like how how is it with adoption? Like with how are farmers and people taking to adopting technology and getting them to buy in? Yeah, some part of it is actually change management, right? Means not not everybody is used to autonomy. And like, but we convey that like our primary focus is on safety and providing cost-effective and easy-to-use solution. So, we have a we have a solid solid grounding in one of one of the industries like turf and sod. And we have, you know, people really utilizing our autonomy technology there. And and we have farmers been able to adopt for some other operations as well like tillage and you know planting. So we have we have like tested our autonomy capability in various different farming operations. So I like I would say once they see the actual tractor go it's it's a totally different ballgame. When I was working remotely I, wasn't I wasn't able to see the tractor in person. But then, when I actually see it go and see how easy it's to set up the, you know, the plan and, and the path plan and, and actually launch it and, it's it's a totally exciting thing even for me as an employee working there for a few months. Means when I actually saw the tractor but farmers when they see it they get really excited and we have a lot of adoption of our technology and I think in terms of acres we are doing way more acres than what we had done before. So it's pretty exciting time for the company actually. That's awesome. So I I think one thing we didn't really get into though was like so is this like a like a module or a box that can be tied into a John Deere or a Case International or whatever or are you manufacturing your own tractors? Yeah this is a a retrofit kit. So there are various platforms we support like John Deere Fendt and then Kubota. So these are different tractors and and we we we have a technician who set sets it up and we have a complete closed product right it's connected to a tractor and we fit the sensors and and cameras and everything and and it's good to go and the farmers can operate by themselves after that. Great. I mean and these are big tractors right? We're talking like the big tractors or even like it's combines and stuff like that as well or um Cuz I'm thinking some people may go all the way down to say a lawn tractor. You know what I mean? I don't I don't know, but I'm assuming probably a larger tractor is what you're talking about. Uh I'm not sure what I Is it combine or not? Yeah, um it's it's reasonably big size tractor. This This was the first time I've seen it. So, I I would say uh A farming tractor. I wouldn't be the best person to answer that. Yeah. No, fair enough. Sorry about uh my mom's old side that they were farmers, cotton farmers, and corn farmers. So, I've got a little [snorts] you know, a little bit of background. I I could talk enough to get myself in trouble. Um What about You mentioned uh You mentioned earlier some of the people run into issues with, you know, productionizing models and things like that, right? Like, I think you're probably well aware of about all of those problems that we have in in the energy space doing the exact same thing. What are some of the the things that, you know, people should be aware of, should consider strategies on on how to do that? Cuz we see it all the time, right? Like, Bobby and I have both have seen it where it's like, "Hey, I've got this Python script that I wrote that's doing a really good job at X for this well basin, pad, whatever, small scale," and then they try and go deploy it, you know, in a an asset or just in a lease, or they, you know, they tear things up and it just completely collapses at some point before it gets into full company production. What are What are some, you know, lessons learned um and best practices that you've seen on that end because that is a very real thing, right? Like, we've got a ton of intelligent people that are always working on optimizing things, making things better, making them safer, automating, etc., but scaling that out so that the whole company or a whole asset can take advantage of it is still a very hard thing. And I don't I don't I think I don't think a lot of people really appreciate what that takes um until they are looking at it, you know, it's looking them in the face. Yeah, so there there are two piece on a high level, right? Uh one is why why this gets killed at the end is um one you haven't communicated the value to the upper management. So, before even doing um starting anything like can you put a dollar amount to this how it's going to contribute to the bottom line, right? Uh not every problem will be that will be very obvious, right? Um so some some problem is ambiguous, but can it affect the bottom line and can you communicate that to the management and can you buy in get the buy in and get the resources uh through to take it to production, right? So, that is one thing. Another thing another reason which which it gets killed at the end is uh you know, the change management aspects of it like not everybody uh there are other stakeholders who are more used to like physics-based modeling or rule-based modeling. They they they don't really want to use machine learning. So, can you get their buy in, bring all of them in and like really understand what what what would success look like and what would they will they will be willing to adopt and use. So, the success metric is people actually using it uh using it successfully and being happy with it, not really a 90% model or 95%trained model. So, that is that is one of the key, I would say. Uh so that that those if you do those things at the beginning uh more more correctly and then uh think about like testing the model in a non-linear way. Uh so, testing as a system is very important uh before you even uh say anything. Like for example the model can model can be this much level of like 90% accurate um in like when I can do it on SageMaker, will it be that much accurate when I transfer it to edge? So, those are those are things which which gets in the way. Um and like people um I've I've made that mistake as well. So, uh think about like the end goal and kind of work your way way back. Uh so, that that that is kind of helpful. And uh uh again, uh also like starting simple and then like making it more complicated as needed. So, more of a continuous improvement kind of model uh development. So, push something uh get it into production, see how it's working, maybe in shadow mode, and then revert it back, and you know, build a much better model. So, that that is how I would build it. Uh and then uh also like not oversell it, you know, just communicate honestly what what the value is, you know, not not really uh like say this is going to solve all your problems. Means, I usually try to undersell it a little bit, you know, this might solve a little bit of your problem, and then if if if it gets better, then then it's it's great. So, that that's that's how I I would do it. Uh again, different people have different way of doing things, and this is one one of my opinion. Yeah. Do you have Yeah. Just a follow-up on that. Do you have any Cuz kind of what you described right there is like the CI/CD process, right? Continuous improvement development. What Do you have any specific tools or anything like that that you would recommend people use for that um kind of s-making sure that that's set up to, you know, for productionization, for scalability? Cuz I've I've never used them, but I know that there are quite a few tools out there that, you know, try to help with that in some aspects, but uh I I still think that CI/CD concept is still fairly new or fresh in the the energy space, at least. Um especially with everything being in the cloud now. Yeah, SageMaker is one one such tool to easily help you with with that, right? Actually build a and even deploy the model and then improve on it. Uh get get feedback and uh there are there are tools where like there are tools like uh experiment tracking and that, you know, see where you are at in terms of like accuracy and and then build on it. So, there are there are there are already existing tools. This uh use something simple, I would say. And whatever environment you're already in, there are there are tools like this. Yeah. Well, cuz I, the reason I asked that kind of loaded question, uh, coming from my last role at the at the edge company, is that's one area where people don't even know. Like, on, it's one thing when you're doing it with cloud deployments or on-prem deployments, you're just your CICD pipeline, but then once you get to edge, it's a whole 'nother can of worms, right? Because it's like, well, hey, now we're we're hoping that there's connectivity or we need a minimum amount of connectivity uh to be able to deploy this even. That's one the issues that my prior company ran into. We were doing push-button Kubernetes deployments on these edge servers, and they would work great if you had, you know, three three to 500 megs of internet connection, but if you got much below that, then or maybe it was a meg. Um but whatever it was, there was a there was a very clear threshold, and so then it's like, oh, well, [ __ ] How do we how do we architect the device so that if that happens, it's not going to just brick the device and shut everything down, you know, in the middle of a up you know, simple up to software update, right? Uh-huh. I think that's a a it again, this edge problem. The edge is coming to the oil field, right? Just like it is in most other places, but I see a lot of people I foresee a lot of issues that people don't know that they're going to have until they get there, and that's why I ask all these these questions. So, one thing I want to hit on, on the problem you talked about, John, too, and I'm going to get Naveen's opinion on it as well, as I think, again, people can do this really cool model and it works like in that localized area. And then [snorts] you get, and I think we talk about problems of that, maybe how it doesn't generalize, like, you know, maybe it worked really well in a cotton field but not in a corn field, or, you know, just certain things that you didn't anticipate. But I think another thing that people really understate and I think maybe it's better in Naveen's case because they're the company that owns the data model and the schema that the data is producing into but is like the data engineering aspect of it. is like, yeah, again, you can have a an engineer or a data scientist that does a really cool multivariate analysis, but then do you have the data infrastructure and have the data engineering set up to feed that model continuously to be able to run that in production? Because, again, it's one thing to do all the pre-processing steps and everything, but then you need a pipeline that's doing those processing steps in real time that can provide that data consistently and kick out outliers and whatever else you need to do. So [snorts], I think that's huge, cuz I mean, I I I did a really brief consulting thing with an autonomous pipeline company, and they had grown by acquisition, and the data schema for each company they acquired was different. So, it was like they built the auto pipe thing for, you know, for one pipe, but then they want, and they're like, "This is great. Well, we need to move this over to this pipe." Well, guess what? This this is coming from a pie Historian; this is coming from some other totally different system. So, they had to rewire every time they wanted to go deploy the auto pipe to the different, you know, sides of the business. They had to re do all the data engineering because it was coming from a different source system. They they never consolidated and, you [snorts] know, aligned the the data systems. So, I'm just curious what you've seen on that side, Naveen, from, like, even the data engineering side and how that, you know, plays with the data science side. I would say first of all I'm very thankful to all of them like without them there is there is no data science but yeah on that end like for example and are there are multiple different tables scattered uh with different information and and they don't talk to each other. That's that's a big problem and I've seen you know data with no entity table and like you know, I I cannot really make sense of what what those data are. So there I've seen all all those type of problems and the pipelines are not really you know robust enough to handle the load as well. That is that is another problem and like I feel like you know it's it's not just a separate person job you just throw throw it at them and like let them let them do the you know data engineering where every every stakeholder has to be involved like you know from from the beginning to the end like even even the business stakeholders and and like it has to be involved on the on the first first when you problem solve it and then set up the pipeline you know like kind of what what schema you need and stuff like that because down the line um you know you never know how how badly you messed up like you know. So you have to go back and redo it and all these things. So it's it's always good to be like you know good to have as as many hands at the beginning and like you know discuss things over and like you know maybe argue over a few things so that you get the things right. Yeah. Yeah, 100%. Okay, so let's kind of change gears a little bit. Let's kind of go back to the to the beginning. Like how did you get into you know oil and gas and or technology? I mean were you the kid that was ripping apart computers or when did you you know see your first computer or when when was that kind of ignition moment for you? Yeah, oil and gas. My bachelor's was also in chemical engineering, kind of like slash down downstream, and, um, like I came here to do masters in petroleum engineering, but then, like, after a few few months into it, I started working on an experimental lab, and then I met some I met a mathematician who was like a mentor to me and, he he kind of helped me transition into technology space. And like I didn't know even I never knew about Python and stuff and before I met him. So kind of like I think it was like 2015 and that's that's when I first started you know really coding coding for research purpose and then that's that's a gradual change. I wouldn't say, like, um, growing up I always wanted to do, do this. But, but when I actually started working in the data space like when I started working with the data in 2015 kind of made all the difference. And I was, I know that this is something which is going to stick around; and, and because I used to spend, like, hours, and it, it didn't feel like, you know, I'm putting in the work. So that's how the transition happened. It was it wasn't like all planned and stuff like that. It just happened by itself. [snorts] So looking at your like the new year of age I know John and I had a gentleman there that we were you know really wanted to get to come work for us and he ended up going to work for Corban instead. But did you know Fahad Siddiqui? Fahad Siddiqui? He was one Oh yeah yeah I know him. He was a fellow I think a PhD fellow I think. Yeah yeah I know him. He was a post doc. Yeah. Yeah. Yeah. He was pretty slick. So I don't I don't know if he was he was the one or not, but he'd done some pretty cool stuff with with Yeah. computers and programming too. Yeah oil and gas I have I met a like a lot lot of smart people and a lot of people I admire are are in the industry. Like for example NMR like the tool I used to specialize in, there are many people in the industry I used to admire and I still come I still volunteer in in SPE and stuff like that. And uh uh like in the mathematical problems which which the which oil and gas people solve is uh it's it's not well known, right, outside the industry. Uh the there is a for example, pressure transient testing is is is not the mathematics behind this is not very straightforward. It's it's very complicated and So, as far as uh because I'm always interested when I think like when a solution from another industry can be, you know, used across industries. I mean, have you seen have you seen that where, like, you were maybe able to bring, like, a solution or maybe something from the oil and gas industry to a consulting side that you maybe they weren't thinking of it from that viewpoint? Or, conversely, you know, are there some stuff that you've learned now, say, in consulting and, uh, at Subanto that you're like, "Man, if I ever went back to oil and gas, I would love to kind of bring this solution or bring this way of thinking over to the oil and gas side." Uh I would say uh like I I never worked in like oil and gas like full fledged I was doing my PhD and then I was had an internship briefly with Halliburton. So, but uh I I I I I don't think I know enough to see what I can bring back. Uh but I I I have like uh but oil and gas and that 3 months I spent was also a starting point for me to get into data space because a lot of other interns were working on data science projects and I could see the potential of computer vision and and things like that. Uh, but in terms of going back and bringing skill sets back, uh I, don't I don't know what what what people what problems people are working on and, and I haven't haven't really caught up with all these things. But I've seen people, like whatever conferences I've attended, like I've seen people working on a RAG. You know, I've seen people work on, like, as one of the problems which people keep always talking about, is bringing AI into production. So there are some of the things which I which I feel like, you know, can be applied back by everywhere not on just one of these. I guess, bringing up AI I mean so I have to imagine that in your current role, I mean, obviously you're doing a lot of deep learning. And I feel like, again, people I think rush to the new shiny thing, which again, deep learning was that, you know, a few years back; but seems like you guys are still leaning into that. But is there also are there applications of something like this LLMs or AI in the work you you're all doing at Sabanto? And you know, do you see that supplanting or working with, um, deep learning? Like, you know, does one make the other one obsolete or or they work well together? Just curious what you think there. Yeah, currently it's it's it's deep learning, but it's computer vision convolutional neural network, so nothing haven't haven't done a lot of LLM work here, but in my previous role at Credara I did a little bit of NLP project. It was more on topic modeling, so just just a basic project, but we had like, you know, internal groups where we try to learn and specialize on NLP and stuff like that. So I was part of it and I was I had also taken courses to do like train myself in NLP, but I haven't really got a chance to work on it and means the I was talking to somebody from the technology space as well about this. Like NLP you need to be a little more curious about the language itself and and like understanding of linguistics and stuff like that. Um I I I feel like, you know, I wasn't really into that. Uh I find computer vision problem more uh fun and and interesting like so. But but would love to try it sometime, but uh right now I haven't I haven't I haven't said and done much in that. Mostly just learning and self-learning. So. I think that's the interesting thing about the language models and at least the language models today is and I didn't realize this until someone explained it to me last week. Um but like a language model isn't isn't just one thing, right? Like, there's a bunch of stuff going on under the the hood of that model that understands sentence sentence structure and types, you know, forms of uh of word like the types of words, what what they're being used in, how they're used; it can predict where they're going to be and all of these things. Like, there's so much going on underneath it and uh and yeah, like people, just it's all lumped together, right? Under a language model, but there's all these things that kind of compounded that have been rolled up into that model that are happening all at the same time that are really kind of incredible actually when you kind of break down what it's able to do. But it's [snorts] uh, I, I'm curious: walk people through, cuz we've talked about computer vision kind of high level, um, and I think people generally understand what computer vision is, but walk people through what it actually takes to build a computer vision model because and yeah, it it's not just computer vision, right? It's basically any kind of model, but I I don't think a lot of people truly understand or appreciate the amount of work and effort that goes into building your own model out of, you know, from nothing essentially. Yeah. Um computer vision um as as a field is like, you know, bringing eyes to whatever things you're working with. Basically, you're using a camera sensor and that's capturing uh all the visible part of the spectrum, right? Into into images, and then you actually process these images to run inference on them, right? And there are certain there are there is a certain certain type of deep learning model called convolutional network, which does extremely well on like processing these images. And now they have started using transformers, which are also like, you know, used in NLP as well. So, uh those are the architecture. What it takes to build a computer vision model is um sometimes it's not not too difficult. Um it's it's like it's about like understanding what exactly you need. So, there might be already pre-trained model, which is exactly suited to your use case, you know. Uh you don't have to uh train anything. Right. You can just use that model and deploy it. But uh there might be like a lot of cases like it's not that ideal. The the environment is totally different, you know. Uh for farming there there is not a lot of data sets for farming, and there's not a lot of models trained for farming. So, like basically it's like teaching a child, right? Uh if the child hasn't seen those, you know, apple or [snorts] what whatever you like show your child when when it's growing up, then it won't really pick it up, right? Yeah. So, it'll take some time to train this model uh with with the appropriate amount of data. Uh and uh uh the difficulty I would say um is understanding what exactly uh the problem you're trying to solve for. Uh so, some might be a classification problem, some might be a semantic segmentation problem, some might be a uh instance segmentation problem. So, there are different types of computer vision problems, and there's a post estimation problem. So, that is that is one part of it, framing the initial part of the problem. Uh and part is like what Bobby said, like bringing the data, putting the pipeline into place. The whole ETL process. Seeing how actually things scale. But understanding the fundamentals is very important because when it comes to tuning the model, if you don't know what what like you know the whole overall architecture is, then you might be stuck. So you might not be able to improve accuracy more of the model beyond a certain point. So Yeah. Yeah, I think it's I think it's really cool on one hand in the the tech community that there are these kind of this, at least on the computer vision side or the data model side, there are like base training sets or base models that you can use to, you know, get a head start, so to speak; but you hit something perfectly on the head where it's like they don't know what if you don't teach it, it doesn't know. Right? Like it is literally just like a child. And so like when people see things from the [clears throat] language models for example, right? They they're like they think it's sentient or they think you know they're the answer's always right and stuff. It's like it doesn't all it's doing is retrieving information. It's not it's not coming up with it doesn't have an intelligence. It has knowledge, right? Like right now that's the best way I would describe like GPT is it's got a a ton of knowledge, but the intelligence, the way it understands that knowledge is still very young in the kind of Mhm. [clears throat] life cycle of of what whatever that becomes, but uh people you know people confuse just good programming with sentience or like true you know it's it's it actually understands what it's doing or it's creating or generating things and uh that's that's not how they work. [laughter] It has memorized everything and it's spitting it out. Right. Right. It That's all it's It's just it might rephrase the way that it's written, which is cool, but that's again, you we've just coded language and language structures and and all that. Um, one thing that you brought up that I kind of wanted to pick your brain on, too, is cuz I mean, again, you guys have a a commercial product, and again, people think it's open source, I...I can just take these, cuz you said maybe there's a model like off-the-shelf that I could use. But I mean, um how familiar are you with like some like see the different licensing, and then like being able to identify [snorts] like this is something I I can actually use, or is I can test with this, but I probably actually can't use this in production cuz like it's protected? There are there are various types of licensing. MIT is usually, you know, uh the where where it's very liberal, and then you can actually use it uh for commercial purposes, is my understanding. But there are Apache, and like there are few other licensing, and uh you you cannot uh you you might get into trouble if you even with libraries, not just models, right? Even with libraries, uh not a lot of libraries uh let you use for commercial applications. So, um I think I I ran into that problem a couple of times in my previous job, as well. Uh so, uh that is definitely to be aware of, you know. I've got something that made me think of this. How do y'all handle weather with the computer vision stuff? Right? Like cuz I always whenever people start talking about models, the first thing I go to is the outlier cases, because that's the hardest part, right? Like that's what a lot of people don't ever think about, is it's like I can train uh making this relevant to the energy side, right? Like I can train my model on thousands of frac stages, but if it doesn't know if I didn't give it a screen out and show it what a screen out was it will not know what the hell to do when we encounter an actual screen out while we're doing a frac job, right? And so, it's like you have to give it those edge cases, or you hear about like the Teslas running into the back of, you know, uh garbage truck because it didn't have a garbage truck programmed into the computer vision model, or whatever. Um but like tell me kind of I'm very curious how y'all kind of deal with that, and then making it, you know, kind of higher level, how what's one way that people can kind of approach those anomalies and stuff? Because those are truly the you know, the the things that make the big differences in these models at the end of the day. Yeah, I have two answers for your question. So one is like it's just not weather, right? It's various scenarios. Like dark and stuff. So you need to do proper scenario testing and like train for those scenarios, right? So that that is very important. Like you could, if you if you miss a certain scenario like while you're training, you might actually miss it. But [clears throat] when you do an extensive like testing at a scenario level that, for example, person could be it could pick up a person standing straight; can it pick up a person standing sideways, back? Means what are different scenarios you could be facing? Can it just pick up an head? So those are those are different things you plan for. And definitely like you know, lighting condition and weather these are different things you can train the model for. Another part of your question is not one sensor can do everything. So there are thermal sensors, there are radio like you know, radar sensors and there are lidar sensors and you know, there there are different and and there is camera. So there are different types of sensor, each one is good at different things and you need to put together all of them to actually build a state of the art model. I'm not saying we're not actually doing all of this, but that's how our state of the art model is getting built. Also to that question, since you are familiar with automating moving things there was I know there I'm a car guy, I'm wearing my donut car hat. Uh there was a debate for a while I guess on the automotive side around how they were going to auto automate driving, right? Like Elon went full in with computer vision, so Teslas have cameras and they've got their own custom GPUs and all this stuff. But then a lot of other people were looking at like lidar and sonar and radar and all these other things. I'm curious, where where do you from your real-world experience, where do you kind of see that? Is it a combination, like a hybrid of both, you think is the best probable solution that people end up moving forward with or This would be my opinion, not my company's opinion, but yeah, yeah. I would say like my opinion on this is we need more sensors, like all all different kinds of sensors. But like our industry is not that complicated, right? Like for especially self-driving cars, you know, [laughter] it's very complicated. So, yeah, yeah. What's what's the outlier cases there [clears throat]? [snorts] Yeah. I think it's Exactly. And it it means uh uh uh it's also like life. Like when when because there's not a lot of like, you know, farming is is a relatively safe environment, but like when it comes to like highways and freeways and all these things and there are a lot of lot of things to be looked at. More variables. Mhm. Yeah, exactly. Well, awesome. Yeah, man. You want to jump into a speed round now or that's fully buy. We're already at 55 minutes. That's awesome. Um so, this is speed round. We'll just pepper you pepper you with some random questions just to learn a little bit more about you and and then we'll wrap it up. Um Who are Do you have any good uh either blogs, social follows, YouTube, people that that you you like to follow on the kind of data technology side you'd recommend? Yeah, three blue one brown. I think I think that's what that's what the channel is. Like they show you a visual of different, for example GPT how you can see it visually so that that is one of the channel. I'm more of a fan of lectures like I go to MIT open source lecture and I kind of burn through it. You are listening like the Carnegie Mellon like SQL lectures? They were really good. Serial? I haven't seen it but I've seen some I see from CMU I've seen some previous lectures mostly for my own work but yeah and not really SQL. I will check it out. Yeah. A follow up to that question. What what are what are some ML courses or or channels that you would recommend people to follow that are looking into getting in or just increasing their understanding? The Andrew Ng course has gotten way better over the years and and that that's something a good place to start. And I think I've said it in another podcast as well: like just, just getting to know somebody, you know, who's actually working and kind of seeking their guidance, that will, that will be like, you know, a good way, because, you know, everybody has different backgrounds and stuff like that. And every, not, not everybody has to get into, like, not everybody has to do data science; like, they can do data analysis or they can do data engineering. So it depends on what you like and what you are naturally attracted to. Yeah. All right. What's your favorite cloud platform? AWS. Yeah. No question. And yeah that was a quick answer. What what sets it apart for you? Sorry? I said that was a quick answer so like what sets it apart for you? That's something I've used so far. That's something I know very well. And I also like the one that like you know basically the services there is is huge and and it's been there for a very long. I like DC PS well I have used it. As you haven't touched it. So Okay. Um favorite video or board game? Uh board game? Or video game? Or video game? Video game Call of Duty with my friends from college on this. All day long. Yeah. Yeah it's all fun memories. Yeah. All right the last one let's go uh what's your favorite open source library? Um open source library um I would say um It's a pretty hard choice. Pandas yeah. Yeah. Have [snorts] you have you messed with Polars yet? Uh no. I'm I'm a Spark guy. I like I I I like Spark. Okay. Nice. Cool. Sweet. Where Naveen where can people find you if they want to reach out or or ask you any questions? Yeah I'm usually active on LinkedIn. I used to be more active nowadays little less active. I have a Yeah uh but yeah I'm LinkedIn is is is where I'm active. I don't have a Twitter or anything else. Cool. Awesome. Well thanks for thanks for being on thanks for yeah thanks for reaching out to us it's great to have you on and kind of see those uh crossovers between uh the energy and the agriculture business. So excited to see what comes out of uh Sabanto and your work. Thanks thanks Bobby really excited to be part of this podcast. I I listen to a lot of your like podcast. I think I listened to even Jessica's one recently. Awesome. Oh nice. Yeah she's she's very good. Well we appreciate it. I'm, I'm excited to see what, uh, I think like we've talked about: the farming is kind of almost like a a bleed into the energy space because we have a lot of the same kind of constraints around remote, lots of auto, lots of big equipment, trying to reduce people, to reduce risk and cost and all these things. So, I've I've never thought about it that way, but I'm going to be looking at the uh the agritech space to see what's what's coming down the pipe for for the energy side. So, really appreciate Naveen.