Energy Bytes

Dingzhou Cao, Senior Data Science Advisor discuss how AI is transforming oil and gas operations. Dingzhou shares his journey from assembling computers to pioneering AI applications in the energy sector, highlighting real-world case studies and the challenges of integrating AI into traditional industry practices. The conversation covers automation, data management, and future trends, offering unique insights into how AI is reshaping the energy landscape and the potential for further innovation in the field.

Join the conversation shaping the future of energy.
Collide is the community where oil & gas professionals connect, share insights, and solve real-world problems together. No noise. No fluff. Just the discussions that move our industry forward.
Apply today at collide.io

0:00 - Getting into Data Science
1:03 - Path to Energy Industry 
15:48 - Joining Anadarko 
20:08 - Anadarko Company Overview 
21:40 - Data Infrastructure 
23:50 - Build vs Buy Decisions
26:50 - Streaming Analytics Tools 
28:16 - TIBCO StreamBase 
30:30 - Apache Kafka 
32:38 - TIBCO Spotfire 
36:48 - Vendor Lock-in Risks 
40:37 - Alternatives to Kafka   
42:01 - Model Training 
47:35 - Career Advice 
50:31 - Jobs and Titles 
55:30 - Containers Benefits 
56:41 - Tools in Use 
59:17 - Managing Packages
1:00:26 - Hardware Acceleration
1:02:41 - NVIDIA GPUs 
1:03:33 - Hardware Acceleration 
1:08:33 - IT Automation 
1:17:53 - Wrap-up
1:18:03 - Favorite Cloud Platforms
1:21:14 - Favorite Data Science Courses 
1:22:52 - Favorite Open Source Packages 
1:23:07 - Student Career Advice

https://twitter.com/collide_ai
https://www.tiktok.com/@collide.io
https://www.instagram.com/collide.io/
https://www.facebook.com/collide.io/
https://www.linkedin.com/company/collideai

What is Energy Bytes?

Welcome to Energy Bytes with John Kalfayan and Bobby Neelon. Your essential guide to the intersection of data and energy. This podcast dives deep into the world of energy, shedding light on how data, AI, and technology are revolutionizing this sector. Each episode equips listeners with insights into the most efficient tools and resources, paving the way for a data-driven future in energy. From technical nuances to broader industry trends, Energy Bytes offers an unparalleled perspective on the evolution of the energy industry. Join us as we decode the algorithms of energy, one byte at a time.

Welcome back everybody. Uh here for another wonderful, beautiful Friday in Houston, Texas. Um welcome back to Energy Bites. John Cowan, the the Rad here. My wonderful co-host Bobby Neil. What's up? What's up? You're doing good. And today we've got my good good friend uh Dingu. I'm gonna butcher this. Dingu Cow. Correct. Close enough. uh senior adviser uh in data science. Right. So Ding, I've known Ding for a couple years now, but Ding's I'm really excited to have you on the show just because you've got a really cool background and story and you've worked across a bunch of different even before you got into energy and oil and gas. You've had a bunch of really cool experience. So thanks for coming on. Yeah, thank you for having me. Y appreciate uh Yeah, I mean I know I mean John's told me a lot about you and you know we have haven't had the you know the chance to really meet. I think we talked a little bit on LinkedIn. I see your face a lot in the in the [laughter] video. Yeah. And Colleen. Yeah. You guys do a very good job. Appreciate it. We try. Um well, so let's you know the first kind of how we normally open is just kind of getting into your background. How did you initially kind of get into either energy andor tech up to kind of you know current um where you're at now? Right. So like when did growing up were you into electronics and software and stuff? before you I mean I would I cannot tell you the number of computers I broke growing up for one reason or another just because I like to mess with stuff and learn and uh and that sort of thing. So how did you kind of get into into that space initially? That's a that's a long story. Let's say computer science that one. So when I was in college, my major is mathematics. Okay. But we do had a minor in the software development. So that software engineer we could say that one. So actually when I was in college, one of my part job is try to assembling the machine for the li college is still there. So we can get some tips, right? So like a let's say a female's friend asked me say Dean I want a computer can you assembling one for me then I will go to buy all the the hard yeah [laughter] like the motherboard like the driver like the CPU and put them together and then hang over to that's when I was in college that 2000 mhm around 2000 so at least it had gooies and stuff back then right it wasn't it wasn't all command line no it had a windows we also we had inst is not in China is not a [snorts] official is always like the fake one. It's like take it from the coffee from the city. That makes sense like that. But but it's working. Yeah. So actually that's the college when I touch that. But my my college is actually we are even learn the assembling language. So it's like right on the hardware. Yeah. The hardware. So actually I'm part of the mathematics math and also software. That's also sometimes hardware like the seven layer of the network. So that's why I can do what I do today. Yeah. So that's why I go into the tech right let's say computer science or software. But in order talk about the how to enter the energy industry that's interesting story because when I graduate actually I graduate as a industrial engineer they actually the branch of the mass is like operation research. Okay. So do optimization. Yeah. We used to ma make fun of the IE in undergrad as a mechanical engineer. We called them imaginary engineers because it was all statistics and and business focused stuff which is hilarious because it's the most practical like everyday engineering form I think that there is now that I'm out of you know in the real world and stuff right but it's funny too because industrial engineering genuinely you can use those principles across everything. It's like is you can find it's optimization actually optimization you had object function you had a constraint and then you solve the problem even right now for the ML AI stuff it's still optimization process back in the the same right you need to solve the object function to fix the function so that's the same so back to the story when I graduate 2011 as a PhD I actually work as a small company in two Mhm. And I never think one day I step into energy or light or gas. But one day when I was in San Diego, I get a phone call from some like 4 4 a.m. in the morning. I get a phone call from somewhere. They say it's a something energy. I thought it's a DTE energy. DTA energy is a utility company in Detroit because my I get my PhD in Detroit Wednesday. So I thought it's a DTE. So I said, "Yeah, good. I will check my email after later like six or seven wake up and then I say oh this BP so actually it's BP call from UK and why is because like Mando I do statistics I reliability so actually my major in the IE is reliability analysis got when I was in school I actually analyze the the brand tank the tanks reliability at that time for the general dynamic len system at that time So, but uh that's a another story. So, it's like liability. Is that almost like some like actuarial type? Yeah. Like similar like Yeah. So, like let's say when you buy a car, right? You had a wrong they give you five years or like 10,000 miles. Yeah. And when you find out, oh yeah, it's just after five years, it's break. Yeah. Exactly. That's the stat design. [laughter] That's why D at the time is Koft is a small company. So BP at the time become a condo. So they try to they actually create an organization called global well organization. They try to improve the the the reliability of the well equipment or something like that. Interesting. So that's why they call me and actually I don't know vein. I submit the resume. It's I don't know vain. It's just like they pull from the pool. Yeah. And see oh this guy has. So that's the the first call I get from them. I I reply and then after that the hiring manager my manager rhino is still a good friend. It's just we get a 4 hour phone call. 4 hour. Wow. After that they give me the offer. Wow. Not come to us to Houston to check what happened. So at the time and my father is interesting my father reading the news about BP will pay like $18 billion to the government because Mando my father look at the news in China and say are you sure you're going to BP? [laughter] So actually I don't know exactly what happened and I never had a chance to come to Houston because my manager is in UK. So even I want to onside they not on side. So I tried to someh channel to talk to someone in in BP and say what is it look like? Is the company still good to like join the company [laughter] and find the the people is still a best friend right now Chinese people a Chinese friend say yeah it's good everything is good just come. That's why I join join BP and after that it's like using right now. No, that's that's a that's a wild story. That's uh I I vividly remember that point in time because I was working in the field and my roommate uh he like he was interning that summer and but he had he would randomly just like trade stocks and stuff and MacO had just happened and he was telling me he was like I just bought so much BP stock and I looked at it and I was like why it's bought like it's so low and he was like cuz BP has more money than half they're coming back country like yeah BP is too big to fail essentially is what he was telling me and that the price will come back at some point. This is just an overreaction and sure enough, I'm sure he made pretty good money. Shout out to Dennis on that one. But, uh, I thought that that's hilarious. That's that's never waste a good crisis or something. Well, it's also good that I mean, it's also in my opinion good that to know that our in our industry will go reach out and find, you know, experts in those subjects when bad things do happen because they are trying to conduct. Let me give you another story. It's the same from my manager. He hired me. They also try to reach NASA to get a get a reliability guy from NASA because NASA you know that that crop actually the spacecraft that one is very reliable otherwise you but NASA say no you guys not going to do that so they decide to come to BB because they think that we are not going to do it right follow the same procedure actually that that's approved that when I for my three or four years in that company I try to do something but actually it's too big to compare Yeah. So, and then so then you ended up at like at uh you were at Ford at one point, right? You were at a mo were you not at a motor vehicle company? Motor Ford Motors. Did you work? Oh, yeah. Actually, that's when I was in a PhD. That was before. Oh, that's also a good story to talk about. Okay. Uh when I was in Detroit, right? I'm in the school. So at that time uh 2008 actually uh 200 okay that's the let me straight up so the first time I come to Wayne state in Detroit I went to the math department okay and then after one year I just feel is so bor [laughter] I mean I don't blame you either it's like oh pro that kind of stuff I like the application or like apply so I like apply a prime map so I try to switch from the math department to the industrial engineering department. Although it's still the same, it's just like more focus on them. Yeah. Right. I I did a theoretical math bachelors like all I did was proofs the whole time and it was boring. [laughter] It was terrible. It's miserable. So at that time I I went to the professor in IE department. They say I don't have a fun for you right now but I do have friend in in in for the motor company. If you can get a contract position for that one, maybe it's you can come to me. So actually it look like this like fer we had a collaboration with the IE department I just work as a student and go to Florida for work three days a week and two days in school. So that's the plan for so I do the interview I do the testing that kind of stuff and finally I get the offer so I get 85,000 each year from not from my not from me it's go to school I get 20,000 something like that. Yeah. [laughter] But but it's good but you know that at that time the advisor in the for the motor company actually he give me a test [snorts] well you know how the test is like it's like they had equation to generate x and y equal to x something something like that and then plus the noise plus the noise. So it's a y = x function of x plus the noise. Okay. And ask me to predict the the the y. Okay. So actually that's e I just use a supported machine and then leave the whole thing and perfectly. So I get a job. That's awesome. So actually that I I went to for for like three days a week and two days school. What were you doing at Ford at that time? Ah that's also interesting related to the oil price. So I think 2007 the oil price is 149. I remember [clears throat] 149. Yeah. What four they want to do is like they try to use the machine learning to predict the the driver's behavior so that they can fight turn the engine parameter to say the fuel. So let's say you either drive really widely or you drive really mildly or you drive in the highway or you drive in the the local or your different environment then they have different kind of profile for the engine right and different kind of profile will impact the fuel efficiency of the engine. So that is the one year I'm trying to work on. But it's not that easy. We But we do have like cool stuff. We we have uh Mccuri. McCuri is what what brand is a Mercury? Uh oh McCree. Yeah, we have Mercury vehicle. Put a lot of sensor on that drive around Detroit. Collect the data in the snow day in the sunny day. It's fun. But but finally I didn't get a good result out of it too. I mean, I they still, you know, to me that's I as I've I've had a Ford for a long time and I think that's essentially, you know, they have the the fuel uh gauge and then it also estimates how much fuel you have or miles to empty, right? Your range and stuff and all of that is coming from from stuff like that. So, it's it's fascinating to see how you know these topics that's I mean that's really one of the reasons we started the show is the these topics that we talk about or these problems that people are trying to solve. They're not nec they're unique in certain aspects but in a lot of areas they're not unique right like you're trying to improve the efficiency of something that could every industry is trying to do that in some form or fashion it's optimization optim optimization right whatever you use optimization y so we we tried that for one year but it's not not good the result is not that's kind of early times of like true machine learning but even just the hardware that you had available to even implement oh yeah the hardware they they actually take about 20 30 of the same the data from the engine and the the brake the steering right now is look like nothing right because right now we have lot of things like everything there but back to that time right like 20067 is still that's why I'm saying that don't just blame on the the big three we call the big three they tried to do something it's just like yeah this wasn't technology wasn't there yet yeah not yet they had the mindset I have but yeah yeah well yeah I mean I can't you know even just looking at personal vehicles like the vehicle I had in 2007 is very different than the vehicle I have now right and even over the last like my wife we she got a wreck so we bought her a new vehicle last I guess last year. Yeah. And like her car is like a spaceship compared to mine and mine's only five years younger you know older than hers. You know what I tried I tried Mhm. 2010 the last generation they don't even make stops still there nobody makes it's very expensive last time I try to change the what the compressor cost me like 100 I was going [laughter] to say it was expensive yeah no more parts yeah yeah more than more more expensive than car when I try to change the car they say it's 20 2000 $2,000 and then my wife say should we repair yeah [laughter] for sure no that's uh Yeah, it's crazy how how fast te tech changes. Yeah, but the more it changes, the more it stays the same, too. It's like some of the companies that I mean I I've moved on to Toyotas recently and it's like they kind of get dinged because they they stay with older technology, but it works and they almost have the Apple model where it's like we implement it really well. Our reliability is good. Like the user experience is good every single time. Yeah. They are they are still like resist to the EVM, right? Yeah. But but very progressive on the hybrid side. The I think the hybrids is a good one. Yeah. I think I still think that's like the best balance today, right? Like I wouldn't buy just an EV car as my primary vehicle because of the range anxiety issue. But an EV gets you all the benefits that I want locally where I'm commuting every single day 50 miles to and from, but then I also have the ability to drive however many hundreds of miles I want to because I can still fill it up. Yeah. Yep. And so it's but then you've got the physics of that, right? where it's like, well, you've got all these batteries and they're super heavy and you've got the efficiency issue kind of fighting each other there. But how did you how did you go from there and end up uh at Anadarko? Yeah, that's uh that's also a very interesting story. So uh BP actually like 2004 five the oil price going down, right? Mhm. And also I think at that time BP scale back from the solar and the bofuel something like that they sell all this solar and wind at that time because before they call the back to petroleum before they say beyond petroleum you guys know that story. So right now we don't comment what they're doing right now. [snorts] So so at that time they we had stuff from 2015 2016 we had different kind of run of layoff. Mhm. I saw why that's fine because I play with the spoke and I at that time you do in BP actually that's really advance at that time in BP because they have a lot of region and they need to know the global production and I'm use the VBX VB6 in the Excel to write the program to pull the data from different database and send to the spok I'm not sure you use but after I leave BP John and for up to two years. My private colleagues still talk about this one with me and try to get over from inside BP like Helio. The amount of the amount of Fortune 500 companies that at one point in time and probably still are running off of VBA and Excel were unbelievable. No, VB6 is VBA in the Excel. Yep. So at that time uh several round of the layoff and then uh I just a little bit tired but at the time actually create a a right and they try to do something and and even like inside BP at the time BP have like they call it a well advisor system similar to what I do right [clears throat] now the real time system I know what they're doing and then I want to do the same thing but I don't get the chance then when I went to that's one of reason I join and on the dark another is the belway because 99 is go through. Yeah, I live in KD. Okay. And woodland. If that is not go through, I may not join the dark. But that's one of the reason I joy was 99 was open and I was like, well, my family's in the woodlands. That's an easy hop on 99 and go all the way up there, right? Like it uh it makes a difference. Yeah, that's the that's one of the main reason. And also actually that's a that's a good decision because when I join Anadarko actually we do a lot of good stuff in three or four years. That's a shame Anadarko got bought by Ox. I mean they were doing a lot of really cool things like I mean I mean what you were working on but then you got mean Zeto broke off from there and Jose and Diego like I mean some their approach like was like hey we'll go you know kind of pseudo crowdsource this stuff. Yeah they did some of that too. Yeah. and like they did the hackathons and stuff like like innovative stuff now and that was what seven six seven years ago seven years ago and so that's uh yeah it is you know they got bought because they were doing those things but then of course as with most things whenever smaller innovative companies get bought they end up being forced into the the old way of doing things or they just say we don't want that I mean like right yeah or they just drop it off just you know uh had data engineer on my team from nsign and like they did some really cool Andrew Munoz was there. They did some really cool stuff. And they got by by Marathon and as far as I understand, they were like, "Oh, we don't want we don't want that stuff. We just want your acorage, you know?" It's like, but they did some really cool stuff, man. Like, like, don't let that die on the vine. Like, yeah. But, uh, but no, I mean, back to you. Sorry. I mean, like, so at Anadarko, I mean, like, you know, some of the cool stuff you did there. I mean, yeah, you've got I mean, you've got SP papers. You've got all kinds of really, really fascinating stuff. Yeah. At that time, we we do a lot of good stuff and then put into production. That's more power because like I told you when your ML model run 24/7 that's the kind of feeling is good at that time right now is normal even you use a Gmail they had model in the back but that time in especially in oil like in the dueling that's very old discipline you can have ML model running 24/7 yeah but classify the data that's that kind of feeling is good let's talk about that because that was most of your focus well it was mostly the operations side Right. So drilling and completions. So I know you have papers on drilling models and some on the frack side as well, right? Yep. Talk about kind of your experience with that coming in. I'm very curious personally because I know lots of pretty much every oil and gas company is in this spot of like all right we've got a bunch of data. We know that we want to try and optimize these X Y and Z processes because these are the most timeconuming, biggest cost etc. most highest risk and but a lot of them, you know, they they skip the foundational part of you have to have good data that's structured and and tagged and quality and gone through all the ETL and and everything before you can even start to train a model. So, I'm very curious like your experience coming into, you know, what you would call a traditional oil company, even though Anadarko isn't what I would call a traditional oil company, but how how was that coming in? like was the did you have that kind of foundational data problem out of the gate or were they a little bit more ahead of the curve in that? So you Yeah, but to be honest when we are in the data foundation is not that good actually. Yeah. So like like lot of the context data like the metadata like the well BHA June string that kind of data we do have like I think in the code we use the will or open world I can remember we do have the database there but nobody put the information there so that's one of the problem uh and another side is the real time side actually the we had you know that Julian side we always have VM although it has 30 30 second latency always but we had that But the the the data foundation for another is not that good. It's just like we try to do something back to that time. Either you buy or or build them, right? So at that 2017 there not so many stuff in the market. One is a mobilized you know that they are actually the best at that time. Ka actually at that time is not that good. They come from 2014 and then almost die 2017 and then Ryan do something I don't not sure but right now yeah it was interesting because we were in the kind of dabbling in that space of the completion streaming when RDS so we were pseudo competitors I would never you know definitely not on the software side but on the streaming side but we talked to mobilize and talked to corva and even well well data labs later is in the in the completion side not in the during side yet but so at that time you can see that is not very good option Right. So in Anadago actually we have enough rig to support like let's say you if you have fire rig you can support right and so that's why we decide to do it internally and uh and also you can see that the the benefit of the internally build is like you can customize you want something you just do it [laughter] right you don't wait for ka to do it for you because that would take a long time. Yeah, because they need four other people to say they want it for them to do it, you know, understandably like. Yeah. So, so at that time that's why and also at that time typical typo actually come to us. They want to break the market in the bright ice of the market in so they need a they need a panel to work together. M so here I need to appreciate my uh my collaborator from typico Steve actually he teach me a lot of stuff in the stream side stream processing because the typical stream base right now I call stream actually use in the trading system right you know the trading system is like second latency they use that one is very good so they borrow that one come to gas work with me and then we build that from the stress with that one so they actually I learn lot of that principle from Sure Yeah. And so that's why later when I move to Kafka actually I I just changed the tool but the principle is still there. So that's why we come to end them for on gas you can see in the Julian side to be honest the data is real time data or like context data and also the model is there not much too much the new thing you can do but the most important thing is the automation right because right now like let's say the content model or like the hydrogen model if you don't make automate the engineer need to spend two hour to copy paste all the input from the open to the to the to the well plane we call the well plane and kick the button launch and nobody will do that right if you don't have special need but if you automate then the people will use it every single day and it's bring value right it's depend so that's why we say we need to make a lot of stuff automate we we build some ML model because like the old one like the lubase one is not good enough but for other model we just borrow we just lease like I always juice again right now part of HMP. to release the model from them. Just pay like 20 30 40 I cannot remember $1,000 each year. Just you you just plug the model in plug the model in we just send the data we get result back. So I don't need to reinvent the wheel, right? We just do make sure that UI is good. So but I think that is the the beauty right now still ongoing. Yeah. But no, that's a good I think that's great advice because it's I think everyone regardless of what side of it you're on, everyone runs into that build or buy question at some point, right? And this concept of being able to build where you need to, but also license and buy where you need to. Like it's not either or. There's going to be times where both make sense, right? Build these parts of it that don't exist that you can't buy because it doesn't exist. And it's totally dependent on the operator, too. Absolutely. We're smaller than you were at Anard Darkco. We're probably going to lean more by on those things. But I mean, we like to do some stuff in house and there's even stuff we're talking about that two years ago we wouldn't have built, but now we might build now. But I mean resources is a big part of it. But then you know Exxon can throw 20 data scientists at a problem. I mean like it's just you know it's a different Yeah. That that actually like after I lived and actually I joined the company in Tulsa. So when the my manager call me I was in Alas I think in I was in Vancouver in the conference call me invite me to join the company the first thing I asked him is how many rig you have yeah [laughter]

right [gasps] you need to have at least five rig so you can support otherwise you hire me it's meaningless right so so that's the but it's come out good we build something in the completion sign after 2020 that that. Yeah. So, so it it's a good one. So, let's talk about uh the model side of things because that's that's that's obviously super buzzworthy and and hot topic these days. But let's talk about, you know, because especially on the ops side, right? because your f a lot of your background and focus has been on drilling in completions with you optimizing those things with AI and ML and those things traditionally are not the best from a data perspect it's it's not that there isn't data it's that there is data but it's unstructured it's in a bunch of different places the context exists but it's somewhere else from the actual data if it's complete right and there's all these nuances and stuff so let's talk about like the the streaming side of things as far as the data uggestion piece goes because that is the source of any kind of ML model that you're going to have and then let's get into kind of some of the different models and techniques that you y'all used and played with and how you figured out the best ones and and things like that. Yeah, I'd say let's chop it up there, but let's focus on like I guess that stream processing piece and see how how it's evolved because like you said you were stream based with Tipco. So let's talk about you know a little bit more about that just kind of what it is and then relative say that you sounds like you moved on to CFKA and if you've done anything else let's talk about those one two or three tools and then like how they compare contrast because you know as I understand one's more it's more of a you know drag and drop guey and the other one's much more you know kind of code based so let's kind of jump into that that's a good one let's talk about that one well and like the pros and cons from someone you know if whoever is watching is a data scientist at an operator and they're trying to you figure out which tool they're going to go with. Yeah. Hearing your direct feedback of like, well, this one is good for this because X Y and Z, but it's not that great for this because it, you know, is batched instead of streamed or whatever. [snorts] So, yeah. So, uh, typical stream base they right now they call stream. I haven't, uh, pay attention to them for a while. I'm not sure how they evolve right now. It's actually is like a Lego. The each of the is a graphic. Yeah, it's like the no red. I'm not sure you guys play with the note is like a block a block a block. Each block is do a different kind of stuff. It's visual. It's almost like SSIS or like a Yeah, you just need to put the block this the data is flow from the lab to the Y. You just plug in all the stuff inside and go through it. Yeah, they use this one in the financial market for a long time and is very good but it's very expensive. Yeah, [laughter] typo is always we looked at that when I was at RDS. Yeah. And then we're like no that's not going to work for [laughter] us. The reason why they kill another one is they cannot pay the latency. Yeah. Because when another tip call for the first three years it's almost free. Yeah. Because a partner. Yeah. It's a panel. I use them for free. But when I left is not free. So that they cannot afford it. So that's one of them. But another one is because the I'm not sure they change anything right now. That one is not disputable. It's not like the the cloud is like because there's a driver behind the scene the typical stream base. So the guy told to me they say we try to make the distubable dist processing okay in the cloud. So you can uh scale scale but they say they cannot do that. I'm not sure they can do it right now but they are very useful and very easy to pick up and they have a lot of the block you can just plug in your model. So you can see that data coming you put a box here what kind of equation here it's like a water and right water flow here you put a red and it's coming here. So that actually that's the the lot that's actually the concept but the capka is different capka is like a palm it's like a big palm or lake you put all the data there you have small app to get the data out and processing and put back in a real time. So these two is the two different concept of the handling sharding and like and then writing the distributing out it can kind of fan out to different sources. So the stream base is left to right. It's a sequence processing sequence processing. But you can have different kind of light to do that. But the kafka is a different one. Kafka is like a big lake all the water there data there. You just get out process come back and next one just get back based on the topic. So that actually is the the big different. Yeah. But also I mean Kafka is much more code based. I mean like originally was Java and then they built like KSQL and they built you know Confluent obviously has like the CFKA connect where it makes it a little easier. You can define things on either side and it'll kind of run those but there's a lot more overhead and even just you know even just like deploying CFKA like you said you know streambased maybe doesn't distributely or horizontally scale and CFKA does but it takes a lot more know how to how to deploy CFKA. Yeah, but the Kafka right now you don't need to manage your own one. Like when I was in the private company, we actually we work together, right? We try to put the Kafka in the field. Yeah. And that's cool, right? But it's a little heavy like we talk about is before is not designed for the field because designed for the data center. It's not for the the edge. So, but I'm not sure right now how it's doing. Yeah. But like you say, yeah, the deployment is a little bit. But I right now it's all cloud. you just use a conference you never have any issue they give you the connect stream like URL plus the port that's right you just talk to it get back but stream base you need to install that in your local machine or in your VM and or like kubernet I don't know and then you run is there and you connect to it and do all the stuff and deploy it it's a little different that yeah stream basics is the old days of the software development and the k actually is more than life that's why I can say yeah and they've got the manage version like I mean I I think it's worth talking about too and I'm not trying to be a typical salesman but is that streambased connects right to spotfire too. So then you have end users you know your drilling engineers or completion engineers that are familiar with doing stuff in spotfire that can be a source into spotfire and they can do that's one point I want to mention back to 2017 when we work with the tip code they actually at that time the sales ladies Kim and I cannot remember her full name right now she is with Microsoft right now at that time we give them the suggestions like you need to hook up the stream base with the tip uh the sparky and let the sparky as the leot visualization. Yeah, like the tick tick tick. You can see the clock. But I don't think right now they may happen yet. No, I think they've got it now. But at that time I give them we give them the suggestion say we want that at that you have lot the market because like guy always use the spark file right is and if you can get the real time data like a grandfather right it's as useful as a grandfather then everybody will use it. So yeah. So that actually is the if they can integrate like that way. [clears throat] I think they got it recently. I mean like I know previous group you were working with I think they had the connection and that it was connected to the stream base. I know it it is a source to spotfire now like but I'm not sure that Spotify can the real time actually show that like every single second you see that I mean we saw someone do it but again it wasn't native. It was because you know the text area in spotfire you can use as a web page. So they they installed Yeah. So they installed in you know I don't know if they're using like um I use a web player like plotly or something like that play a lot. Yeah. Well but I mean the text area you can like you could install like plotly or pull the CDN. Oh and then you can like visualize it but you're using JavaScript behind the scenes to connect to that not truly the native like it's a work. Oh it's a workar around. Okay. Yeah. But but even right now I use Spotify [clears throat] I like because it can view million of low. It's such a great visualization tool if you have large data sets that you want to visualize like I tell you right now we had a real time model 247. So what we do in the morning we will view the past 24hour data past 24 hours how many data point you see is each hour is 3600 times 24 36400 and times the 25 ring is a bunch of data [laughter] and then we view that and try to see what's the model look like is a long prediction then we collect that one so that's a spot file we use a spot every day spy every day yeah know there's that's one of the the consistent points I feel like across the show is that if you can if you can make it easier to use whatever it may be in the place that the end user is comfortable whether that be Excel, Google Sheets or Spotfire or whatever. Yeah. Whatever tool they're comfortable with is normally the best way you're you're going to get buy in. Yeah. Get by. That's the way I like recently maybe out of topic for this question but right now we had a stealing team. Mhm. They want to you they use a star steering. You guys don't know. Yeah. We had Eigor on. Yeah. So they need rely on the rig to send them the data and plug the plug because they need to send out the pro the the report every three days three times a day right but they rely on the rig to send them the slide data or other data right now because my system we just push the data to the to the star studio and they just hit the button and they get all the data before they never know that that's e and also our quality of the data is good because He had the ML slide model which is running very accurate like 95 or 96 97 something like that. When they get the data from the rig they say okay 2:00 to 205 is slide. Right. Right. Right. Right. 206 to 210 is not slide. But actually it's not. It's like maybe 203 to 205 is a slide. Yeah. It's the difference in a human writing it down in a spreadsheet which is probably where it's coming from because that's how all the field stuff happens, right? And we get to every single level. So single level uh single second level. So this second is a slight or not we classify. So that's it. Okay. Let's go back to the sparky not the spy stream base and the copa. Yeah. The another thing is like I think it's like when when I see what happened to the system we build I feel that we need to get that kind of technology which is open source but you still had managed service right you had support support that's why Kafka come into the spot greatly because like you never know next time who coming to get the current company And the op similar to like the cloud is the same thing right now a lot of people like to write the application into one cloud. Yeah. Like either AWS but then you're locked in the login that's what happen you can move it around. Yeah you we use the Google cloud at that time but when oxycom is a so if you do anything use GCP you need to redo the whole thing in the Azure. So that is crazy. So that's why right now I try to avoid this kind of use the service in the cloud a single but even like Kafka as containeriz even different cloud service they had a cafka manage service like a has like Google has and also comp has so that's another we call a v window lock we don't want to be locked by the window that's yeah typico you're tied to them because like because that doesn't go anywhere you know, like at least maybe like a Python script I wrote in a Google function, I can maybe tweak it a little bit relatively quickly, but like if it's streambased, you just got to refactor everything, you [snorts] know, and that's a it's pendable. It's pendable. So, well, that's what they're banking on though, right? Like that's that's a lot of their business model. Yeah, that's sticky because it's working and like that's why it's going to be a million dollars to cut off it anyway. So, that's three four years of OS, right? I know a lot of people had an issue with that one. Yeah, the Aviva try to face out the pie. I move all the customer to Aviva. Yeah. Yeah. But but people don't want to use Aviva because it's more expensive. But you you rely on the Pi VB6 and everyone uses Pi too. Like it's it's very similar to Spotify, right? Where it's you've got a license pretty much everybody's. The problem is the Spotify keep update with the technology. Yeah, that's true. Yeah. I don't want to know. I don't want to say name. No, I mean all the users of it know. They know exactly what you're talking about. Yeah, it's crazy. They don't keep update. It's still VB6 in the back end and never invest their dining on that one and they just try to push all the pie customer to the one. Yeah. Yeah. Now that's Yeah. Yeah. No, I mean this is exactly why we have this show because a lot of people you don't know what you don't know until you go through that. You feel the pain, right? It's going to be a million dollars for us to migrate this or we just live with it, right? And so no, that's I mean that's definitely if you all haven't if whoever's out there that's in the data ingestion side, if you haven't looked at at CFKA and Confluent, it's a very very useful tool for that real time. Uh and we use it every single day, right? Uber bang is the cafka. When you place the order for a ride, that's the cafka. Send the message to some place and do well and that's the other thing too is if you're as you're evaluating those solutions, look at who else is using them because Uber has a much more difficult use case than we do in oil and gas no matter how you and Uber has, you know, tons of Silicon Valley software engineers that are focus on that and they're writing open source software and like that's why I mean things come out of those companies that other people use because they build built in-house solutions because they have the team to do it. Mhm. Um, all right. So, let's move downstream a little bit from there. So, we're we're it's into CFKA, let's say. Now, you you've moved on to CFKA. Where where are you landing the data? Uh, does it go into a database? Does it go into a data lake? Go to Snowflake. Snowflake. Snowflake. Okay. Yeah, we we have a Google Enterprise Snorre all my data is stopped in the Snowflake. Whoever want to use it, they just go there and get Okay. Yeah, that's that's what we use as well. So, I mean, yeah. And we use the mango before. We may still need to use a mango one day because mango is a is a nonSQL. So you don't need to defy the table. Yeah. But but you can drop your JSON blob into a staging of uh snowflake and then they have the the parsing that you because you can do JSON parsing in uh in snowflake. Yeah. Because right now that's back to the VML is more detail like VML you have 400 or 500 column right but we only need 50 of them for the [clears throat] remaining 450. Some of them have data, some doesn't have the data. That we call the dynamical column. Right now we try to get that one. If they had the data, we still want to get and see whether it's useful in the future. But we cannot build [laughter] right table for that because each read the name is different something like that. So that's why maybe one day we need to use the mango or like like you say the the block and passing the frame to that one. Yeah. But uh yeah. Okay. Okay. So, it's in there and then so now let's and let's talk about what you've done in the past and what you're doing now. So, we're doing you know training real time models, right? Or so do you all train the model and you just let it run and are you just passing in and getting the outputs or are you also training it in real time and is it is it you know once it gets to a certain confidence level are you deploying a new model or are you just is it more kind of a manual process for you? Yeah, that would be an interesting topic to talk. Uh so right now we right now for the tuning side we only have one ML model running 24/7 in production for almost two years that's ML slide right the slide model uh we build that in anago and I rebuild the whole thing but the different architectural we use a convolution neural network right now we use rocket that's another uh model but my my goal is like I always try to see what is available in the open source community and try that one. Yeah, like let's say meta had a new model coming out maybe can be used by my use case. I will give a try whether it's good or not and continue. So that is one side we try to see whether we can use a different kind of new model to improve. Another side is like like you say uh we don't train the model in real time. Okay. So the model will train offline and put in real time inference but we retain the model whenever we feel that is needed to for example let's say there's a a unique data set yeah let's say right now we have train tomorrow at two rig the two rig we never see the model never see the data for the two rig before and also maybe the location of the well is never had the training data before so we need to monitor that two week and see how good are they lot time we use our eyes sorry we don't have a good stuff because they don't have one tool right when the real time data coming you never know is a slide rotate until you use your eye to see or you use a model to classify sure so that we don't have um automatic matrix we use the eye like me or my one of my main we use a spark line yeah [laughter] and then that's we have a funny story about when Bobby and I were working together we were he was messing around with with building some uh some frack models to identify the different like classify the different parts of the frack stage like build up you know breakdown year [clears throat] that's very useful right well because we were trying to get cuz like it was either that or we were going to have to have like our account manager or someone looking at it and brushing and marking you know so we were trying to like classify that so that because a lot of people then wanted to be able to you know align you know normalize wells to breakdown or normalize them to frack start or to ISIP but you have to have that classification All the metrics that that you want typically on an on at least on the frack side are very easy to back out when you know the start and stop, right? But the start and stop or which part of the start and stop you're in, you know, are you in the pad or the ball seating or breakdown or wherever you're at uh is important. And so that was we were having these conversations and I was like I asked him I was like it's almost like we need a computer vision model because like visualizing that data is so much easier to say for human to say this is where this happened, this is where this happened and this is where it started and stopped and all of that stuff. And so like Bobby just built a little uh I don't build an R like plotly and I just I would just like it was just you would brush it would pull up the data set and you would brush it and then you tag it and it would save. I was like, that's a brilliant way of doing that. That's labeling, right? That's labeling because labeling. Yeah. That's what we do. But but you have to do that up front. I mean, someone's got to train it. I mean, so whatever was was trained off of my eye, which I generally knew, but but then then you bring in bias, too, right? You like um because you probably have a handful of people doing it that way. You get a few different experts opinions on it. But man, people don't think like people think that it's like, oh, I've got this data set and then I can throw an algorithm at it and then it spits out all good [ __ ] right? Yeah, it's like no, you manually have to go through and tag thousands of stages, right? Or if you're good like shout out to [clears throat] well data labs, right? Like they built their interface so that the user is tagging these things and then it's getting stored and then you get enough users doing that. Now you have federated learning feeding your model constantly, right? Doing it. I think more companies in the energy space on the software side specifically need to understand like the processes that their users are doing and how and if there's ways to use that to build and train models because that's happening literally every day. So instead of you feeling this overwhelming like how do we where do we start? we've got 10 years of data and thousands of data sets. Like no, just start now. Small pieces small. Let the users like as much that you can do where it's being crowdsourced from the community and the users that are using your stuff every day, the better because they're the ones who are going to they're themes on it to begin with, right? Like well and you talked about like, you know, hunting for you and again technology changes so much because now like neural networks are like old news almost, right? But like at the time, but I mean I end up searching around trying to find a good and I found a good neural network for classification that some PhDs had written. So I was like let's try it out and you know and then had to work on the kind of overlaps and finding the right distances but like they were doing basically what I was doing for a different application. Yeah. More I think medical or whatever it was but then it was like well they're just looking at you know different streams and like training it and like it was a recurring neural network and I mean it worked pretty dang well. I wouldn't I didn't say long enough to see it through, but well, so what's your what's some advice both, you know, things that you had to learn the hard way and things that you uh found out the easy way, but what's kind of some advice that you have for people that, you know, are data scientists in these different companies and stuff trying to, you know, start using [clears throat] ML? like what's what's what are some of your big you know pain points that you ran into that you know now you wouldn't run into but you only know that because you experienced it firsthand right like yeah my first uh uh advice is always the use case on the use case don't care about the fancy model so like right now a lot of people think that yeah the the the model is more important but actually not like right now they that they had always had two uh opinion. Why is the model central or like data central? That's right now model that means you always try to build a sort that means the state of the art model is always good use the same data set and build a good model but the other side is like they always try to get the best data and use because a lot of model is not that huge different right so so that's what I want to tell the young guy if they want coming in say whenever you try to do something in visit mail you need to focus on your use case. Yeah. Use case, use case, use case. Like when you buy a house, it's always location, location, location. Right? So here is a use case. When you get a good use case, either you use a very dumb way to happen or use a very fancy way to make happen. Doesn't matter. Either either like when I was in uh BP, I use the VB6. No, the VBA and Excel to make working. But I made working like solve someone's problem. You solve their problem and I get I get what we call so why in several wrong of the layoff because the stop right although I finally I leave but but it's not like technology sometimes is good but you cannot just use the had a hammer and try to find a nail. Yeah, you get the nail and then you use different kind of use the right tool for the job right tool for the job. Even it's a very dumb one. Even you are PhD don't make doesn't mean that you don't play with the Spotify or like the VBA, right? Yeah. So that that's my suggestion and and that is very important. Yeah. I mean I was talking to Colin out there about it, you know, like when a user uses a website, they couldn't care less whether it's Ruby or Python or Java behind the scenes or using React or Angular. when I click the button, does it work? Does it just spin fast spin up fast? Like, so I mean, use whatever tool that works and get gets it done. But the end users, you know, we we live in this technology world where we know all these different technologies and think, oh, we got to use it. 90% of the people out in the world don't care. They couldn't care less, you know, what you're using. They just want it to work and to solve their problem. Yeah. Well, and I think the other the other thing that that people run into when they're chasing new technology is that like uh we were having this conversation the other day, but like vector databases, right? Uh there's a group here that I've been to a couple of their meetups. It's a Houston AI group or whatever, but a lot of the conversations are like, well, there's here's a list of a dozen vector databases, right? Like [snorts] which one should I use? who's going to and like when you're that early, when you're almost like in that bleeding edge space, you don't know who the who the Google's and Microsofts are going to be. And so if you choose one and you go build your entire stack off of it and it doesn't survive, then you're screwed anyway. So understand the new things that are happening and what's happening there, but don't immediately run to them as the first thing you're going to use. Because guess what? There might be something that has existed for 20 years that has full support, all the documentation in the world and is super easy to use, but it's not the sexy name. Sexy name. Sexy name. Exactly. A lot of times it's just the sexy name. Right. Like the day I reading a news MIT professor just changed the name of his class from linear algebra to AI foundation. And the class [laughter] is just like got huge. your job. So, [clears throat] so the name is just like well I guess I guess I took a AI foundations back [laughter] in 2008. I guess it looked good in the transcript, right? Colin has the same [clears throat] story. He said uh there was I don't remember what company it was but they were at a you know doing a talk at a university a presentation for the students and it was something about you know using it it wasn't it was very technical the the original time that they presented it and then the next one was like leveraging AI for this and it was a like the first time they did it there were like two students because it was talking about mathematical application of etc you know as we do in the technical community we're very technical with the names normally but then the second time around the marketing department had gotten a hold of it and they molded it into this you know whatever with AI and the CL it was full right like and it's it branding is a huge in a very real marketing and branding is a very real thing but when it comes to dev and stuff like just because it's old doesn't mean you shouldn't use it or that it's bad or that it's not like there's a lot of value that it brings and it's funny because when I was younger I feel like I would be slapping myself for my current self for saying that because I love the bleeding edge stuff But it's when you when it comes to application and real world stuff, it's a lot better to have. Well, you need things that are hardened. I mean, there's a reason Java is still running a lot of things. It's I mean, it's tried and true. It works. C C I mean like Python's fun. JavaScript is a can be a mess. You know, it's like you get all the you let someone else learn the lessons, right? So like the scalab, you have scalability, you have security, you have all of these things that you don't know about when you're testing models on your laptop versus when you deploy it full scale across an organization. Well, and there are pitfalls with open source sometimes too. I mean like wasn't there I mean it was probably five six years ago they came they found a problem in numpy that and a lot of people had built their statistical models for their dissertations and this that you know or published papers based off of that and they had to walk some of it back because the numpy you know model wasn't actually calculating properly. So I mean there there is you know pluses and minus you got to know the tools or you know some things that you want to believe. I mean, I guess it comes down to the whole licensing versus, you know, having something, you know, do I use SAS or do I use R, you know, I mean, I I don't want to use SAS ever. But, you know, the liability is on SAS if it goes wrong. Like, but, you know, if I take an R model that some wrote and I don't validate that it's actually doing what I think it's doing, I know it's backed by IBM, right? Uh, no, I don't believe so. I mean, deposit. They rebranded. There's R Studio, but Oh, they rebrand it. Well, our studio rebranded Deposit so that they're more like because I think they do some Python now and they support Julia and all that. Oh, I see. Um, but it's still like the IDE the open source one's called uh R Studio still but um but um yeah I mean I think there still R is still open source. It's not you know and actually Microsoft it's funny talk about spotfire you know Microsoft did that whole Microsoft revolution analytics and they had their own repository you know there's cran is the main repository and then they did like mr an or whatever so a couple months ago my one of my R data functions stopped working because like for some of the users because they had got the newest version like so typco was using Microsoft's repository to like load it so it was loading the newest version of like dlier in whereas everything I had written was based off of like version 0.9 or whatever that typical had pre so they had to write a hot fix to repoint it to cran because Microsoft shut down that repository once they discontinued supporting it and then like when tear pulled it for all the new users they were getting the most recent version of because I guess it would default to the the major cran so it was a interesting uh benefits of containers right yeah that's no I mean I run into that all the time that's half the reason I I In my opinion, I don't code more is because I'll go try and, you know, install a new package or whatever and it's like, oh, well, this package is out of date or this one is too new and it isn't supported by our current like you're like, yo, that's another one I want to mention is like right now we have so many packages. Yeah. And sometimes old sometimes new like the Py Python, it's better than Python. It's like if you don't specify the exact version, you always run into issue. Yeah. But Python at least I mean in my experience between Python and R like I mean I used to hate Python because I couldn't ever get started and I had and I had this and like installed five different versions I didn't know how to make it work but now that I understand virtual environments and but that you can define the requirements.ext file and you can say I want your request you know equals equals the version whatever and you can lock it in for that environment. I mean that you much better environmental control whereas I think R is trying to get there with some re envy but it's not nearly as good but like R I mean I can install R and R studio and get someone up and running like that you know so there's pluses and minuses there and and I did want to ask I mean because and maybe you probably answer is probably both but I mean what have you used like is it mainly has it been R has it been Python or are you using that now or we are mainly okay let me put it this way my the whole ecosystem in the tooling side we have different kind of languages It's because in the uh company like 50 years old you had different kind of system. Sure. You need like we had a C# to get the data from some place and we had a no to get the data from other place and we had a Java as the the main for the capka because Java is capa capka stream or like capka consumer and producer is a capka is a Java and also the model or our model is in Python. Okay. We write either the ML model uh ML model actually is the file right you just embed the file and then you build a rapper to get inference that kind of stuff and also like engine engineering model like the to model is equation bundle equation you just write in the Python sure and pack as the API right now we pack everything as API and and run as app so everything is containeriz even all our other app like the influence app or like the sync app what is the sing app that means like we get the [clears throat] data from the model we need to write back to cop cara or we write back to snowflake that requesting app so we have more than 40 50 app right now include the model so right that's the people come to me and say ding how do you do mop I say I don't really do m because my whole ecosystem is follow the software development cycle my all my app is software is not just a model model is just a small piece like especially model you have to write a flask scap or fast api I need to manage them in the same way. Otherwise, I just specify something for the ML later. It's just harder for me to manage it. Similar to like you had a free like American Aine, you have half the aircraft [snorts] is Boing, half the aircraft is Airbus. It's harder for you to maintain that kind of stuff, right? So, yeah. So, that right now we have different kind of language. Whatever old system we need to talk to use what kind of language, we will use that language. So, Oh, of course. Yeah. Yeah. Yeah, I mean again you've got all different kinds and again probably C# and Java have probably been the most traditional like in the past as far as like software engineering and data like integration type stuff but I mean yeah I was thinking more on the model side so it sounds like more Python than yeah model like I said is more package you don't need to write like if you want to write a model it's a Java you need to write from scratch for some of the basical calculation but in in the Python you just get the package in long asm right but that's another story I want to share is like 20 24 years ago when I first played with the ML you know the back propagation algorithm yep I learn in the C by myself and and that was C not not even C++ no it's C it's not even object oriented and you can see that how I sometimes I feel that I appreciate so many package oh yeah but sometimes too many package had wasn't issue but when you come back look back 20 years ago you need to [laughter] write the the equation yeah I get the memory and like I get the paper all the equation I need to convert this equation into compile I compile them and then learn that and also at the time we have like the parameter is like several thousand I learn to change the neural network is about eight hours something like that and also we had a support machine at that time we learn the the the sample we has 200 sample but look at here right now like what the open the check GBT how many problem oh god hundreds of thousands to billions now. So you can see that how the hardware actually the hardware enable a lot of stuff. Yeah. Hardware and memory. I mean like memory hardware because I mean I talked to people I mean like you brought snowflake. Snowflake is $23 a month per compressed terabyte. I mean I remember talking to like Jeremy Funk and some other and a couple other people I know that were doing software and hardware sales like 20 years ago. It was like a hundreds of thousands almost millions of dollars for terabytes of storage. Like but like I told you when I was in college I assembled in the computer for my classmate at that time I have stand up configuration 120 gig 120 120 megabit memory. Yeah 30 gabit hard drive. Yeah I think the CPU is 800 or 1g [laughter] not multi-threaded. Yeah. Wasn't Was it Bill Gates? Didn't isn't he famous for saying something like you'll never need more than 256 megabytes or something like that. It was literally it's it was shout out to my uh high school uh computer science teacher. She had that quote on her wall in our computer science lab and it's I think it was Gates but it was like yeah a user will never need more than 256 kilobytes of RAM. But at the time we run you've obviously never seen Slack. Yeah. [laughter] But at the time we run Windows very smoothly. Right. Yeah. Right. Window A, Windows A or Window Net NT something like that or Windows 2000 it's running smoothly but right it's I look at my memory right now is like my desktop is 32 GB oh yeah you see well I mean but think about the but this is more powerful [laughter] than any computer I had think about inmemory processing now I mean that's a huge deal like if you can get data in memory like you can process it much faster than like you know off the hard drive now and everything right and now you also have solid state versus, you know, spinning discs, which is a better deal. But like if I can throw it into RAM, I can process it really quick, you know. Um, but yeah, no, it's No, I mean like definitely like the hardware improvements have enabled so much and and again what was big data when I even got into the industry and big data is not even like a thing anymore because it's not it's not big. I mean like no seismic would be the only one that might qualify for that. But not even for us. I mean like I just don't think big data is a thing anymore. People don't even say big data anymore. That's very true. There's like it because it should there's always data. There's basically infinite compute. Yeah. Knock on wood. I mean like Yeah. Right now when you look at Nvidia's new GPU is like longer than truck, right? I think like is it that big? Oh, it's big. It's like to here to here is the that high. It's like I cannot remember the model name. Uh but it's very from here to here because like for the GPU the most important is the GPU communication to GPU. So like you send data from this to that one. So you need a bandwidth very big. So that's why that's the key and also another is like Nvidia already had a CUDA ecosystem like the Windows we have right now. So if you want to break into that one and get another empire that's difficult although the technology is not that but people get used to load the library like you say the library like the package is already there. you can spin up in like one hour, but if you get a new one for AMD or something like that, you take a lot of time to see. Speaking of CUDA, you shared it with me and I had seen it through another one, but did you see like the news this past week where like you just feed like there's the package, it's a CUDA df package in Python and it'll speed up your uh pandas by like 150 times. Whoa. I I Yeah, we'll share with you like it's and it's a simple like you can keep your pandas code. just define it at the top of your Python script and it will, you know, accelerate up to 150 times your pandas code. Yeah. Using G if you have access to the GPU like through I think we're going to see a lot of really interesting [ __ ] around like the the the the hardware acceleration is what it's called, right? But this idea of hardware continuing to accelerate software as well with with all the ML stuff because it's all most of it at least seems to be GPU based moving forward and so those are very expensive and they're they're not cheap they're not easy to uh maintain and stuff and so that's uh it's going to be interesting to see how like the we have worked together on that one like two years ago or three years ago when we don't have a starting what we what we use in the field in the fragment is FTM right FT or like satellite. Yeah. Right now the staring company 200 or 300 up and down it's no issue right now. Nobody complain about that because that day when I talk to neighbor and ka they work with the they collaborate. I asked them how they collaborate. They say that call will calculate the the low map for the new well to J and send to the rig uh so that the check point right so that that we can follow and do that but imagine without the styling how can you do that because like everything the coa stuff is in the cloud and you send that because the start I mean the satellite stuff we were using five six years ago was slow super expensive I mean like that's Joe and me work together at that time it's a high sale we try to make more stable table and then use the LTE like the poor I think Paul the founder of the high sale at the time try to put a model there so in the high so but right now is not issue anymore you can see how the technology yeah masco I don't have too much feeling on him but sometimes he did a good job so starting is solve all the oil problem right now no I mean that that has historically been one of the biggest issues is how do you get the data from the field to a bigger computer processor to the cloud So it can do whatever it needs and then send it back and yeah. No, I I think that's that's a huge development moving forward is the Yeah. So right now nobody complain about the internet right now. No, I used to it's it's crazy how that's changed. You ads before even before that when I was in the field, right? like everyone, the engineers and stuff and supervisors would have hotspots, but those would be capped at, you know, a gig or whatever it was back then because you paid by the megabyte. And then you'd get, you know, the guys in the field would be in their van watching porn on the cellular signal, which is the same one that you're supposed to be pushing reports and stuff. We weren't even streaming anything. We were just pushing reports and like files, but even those would get bogged down because they'd use all the bandwidth watching TV or streaming, trying to stream stuff. remember the guy just call uh co say ah I just watching a movie I [laughter] like you can't do that man it's funny I mean kind of brought up kind of clicked for me a little bit there are some parallels even say you know say we we knew shale you know there oil existed in the shale for but it wasn't economically viable or technologically viable you know and then you see some of these parallels like you know talking to you about what you were doing 20 years ago or you talked to Talal and he was doing computer vision vision for his, you know, master's thesis back 20 25 years ago, but like again, you just didn't have the hardware and stuff to really do it or productionize it. But like all of a sudden once it catches up, you're like these are not new concepts. It's just we actually we can actually do them or apply them at scale and affordably. And yeah. Yeah. It's pretty interesting. It is. It's very true. It's uh it's new and sexy to the average person, but for the OGs, they like I've been working on this for 20 30 years now, you know? It's uh but yeah the the hardware definitely enables the idea like to your point those ideas aren't new new at all. Yeah, the idea of ML isn't new at all, right? Goes back decades. But the actual you know commercial implementation of them gets enabled by the technologies that yeah like two years ago you may hear about me about all as computing but this two years I don't say that why is doesn't need because we had starting why sometimes we still need some of the model but for most model you just put in the car and just push because stalling is really stable and very fast. Yeah. Yeah. That's I mean talk about snowflake. I mean, if I want something to run faster, I click a button and within a second and now I've got twice as compute, right? Yeah. Like it run and it runs twice as fast. Like it's just Yeah. What uh I want to ask you this because we gota be wrapping up here, but what how do you see um automation moving forward with because to me, you know, the first step once we now we've got we're focused on data, we're focusing on ML. Mhm. The the one of the big kind of visionary things with ML is automating processes, things that happen in the field, whether that be, you know, just on a regular rod pump well, shutting off the motor or closing a valve or whatever, all the way to, you know, the dream of a completely automated drilling rigs and fracked sites and all of these things where, you know, there's all this risk associated with humans being out there. How can we eliminate as many humans as possible in the human era element? How far away do you do you think we are from some of that on the on the ops side at least? Um, you know, because like to me when I got into the edge side, that was like the the immediate click for me was, oh, if we have this computer here, then that gives us that enables us the ability to potentially automate these processes because you take the internet connection liability out of the equation, right? Like if that goes down, you still have the ability for the entire process system to operate because the computer is local. It's not and the model is local. It's not up in the cloud. But just in the last two years with Starlink and things like that, like how do you I mean has that changed your perspective on the the automation side of things at all as far as like we might actually be able to run some of these things in the cloud or do you think there's always going to have to be that level of redundancy of having local compute? From my perspect we we always need to have a local redundancy. Yeah. The another one is like from the technology perspective I don't see any issue. Yeah. Because the automation I I I don't say that you auto jud well like just like blind Jill the well that's the whole story but for other like pie handling like that kind of risk that kind of stuff I think for a technology perspect we can do it it's just depend on several factor one is like the culture we have a lot of people working in the field like 30 years they are not going to change the mind right like the second one is like how much money we're going to put into it yeah Mhm. That's like my VP when I was in the private company in Tulsa, my VP of the the during completion, he say right now the the vehicle can auto drive in the in the street. Why cannot have auto during week? I say yeah I I cannot say that the technology is there already but with the investor we put in the money into the during competition or oil gas and we never go to imagine when the car auto driving in the in the street how many money we put each year globally right billion of dollar a lot of stuff doing lot of people doing they custom develop the whole edge GPU the test but for guys we everybody think we the sack industry who going to put the money and if you no money how I don't say the technology cannot be breakthrough but you need to have the investment I don't think it will going to be happen one except that we see the jeweling technology can to use to drill another kind of mind like like the different kind of something like I think people want to rush to it too I mean like you got to take baby steps like can we remove half the people off the rig I mean and I don't mean that from like I want to replace people's jobs. I mean certainly you can save on GNA costs and stuff from that standpoint but there's also just inherent risk of you know you know life limb you know safety side of it. There's you know large implications but I mean like I think you know just there's got to be that you know those baby steps of like you know just the augmented intelligence where like these are informing them to do things better. Yeah. Augmented turns into automation. Yeah. Yeah, but I think that as the time go by, we can see that the people in the rig or in the field will get smaller and small and less and less because we automate a lot of stuff not just from the the data perspective but also from the operation perspective like the neighbors new rig like the pie handling you know that neighbor and extra mobile they had the new rig do the pie handling in the in the per right now because the first time when come out a lot of people come to me say oh yeah neighbor and extra mobile had automate jewelry rig and when took when we say oh no that's just a handling of the pie right that's everywhere like toss the fac gig factory they use the all the mo the robot to do the handling it's nothing new around you've been building cars and robots for years [laughter] but but I can see all this kind of like that this kind of stuff will gradually coming but not going to be happen one day yeah so it's got to be like t 10 years plus or minus. I mean like plus years. Yeah. No, I think I think the big thing there is that you know the companies the perspective of the companies need to be this is a long-term investment. This is a long-term thing and we're not we don't need to come in and say how do we replace the entire rig today. It's how do we inform them to make better decisions so that it's more efficient? We can drill faster. We can drill cheaper. We can get more oil like whatever that may be. But then over time you know you get those small wins and they compound into a solution at some point right yeah that's actually exactly like 10 years ago how many rig we have right and how many week we had right now and was the production 10 years ago and was let's say we just use and per so you can see that the technology actually improved in the past 10 years dramatically in oil in the DNC you can see that right yeah you can't use rig rig count anymore as a as a direct factor of of production because of how efficient. Yeah, I mean all of it. I mean just being able to drill all the step outs and everything from one location and then like even zipper fracking and just I mean simal frack now and all this. It's just but I see that the automation will be still ongoing no matter like like the data one or like the operation or like the execution it will gradually move on and then we see less less people in the field and then technology actually made this happen. Yeah. Yeah. I and we also have to step back and realize I mean because again we're very techn all of us in here are very technologycentric and we've worked at companies that are pretty progressive technologically like never forget there are still operators out there that are doing everything in Excel still like they don't even touch spotfire you know like and that's a 10-year-old deal in oil and gas I mean there's so many slow you know people that are still slow to adopt yeah you're going to have your BPs and kicos and you know whoever else like pushing the envelope right you know because it's incumbent on them you know and they have the money to throw at it. But yeah. Yeah. But you need to like you say they have some people is slow adapter. Some people is open mind. You need to consider that one. It's not everybody had the same mind. Right. Some people say, "Oh, I doing this for 30 years." I really don't want the chance. I'm happy with Yeah. If I don't do that every day, I feel not good. Yeah. [laughter]

So, but anyway, it's like I see that analysis is good. Like there was a when I was in China when a little young boy I see the TV had the advertisement say technology made life better. [laughter] Yeah. I mean, that's uh it's gonna be interesting moving forward to see because of all the uh the doom and gloom with AI, you know, becoming sentient and all of that fun stuff, which I still think we're pretty far off of that. But yeah, if you've ever used even just like the agents right now, like I've done quite a bit of messing around with them and sometime a guided agent can actually get you pretty far, but it still needs a lot of Yeah, I equate it to bowling with bumpers, right? like you pull the bumpers up for the kids so they don't go in the gutter. It's the same thing with the AI bots now, right? They're very young. They don't they're still figuring things out. They're And those bumpers are the human guiding it and forcing it back so it doesn't go off on these rails crazy. Yeah, the [laughter] crazy stuff like the Terminator that my favorite movie is Terminator. Yeah. The Terminate 2. Wasn't there something like an like Air Force like AI thing? like it was it kind of went off cuz like it was told to like attack and then when the people that were telling it to back off like it went to go attack that like like they were doing some testing and like you know it was AI and like they were it it saw them as a threat because it was telling it to stand down like and they didn't consider those implications of how it was trained and then it was trying to attack them because they were the ones telling telling it to stop and it saw that as a threat like yeah we we had the uh I think it was Deote brought their little uh Boston Dynamics dog to a couple weeks ago and it was watching it walk around. I was like, "This is so awesome." But this is also the most terrifying thing in the world because [laughter] it's like you could imagine a battlefield of a thousand of these things with, you know, guns on their backs and rockets and all kinds of Star Wars. Yeah. And it literally looks like, you know, Clone Wars where there's just robots fighting each other. And you're like, that's a very real potential thing in the future, right? Like it's going to happen. But it's going to happen. It's terrifying. This is a [laughter] foundation model. You can ask us to do anything right now. It's like Yeah. No, that's it's going to be fun and hopefully mostly more fun than scary, but uh we've got a we've already man, we've blown through this hour, but I knew it would be good because you're such a cool You've got such a cool experience and it's good to come here. Yep. Yeah. Appreciate it. We'll uh we wrap up with some just quickfire questions. Just give us just random stuff. Just give us kind of a short answer. Uh the first one I'll ask you is Azure, AWS or GCP. Which one which one is your favorite? I I never play with the AWS but I like I think if the next one if I want I had choice to choose I will use AWS. Yeah. Yeah. I never use it. I use Yeah. It's funny because I mean I've never used GCP but I mean I've heard good things but maybe it's like you you know always greener GCP. I'm not sure right now I use this in ano they are [snorts] sometimes they are not that update let let me give you an example at that time GCP like the stream processing GCP had a data flow or data block something like that is similar to the stream base or like the cafka but at that time when they say at 27 they say uh we ask them are they support python they say yeah we we support python too oh that's a one okay that's one example Another example that's five years ago right now five years ago we build ML model we want to use they call ML engine in the in the cloud is I'm not sure right now call AI vertex cor something like that they change the name and we want to use a tensorflow at that time the tensorflow is 21 or something like that but there wasn't in there it's like nine or 10 okay yeah what they supported and those are very different too cuz I think even when I was doing that whole TensorFlow like the model I was talking about earlier, it was right at the time where I think they switched from maybe like you know one version to another. So like I had to again I had to specify the version that it was using because it would totally break because like me I we always try to test the latest open source model. Yeah. And if like this tool we cannot deploy to Samsung that's why at the time my VP at the time said D let's use the uh the cloud ML engine. Sure. Because we sign a deal with that and we need to use a credit but I say I cannot use it because it's too old. Yeah. And then people don't understand why the why the is the attention flow is from Google. Why it's so old? But you need to understand it's from Google brain or Google AI. It's not from Google. It's a different company. It's not all under Google. So not Google. I think No, I think that's a very common. No, I mean AWS is a it's I mean it's a gold standard. It's a common if we're being honest. I mean like it is what it is. Also after that they release a news. They not going to do any AI stuff for the oil. Come on buddy. You guys are fail in oil. And then four months later, oh we're going to do something with the Ramco. Yeah. They back out of that. It's 2020. At that time we all the whole company is is discussing this news and ask me ding are they doing good stuff for us? I say zero. Yeah. Yeah. No, it's it's very true. I think they really shot themselves in the foot with our industry because of that. But it's a also a very common misconception that it's like oh well it's AWS. It's like well AWS has a million departments because there's a million different tools there and all of them do right. Um, I was going to ask you what uh what's a what's a really good, you know, book that you would recommend for for someone either getting into data science or energy or just in general? Uh, not a book, but I know the angel that's the standard professor before he had Yeah. Angel

say in Chinese know but that's a Chinese but his class like in the coella mhm there they call the uh ML specification or they had a series one series is called the ML or like the AI but they have four course there that's the best course I see explaining the neuronet the the machine learning is the best course. That's the good one. I always recommend the people. Yeah, we can put in the link in the comment that one. But that's actually the best course I see in my He's the gold standard, you know, with all that kind of those courses are really good. Yeah, Cosero. He's the founder of the Cosero. So he actually is the big guy in the AI side. He used to be a professor in uh Stanford, but he came out and and found a lot of company. He explained that one very easily like linear algebra because all the back end is algebra. So that's a good cause me that one. Yeah. He's he's got a really good course. I watched it recently right when all the LLM stuff was coming out but it was he had one on like uh prompt engineering. Yeah. Yeah. That one. That one is also on Google. Yeah. So you want to wrap us up? Yeah. Uh one more. What's the uh favorite open source model or favorite open source package rather? Sorry. or pandas. pandas. Yeah. Yeah. Pandas we use every single day. Have you messed around with polars yet? No. No. Yeah. What's uh this will be the last one and we'll wrap up. What's what's one piece of advice you give to someone either trying to get into trying to get into energy and data specifically. So assume they're a college student coming out of coming out of school and they're interested in getting into to being being a data science a data scientist a data science adviser at some point. What what would you what advice would you give them? I still back to the original advisor just like find a use case. Mhm. Yeah. And so is deliver the use case. Yeah. Use any method you see f no over cube no under cube. Yeah. That's it. Yeah. That's the my that's the secret. I get so wide by several of the lay 100%. I mean like I came into a didn't know anything but you know you get you make yourself useful and you solve people's problems and they're happy like Yeah. Yeah. That's that's very true. That's the best one. I because I in my short career like 10 years in OS I experienced the more than 10 long layoff and also too much I survive. So that's the way you you need to do just find the use case to solve somebody's problem use the most proper technology. Yeah. No, no. Yeah. 100%. That's great advice. Ding. Thanks so much for coming man. Thank you. Great talk it on there. Yeah. Thanks, man. We'll have you back.