EP 76: Christy Hopkins and Philip Heggelund from PandoScape

0:00 All right. Well, welcome to another fantastic episode of energy bites. Fantastic. Yeah, we're always strong fun, new and exciting superlatives.

0:10 But, you know, I'm Bobby Neil. I got my custody trusty co-host, John Calfan here, the Rad Dad. How are we doing? Good, man. Has Arkansas been without a coach since last time we were? I mean,

0:22 we've got the official interim of Bobby Petrino, but nothing has changed No, we will not win another football game. It's probably for the best, because otherwise, if you start winning games.

0:32 Absolutely. I'm all on this train of like, listen, burn it to the ground because I don't want him coming back. I don't like it's not. No, we're moving on. So no, just like it's good. Coach

0:42 Carousel and October has been crazy across the board. Yeah, that's the thing. It's like we need to hurry up and move. Yes, there's not less coaches. There's more and more had coaching jobs

0:53 coming out. So but obviously that was why we're here to talk to Yeah, we're a sports podcast now.

1:00 But no, I'm super excited to have, we got Christy Hopkins CEO from PandoScape. Yes. And Phillip Hegland, the CTO. Yeah. And I'm so glad you had to do the intro because I would have said Tom

1:12 again. Yeah.

1:14 It's just like my ADD is so bad. But no, but welcome. I'm glad we could finally make this work. So yes, yes, I'm excited to be here. Yeah. And so you've been in Houston and like you just moved

1:25 back, right? You were in Brazil. Yeah. Correct. Yeah I was living out of Brazil for two years, and then before that, Portugal, and before that, Italy, France, a couple of the places. So

1:36 you definitely speak Portuguese. I was going to say you're Portuguese. Yes. I am fluent in Portuguese. My wife is Brazilian. OK. And we speak in Portuguese. Like that's our, I guess, language

1:46 that we speak in generally. So yeah, it's had six love languages. Yes.

1:53 Portuguese is definitely one

1:57 of the best It's awesome and then was that purely like. Because she's from Brazil, or do you all have any developer talent down there? No, so it was just because she was from Brazil. We met in

2:06 Portugal, and then at the time, she wanted to study aesthetics, and Brazil has a program for that, but Portugal doesn't. Okay. And so we went back there, yeah. Nice. That's awesome. My

2:20 grandmother taught Portuguese around a bunch of American embassies and schools and stuff across the world, so that's a random She taught in Angola for a Chevron plant, and she taught in Portugal. I

2:34 don't know that she ever went to Brazil, but she taught all over the place. She has a place outside of Lisbon on the coast that I was supposed to go to. Instead, I got a job and I still haven't

2:44 been - I'm gonna guess it's a Kish Kish, so like one of those. It's just like the west of kind of southwest of Lisbon somewhere, they'll be each time. Oh, okay, okay, yeah. That's awesome.

2:56 'Cause I lived in Lisbon for about two years and fantastic city. Yeah, that's what I've heard. But it's overrun with anybody in Europe who's trying to not pay taxes on cryptocurrency isn't in

3:09 Lisbon.

3:11 Big crypto draw, huh? My friend was like, I'm proud to be a sweetest person until the tax bill comes in.

3:20 That's hilarious. And obviously definitely want to get into a panda skate, but you know, sometimes I start off with some current events and I think it's relevant for everyone. Yeah, everybody And

3:29 I used as the internet. I couldn't access TripAdvisor and didn't know why. That was my first experience with that outage. TripAdvisor. Yeah, I went there. I was like, why isn't, this is like

3:39 one of the biggest quotes I've said. Well, we had like no shit down, gira down. Yeah. So much. It was like, oh, well, I guess we're not doing much work today. But so as you may guess, we're

3:47 talking about the AWS outage, what about a week or two ago, or by the time we released this a couple weeks. And then just yesterday, Azure was down for a good bit of time. And I think still we're

3:56 covering this morning a little bit. I was trying to log into it. Azure Data Factory for our customer. And I was like, this wasn't quite right. It was like, I hear go to the simplified page, not,

4:05 you know, we're not quite ready for this yet, but, but yeah, so I mean, just, you know, curious, you know, if it had any impact on you guys, whether, you know, personally or, you know,

4:12 for the company, but just, you know, even have how things like that have come up or, you know, conversations like that come up. What's your take on that? Yeah. That's the wrong question. I

4:21 think there's a couple of things. It's like we,

4:24 the pendulum swung towards cloud It's gonna save money, you know, you're gonna have more uptime. There's just so many benefits. Yeah, definitely sold on the up time, right? And the, yeah, the

4:37 scalability that most products don't have to take advantage of. Sure, yeah. I mean, the horizontal, like, I don't have to go buy another server, wait for it to come in the mail or go to Best

4:46 Buy and set up a server or something like that. Just click a button and there's the server, right? But then, yeah, what happens when, I don't know. But we didn't have a good disaster recovery

4:55 plan for the cloud because it was. us to never go down. Yeah. Well, it never helps when literally the entire internet puts all of their provisions on to the same AWS instance. Yeah. Like,

5:09 that's. Yeah, I mean, US. East One is the oldest. Everyone does it still. Like, I was thinking about that the other day, like, do we need to just start spreading our stuff out just. Just

5:18 change the defaults to mass people's IP addresses and logging on with it, you know, just like. Something. Yeah. So that if that ever happened again, like, the entire app isn't nuked Yeah. So

5:28 the problem is, too, I think East one gets everything and like there's certain core services that run out. That's exactly what I mean. Route 53 is the one that. It's like you have all these

5:36 dependencies depending on what services you're using as to which regions you can use. Yeah. We have an Azure server in Canada because of whatever. I don't know what it was. Yeah. But we could

5:48 only deploy it on in Canada. That was the closest thing. Yeah GCP, like, you know, certain. regions or whatever, the only plan was to have this model available or this tier of server or

5:59 whatever that it is. Right, right. So yeah, I mean, I guess overarching point is that it didn't necessarily make our lives simpler or easier necessarily. Maybe add a more complexity in some ways.

6:10 Bobby and I are starting an on-prem company. That's what we're doing next, actually. So if you want to get off the cloud and enjoy your data, whenever you need it, we can go back on-prem. That's

6:20 actually, yeah Some of the little in-video boxes that just came out. But the problem, the DGX sparks. Yes. Certainly I've been, so we're in discussions with somebody, I can't, I guess say the

6:31 name, but they're building on-prem AI servers. Yeah. He said that the problem with the DGX sparks, and I did, I don't know enough about this, right? So he's the low level, the hardware, like

6:44 how do you get the model spun up and serving it, and

6:48 we just want to take advantage of the model, right? So once it's kind of available through that to classic completions API endpoint. But he was saying that the memory transfer rate on the GGX Spark

7:01 is only 256 gigabytes per second, which is actually not good for serving up the models. He was saying you need like one terabyte a second or something like that in order to serve them consistently

7:15 with like enough speed or something. I don't know enough about this. He said there were fantastic if you were training. I was gonna say I think their target for those is training Okay,

7:24 specifically, but if you're gonna train one that also means you have the intent of serving one at some point. That's what I thought, right? I spoke with NVIDIA about these like six months ago and

7:34 put us on the list, but when he was this company we're talking to, they were explaining that to us. I didn't know enough to be like, oh, you're wrong, or you're right, you know, but yeah,

7:45 that's interesting. Why don't you tell them the interests of why you kind of talked to the guy? That was the part why we were going to talk to him Oh yeah.

7:54 Yeah, well, I mean, there is an NDA in place, but there was basically what this guy does is he builds like on-prem AI servers for serving up, I guess like, I don't wanna say large language

8:06 models, but you know, right now they're focusing on, I think around like eight to 25 billion parameter models, but I'm working with them to try to get maybe like the OSS 120B on there as well So,

8:21 but the idea is to basically have an in-house air gap, and they're deploying these actively in the finance industry in Japan. So that's sort of like, again, you know, think hedge funds in these

8:33 guys, they're not going to use Azure AI like. The way that it was explained to me by like the head of, head of one of these hedge funds in Japan was they just don't trust it. Yeah. It's like,

8:45 they don't even care what's in the contract, they're just a complete lack of trust for their data leaving their network. I don't think there's that or just any. downtime. That's a true killer.

8:57 You got to control your own destiny in that space. And latency. If

9:03 it's running right there, you're getting asked, I mean, that latency is another very bigger part than we think it is. That's a huge thing. And even just on Wall Street and everything, that's big

9:13 up there shit running right there with ridiculous everybody. Everyone moved away from Wall Street in the '80s and '90s, and then they came back in 2000s, because latency was important. Well,

9:23 think about like with these AI coders, I've used them all just because I think they're fascinating, starting with like, I guess GBT codecs all the way to Ryder and Cursor and all these other things.

9:34 And Ryder was interesting because they built it on MCP. And so you sit there and you watch it sort of like it's basically host, I guess it's hosting its own MCP server inside of the IDE. And it's

9:49 sort of like calling the MCP server. Well, okay. So if I'm I don't know, let's say 100 milliseconds away, like I'm actually just wasting so much time because my AI coding agent is like far away

10:02 or the LLM is far away, right? So again, when we get into AI doing more things, like I think it'll come down to speed just in-house. Yeah, I did a speed test probably about a year ago between

10:16 Azure OpenAI, I think Claude and Grock And it was shocking, I did like time to first response, tokens per minute, just some generic stuff. And Grock was like 10 times faster than everyone. Like

10:32 it was shocking how much better Grock with a Q. And so I was like, oh, that's interesting. So then I went and got a quote for a Grock rack. And I was like, okay, we're gonna have to wait for

10:41 like Series A or B for that.

10:45 They start at a million dollars So, but like, also, if you're trying to stand out Crowded LLM market, so you'd would be one of those things. I would absolutely do that. So I think we established

11:01 Pandascape is an AI company, right? That's what you guys do. It's rewind. What do you guys do?

11:07 We focus. So we have a couple, kind of what we're doing now. And I'll let them talk about kind of rear. We really like to go technology-wise. We focus, my background's more in the subsurface,

11:21 data management, well data management space I know that I've tinkered with seismic in a previous life, but that's not where I focus mostly. And so we do data integration, data quality. How do we

11:32 make all of that move from when data is coming in to how do we get it into all the different apps, like the trial of Kingdom? They're dealing with their masters, dealing with public data as just

11:42 constant nightmare. Think about patrol, like they have their ocean dev kit, which has like almost no documentation that you have to - Yeah, exactly,

11:54 like everything can get into patrol, but like getting it out is another adventure. Just don't cancel our development kit license, please.

12:04 So that's, you know, and then there's all the complications of like, do you understand how to, you know, move a directional survey and well path correctly? Do you know how to do CRS? You know,

12:12 how to rotate on grid and true? Do you have do units? So that was just something I've done with the first company that I started for right out of college and kind of stayed in that space and then

12:24 kind of got talking to Philip and our other founder and said like, y'all are doing some really cool technology. It's so different than it was and you're dealing with a lot more volume. You know,

12:34 can we do something that's easier, scalable, like faster? Like, is there, and that's kind of was the foundation of starting, you know, three years ago is getting into this space, but

12:47 getting connected to the data and doing some integration is cool. It's where I saw like that we never could get to next, which is like kind of being able to serve it up and an easier way to use.

12:59 And that's kind of like where Phillip kind of comes in 'cause he's my techie guru, like, I'll just go in the data weeds. So I didn't know anything about petroleum maybe three years ago. Like I

13:12 knew what a well was and I could recognize like a platform out in the ocean and that was about it

13:19 Which is kind of funny 'cause our other founder, my father had a PhD in petroleum engineering and while that never rubbed off on me, the software engineering did. And so I spent a long time writing

13:32 code and getting into that. And so they pulled me into this and I'm like, okay, so what's a well bore? What's a UWI? Then I'm to the PPDM website. What is a well, what is a completion? Yeah,

13:43 my favorite one with my devs, I have to clarify every time Exactly. You know, like projected versus geodetic systems,

13:52 What's a coordinate reference system? So I've just been learning a lot, basically. Yeah, we both got our math hats on doing, calculating well past like, you know, verifying things in the weeds,

14:01 enough to pull her. Yes, a lot of, a lot of hair has been lost. It doesn't look like this is a toupee. It looks

14:11 way better than

14:13 mine does. I might under a hat. But, no, it's been a lot of fun. And, but I've always been fascinated by data So I came from Food Trace Ability, which is data. And what's interesting about

14:25 that is that we dealt a lot with the sort of external integration, right? So it's like, how do you get this data to actually flow through the supply chain? Which, it's one thing to get, I guess,

14:37 like two apps within an organization to talk to each other. It's a totally different thing to get two apps and two different organizations to talk to each other. And then to get them to standardize

14:46 that through very complex supply chains So I was ecstatic. to realize that we were going to focus on apps within one organization. And - He just didn't know you was trading one set of problems for

14:58 another.

15:00 Yeah. We just tricked you. Yeah, no, but it's been a ton of fun. And so where we really want

15:09 to go - we've been doing a lot of sort of side projects, just in AI, working on different types of projects, sort of like we have this one project we worked on for an old customer, mostly because

15:21 it was the opportunity to just do and work with AI, right? I mean, you've got to get your hands dirty sometimes. And this was all about transforming packets of documents. So you can imagine

15:33 you've got a container ship full of seafood coming into the country.

15:39 You've got maybe 100 documents. Like, how do you translate that into just basically like a story? And what was interesting is, like, I thought you could just throw it into an LOM. and it would

15:51 spit it out. Well, no. No, just give it to that language model. It'll figure it out. It didn't work. That's my favorite. Like just total misconception from all of our clients, both sides of

16:02 this table is it's just like, it's not magic, but they see, you know, the press releases and they see the dev days. And it's like, yeah, but even in the dev days, if you pay attention, the

16:13 graphs are all screwed up. And like, there were still things that were messed up, but the general public is like, Oh, it does this thing now And it just happened, it like learned that. I don't

16:22 know why they think that like - Just magic. Intelligent systems, and they learn these things. It's like, that's not how it works. No, it's as good as the person who programmed it. Yeah,

16:31 programmers on the backside of all of this stuff that have to build these features and these tools and the functionality and all of that stuff out. It doesn't just magically happen. Sorry, my

16:41 rant's over. No, I'm actually so happy to hear that 'cause that's like how I feel was that,

16:49 We want to connect with all the data. We want to do this data quality management. We want to make it feel like when people are working in the app, they're comfortable in working today, that

17:00 they're on the same page as the rest of their team. And making sure that

17:04 the data's complete, it's consistent, it's valid. That interoperability between the apps and that data's not static. It's always changing. People forget in subsurface world, you assume, they

17:17 deal with a lot of public data, they clean it Well, it's clean in that one project, but if they don't share it, no one else is getting that. Sort of like tops and stuff like that, right? Yeah,

17:25 from tops to elevations to surface, they fix data all the time inside of those projects. So it's dynamically changing people. Like, how much data changes every month? Well, I look at my

17:35 customers and I have just one area and I think the data's spread across five to seven different types of project databases. And on average, depending on the size, they're doing anywhere from 2000

17:47 to 20, 000 changes are occurring a month. No, I'm sorry, a week of just random things, you know, operators change this change. Some of them insignificant, but it's easier to just run all of it

17:59 than it is to pick and choose when you're going to, you know, do updates inside of there. So it's been, we had to solve a lot of interesting problems. So the sort

18:14 of technologies that we were replacing were able to, I don't know, maybe process projects of up to 25, 000 wells and all their children objects. And it could take, you know, if they were batch

18:23 jobs, you'd have to almost manually run them most of the time. And weekly, maybe some monthly, depending on your size. Yeah. And now we can do like the entire, depending on how you shape the

18:33 Permian, you know, eight, 900, 000 well-bores. All those child items across six sources. And like a first time run, you know, just to figure out what you have is, I forget, 16, 18 hours And

18:46 then after that, it runs every hour. Okay, so I'm doing a little a lot. Yeah, right, but yeah, just the fact like before like in different things that I've used like that just work would never

18:57 get that You would just never do get a hundred thousand. Yeah Yeah, you have to condense it down right right and you break it down to 200, 000 wells or a hundred thousand depends on what Software

19:07 you're using it just becomes you know not doable or you'll be like okay Only in this field run that like every week in a batch job and then over here run it So now it's like we can just look at

19:17 everything all the time So exploration is always up to date because they're usually looking broader versus like development is like right now Yeah, you know what they they want everything So just

19:27 kind of the fact that we could do the whole thing and look across everything because now you can start having a peace of mind that You know everyone's working from the same right so something's wrong.

19:37 I mean that's someone's going to you know Like I had a VP at Newfield and the reason why I'd even come in is like, you know You're your competitive advantage is your data and he would walk in and be

19:48 like there's a well missing. Where's it at? Like, why don't you have all the log curves on there? Like, I just told you guys to fix this elevation last week. Why is it wrong? Like, oh, that

19:58 was so-and-so's project. Well, why don't we all have it? Like, that's, you know, so not every VP or asset manager maybe thinks that way, but that guy was that, like, why aren't we all on the

20:09 same? Yeah, yeah. No, I mean, that's the same way on the up side, right? It's like, I've sat on many frack jobs where there's sand trucks just sitting there charging us to merge for just

20:20 burning money because someone didn't tell them that we screened out six hours ago. And it's like, okay, well, literally like the money on fire. I'd rather do that than watch the sand trucks so

20:30 that there's no reason. Like, can we go to a nice dinner? Yeah, like something, 'cause yeah, they're making tens of thousands on that. Just sitting there. There's a lot of inefficiencies

20:40 around. And so, yes, you know, how can you, just how can you make it easier? I feel like I look out there and there's so many people doing different things like you said, in field or just

20:50 knowing like. I'm going out to this job, Hey, put all these things, equipment on the truck, because you're going to go ahead and take care of these other tasks while you're there. All the

20:59 efficiencies that they're working on, pretty fascinating. It's amazing what can happen when you have good communication across the company. Communication being key. So I'm assuming your customers

21:13 are ENP operators, right? So what disciplines are, I mean, it sounds like it's centralized, essentially from a lot of corners of the business, but is it more around geology, reservoir, like

21:26 that? Subsurface mainly. The potential is there for it to keep growing past that, but the focus was kind of where our domain, me and the other founder, our domain expertise was kind of in the,

21:37 just starting in the well space and we're also forward thinking will be some seismic space But I feel like it has the ability to do more, but we're just kind of targeting in the space It seems like

21:50 anywhere where there's opportunity for interpretation, so to speak, or tweaking. Yes. Yeah, that's a - that makes a ton of sense. It's just a lot of work and to move it right and that the users

22:04 trust you. Because that's one of the things that's usually a part of a proof of concept is sitting down. Yes, you might be working with the data manager. But a geotechnies to come in, we make

22:14 updates. You need to go in and look at it on a map Look at the bottom, did we get all the way down to the bottom hole correctly on the, like laser eyes, it's usable when we're done touching the

22:26 data. Yeah, well, and then the fact that it's done once and it's available for everyone. And so I can only fathom how much

22:34 rework or duplicate work has been done in our industry on any lots of data things. Anything data, right? Like, who's like, oh my gosh, you took away my loading tasks, like nobody thinks that.

22:45 No.

22:48 I've been on the data side and I've integrated it with Kingdom, so at least through data warehouses and stuff. At least it has a sequel back end. But even with that, I was still hesitant to write

22:57 back to, you know, because like, it's a lot. You're reversing from here. You can mess up the software, basically, if you push it in the wrong or not updating the primary keys the right way,

23:05 you know, whatever. Oh, no, you know, it was great about Kingdom is that if they require the first point of a survey to all start exactly at zero, not 000001, it

23:17 would fling on the map And the opposite direction, like, literally had to write code, make sure that the first row is always zero. I'm like, well, what do you do when you don't have it? Like,

23:25 when it doesn't start, is there, no, we require zero. I'm like, okay. So like, I told the users, like, we'll push it in. The well path will be correct. Your survey will visually match what

23:35 you want. But if you press the calculate button in Kingdom, all your XY's are going to be wrong because we've calculated with the type point before writing it. So this is the kind of knowledge,

23:45 like,

23:47 that we're bringing built into the table is not just built-in connectors to these products, but also the built-in, like, how do you do data management best practices for that automation? Yeah,

23:57 'cause even I was at, I was gonna mix a petra or a portrayal or kind of file base behind the scenes. Petra would be, I don't know about Petra. It's not, well, I don't know what Petra - No, but

24:06 Trev's not a file. No, the

24:08 Petra product is, but Studio's SQL Server. Okay, yeah, yeah. But Petra is definitely file-based Yeah, well they use a, gosh wait, it's called file system or something like that. So they use

24:19 that SQL Server file system as well. So it's like files, but stored in the SQL Server, also makes it a pain to take it anywhere outside of SQL. But yeah, yeah, it's - Lots of opportunities for

24:33 fun dealing with these subsurface apps. Like think about, there might be analytics coming on this, but these true interpretation I time knew one was last the once out Like, packages.

24:45 Like you have Kingdom and Sizewear. But Kingdom is the new one, right? Yeah, 'cause they took on Petra and like, they tried, and they tried to kill off Petra, but the users are like, you're

24:55 gonna pry it from my cold dead fingers. So like, but there's not been a new app. Like, Geographics is still around. I think they call it like G-verse or like, it's got some new flashy names.

25:08 Ask G-verse. Yeah, it's like,

25:12 I'm gonna make that

25:18 But yeah, there's no new subsurface modeling. I mean, you know, pretty much everyone's in like a patrol, a kingdom, a Geographics Petra, but you don't have any, but you look at all the

25:28 knowledge in those packages and you're like, no one's gonna replace all this. They're just gonna keep making this old thing work. No, yeah, I mean, and I've argued that I mean, well now SP or

25:39 IHS or whatever you wanna call it, but like, if they ever get rid of that, their business is done. 'Cause so many people have Enveris and

25:49 SP or I, just for everyone to call it. Yeah. And the only reason is because they have Kingdom or over in the direct connect. And like, 'cause I mean, I think two different spots are like, well,

25:60 we could put Enverisated in there. It's like, you can. But again, just for the same reason I was saying, it's like writing into that schema is not trivial. And maybe you all help people do that

26:09 sometimes too. Yeah, that's so inverse would be like another data connect. So we connect to all these different applications and in the back end, we're kind of by default when you're doing

26:18 integration and normalizing. So you're bringing everything normalizing it together so that you can then bring it to the center and then be able to push out everywhere. And like, what does that

26:29 application need? 'Cause they all have their own little works. But just having that knowledge of how the back end works, like there's not many people will do. And then even like, and I think we

26:38 can get into a little bit like say kingdom, I know like they store like logs and deviation surveys as like the blobs. Yeah. Gotta decode those. I actually beat them down enough where they - Do

26:50 they finally get cave and give you the answer? The answer, but even still I try to decode the hex strings. It was not trivia. You get like the first part of it, but then - Yeah, you can come

27:02 talk to us. I don't know, like we have something that how the blob is organized or figuring out those, 'cause it's the same thing when you get into open works. Like, I think the surveys in there

27:11 as like rows, but like the well path is in a blob, you know? And so like when we talk about when people want to come, you know, you start getting AI and machine learning and stuff and they're,

27:23 you know, inside of a company and they're coming and ask the sub-service team, like we want this data, well, how do you give it to them? Like how do you tell them to join all these tables? And

27:31 that's kind of that next level where we talked about going, it's like the MCP server. You know, instead of telling you here go to open works and this is the tables, this is how you join, This is

27:42 how you rotate this, how you normalize how about like I just say like you say I want all the surveys you know this list that you just go ask it for it and I give it to you do you want it in XML do

27:53 you want it in JSON do you want it in an LES format like that you could start saying what format do I want back you know what is your product oh what is your product you know how does it consume the

28:05 data you know I don't know what Python like what do people use typically in Python is it JSON oh gosh I mean JSON's JSON is that's a comment I find it like not readable but that might be the old me

28:17 like I read XML you just need a you just know XML makes me gives me hives oh see I feel the same about chase to me is what gives me I look at I'm like where are the curly brackets yeah yeah this tab

28:30 based I can't deal with this yeah I mean yamels like I use a decent amount with like dbt for like right for you know it's like configuration type stuff but like I would never God forbid someone That's

28:44 what returned from an API,

28:48 I mean, at that, basically. Yeah, I can't see it. I don't find Jason like easy to read, but maybe it's because I've looked at XML for so long. It's just for XML, like, I'm glad we're

29:01 going to find it easier to read. I think like my problem, straight up, XML is not too bad if it's, you know, but like some of this, like, then people are embedding actual data within like the

29:09 tags. Yeah. And that kind of stuff becomes way more like, you know, it's being on various like open invoices API is this bizarre mix of JSON and XML and like, but it's, you know, but the,

29:21 there's important information stored like in, inside the tags. Yeah. Yeah. You're not where it's not super. You have to parse it out. Like do something. It's like, yeah. Is that like the raw

29:31 that comes out of the model?

29:35 Yeah. I mean, we're just hitting the API like, okay, there's certain parts of the API that return and then like when you actually get like the true like counting data part, it comes out in XML.

29:43 I know that like when you're dealing at the I don't know enough about this, but like when you're actually Communicating with the raw model or something like that. It is like some sort of tag-based

29:53 language that's coming in and out of it Yeah, I saw something the other day and they were I Want to say they were set like pushing XML instead of JSON because it's Smaller, it's less tokens or

30:06 something. Yeah, I remember what it was but see that I thought there's less white space even probably or I guess that's why I'm more recommend like in their documentation They recommend basically

30:16 like here's an example and to wrap it in XML tags called just an example. Yeah Interesting, I thought everyone was trying to go to Jason But I found XML is easier to read and also if you're doing

30:29 any Transforming it to another format you have XSL T which is Like you don't really have a nice a nice version in Jason. I think they have it, but it's not as extensive as um, XSLT, like for

30:45 transforming the, from one XML format to another, you use XSL. Oh, okay. So it can kind of a plot, you know, match a tag and then output, you know, something else. So you can take one format

30:55 and output another really easy. And it's very fast, like, very, like, extremely fast. You couldn't, yeah, like it took forever, like XML, you have X path, you have X, XSLT. I think the

31:07 only advantage, in my opinion, that JSON has is native support and JavaScript basically, our TypeScript, right, and then the

31:15 JSON schemas are better, in my opinion. They did a much better job because with XML, it's so strict that you'll get into these situations where it's like, Okay, I can't really do anything here,

31:26 because it's more exclusive in

31:30 the way that they, I think, that they designed the schema language where I feel like JSON is more about, Well, does it at least have this? And if it does have this, does it also have this?

31:42 Maybe I'm using this in more like inclusive versus an exclusive schema language. So with that, I think JSON schema allows for

31:53 extensibility, right? Like you can, yeah, maybe I have all of this extra stuff over here. But what I really want to know is like, is this here? And is this in this format, right? Yeah. Yeah.

32:04 It's, I mean, they're both useful for taking tab, any kind of tabular or just like nastily structured data and getting it into a format that you can then convert, spread some structure to, you

32:10 know, especially now, like, you guys like parsing stuff out of documents and stuff, like, no,

32:21 I mean, one of the, one of the latest things that we just rolled out is a revenue statements kind of workflow. But all these operators get these revenue statements and it's fractions of a percent

32:31 of their interest, their non op interest in these wells. And so they're getting like really long literally mailbox money but it's, you know, you might have a bunch. And so there are hundreds of

32:41 pages of like, 30 cents. And then here's the tax on that 30 cents. And then here's the severance tax on that. And it's by asset. And there's all these accounting there. And it's in this gnarly

32:53 PDF, of course, tabular format with nested headers and all of this fun stuff in a PDF. And so they also provide a there's a more expensive service that you can upgrade to to get it into a CSV

33:08 format. But the CSV formats are not standardized. And so the CPA is we're just like, there's no reason to pay for it because it doesn't help me. It doesn't make my life easier. And so we're

33:18 pulling it into JSON. And I'm like, how do you want to see it? I can, you know, I can transform it. However you like. And so you're like, once I can get it into something described, now we

33:29 can do anything. Exactly. Yeah. Yeah. Serializable or whatever. Yeah. No, I mean, that's in my, from a data perspective, as a data person, I feel like that's one of the most like slept on.

33:40 interesting things about large language models is just being able to take data from PDFs, from unstructured, traditionally unstructured documents and then say, Okay, well, instead of me having

33:53 Billor the reservoir tech or the e-techhave to manually type all this shit into well view, I can just pass it in and an agent can do thatautomatically by itself with repeatability and accuracyand

34:04 it's not gonna get tired, it's not gonna leave earlybecause they've got the kids football game that night They're willing to work on Saturday night. But it's those are the things though that humans

34:15 hate doing and it's why we invented computers today. It's data engineering tests, like data engineering tests, like lately there's that new term and I'm like, this finally describes what I do

34:24 because it's not ETL, it's that data, like it finally had a name 'cause I'm like, I'm not really an analyst, you're like, you're a data engineer, you're doing all those tasks that people just

34:34 don't wanna do. Yeah, well, I don't even have to be,

34:38 that happens. Like I was expected. They're doing it and they don't even realize that it's like, um, I think the, one of the latest books I read, they mentioned a study. I think it's like, uh,

34:49 machine learning and AI engineer spend about like 80 to 90 of their time just getting the data and prepping it. Yeah. That's it. Um, I think, uh, there was a report and I want to say, gosh,

35:02 2015 around that time, um, that basically, no, this was early This was more recent. This was like host 2020, um,

35:13 geoscientists and petroleum engineers spend 50 of their time doing data engineering tasks. Yeah. I mean, that's like, solidly solid

35:22 and minimum that's what they're spending. I couldn't believe it. And I'm like, yeah, no, the number we use is 30 to 50 of our, of the industry's time is spent looking for data. As I say,

35:32 that's just looking for it. Before they try to load it. That's before they start using it. That's just looking for the data. Like before they try to load it, before they got it, like, you know,

35:39 kind of formatted to get it loaded. I just remember working work that someone did four years ago. Because they didn't capture it back. Yeah, and they laid off half the company to happen and then

35:49 no one knew. Yeah, they hired a bunch of new people. They have no visibility. Someone's like, give me a new project and they start all over again. It's like, before Google, when it was just,

35:59 gosh, I think SGs and Dogpile and stuff like that. I mean, you could find things, but to me, Google is when you could find things. Like I could, it was a, even today, like when they drop that

36:12 AI mode, it's phenomenal. Like, I'm sorry, Bing, just. Yeah. I don't know if this is like an anti-monopoly play by Microsoft and Google or something. Like that's why it's still around, but

36:25 there's no competition. I mean Yeah, my cousin told me I'm bad at, you know. searching the web and he's like, you just didn't enter in the right keywords, and I'm like, I don't know.

36:38 No, I mean, I tell people all the time that like using an LLM today properly is just like, it's the difference in me using it. Me using Google versus my parents using Google, right, where they

36:49 know nothing about keywords or quotes or any of the, you know, tricks about it. They're just going to type in their full sentence. They're going to type in a full question of a sentence into

36:59 Google. And I'm like, that's not how this works. You don't need to put in every single preposition and every single word of what you're trying to get. It's not. Do we get to Loubies?

37:11 But it's, yeah, it's the same thing. Although I will say, uh, perplexity and, uh, open AI have now both come out with browsers that are AI enabled that are taking shots at Chrome. It wasn't

37:23 very curious. Wasn't perplexity trying to buy Chrome or they just made a try and make a I don't know if that was real or not. 30 billion offer and their market cap was 15. So I wrote it was like,

37:35 oh. Yeah, they're just trying to get their name in the, they were getting the name of CR before. Oh, they were just trying to get the drop of the room. That was like, yeah, maybe that was

37:42 their cheap way of getting ads. I'm sure you'd get some PC money behind you if you could actually buy Chrome. Like it's like the most used browser in the world, not monopoly, but somehow monopoly,

37:55 yeah. Well, I mean, like the engine of Edge is just Chrome Like they - Yes, they moved to the Chromium. So much as Chromium. Yeah, it's like, yeah, I could see wanting to buy it for sure. I

38:06 mean, I use Edge mostly because it's like once you're logged in, like it, first of all, like tabbing, you can see the windows separately. Like the tabs are separate when you do the, I don't

38:16 know, alt-tab on Windows and stuff like that. And it's like auto-integrates your Microsoft and try ID that you've logged into the machine with stuff like that. So to me, it's just, it's more

38:27 convenient. but I would have never used Internet Explorer. I mean, when I was like, the old joke was like, Internet Explorer exists to install Chrome. Yeah. You

38:40 should have just left Netscape on there. But even like, I mean, the first thing I'd do in Bing is go change the default search to Google. You're like, please stop. Is this painful? If it's

38:50 never having dial up and using like AOL instant messenger and you're like, you've got mail. Oh yeah, it's a wonderful version Like, yeah, but the fact that you had to do dial up and nobody else

39:01 can be on your home phone. We were just explaining that to my kids this weekend. We were in the car. They're totally confused. So yeah, like there was one phone line. No, 'cause my daughter was

39:10 asking, what is a land line?

39:12 And it was like plugged into the, you know. Have you seen the videos where they give them the rotary phone and they put the kids and they have an hour to figure out how to dial the phone number?

39:22 But then like my wife was like, yeah, and like if someone was on the internet, no one else could use the phone and you've picked it up. Do you know what I mean? Is that still on the intro to our,

39:30 I think so, yeah? Yeah, like the intro, like sound or whatever it is, like that dial up, our podcast intro and exit noises I made of just like all these nostalgic 90s, technology sounds, it's

39:42 like PlayStation one boot up, it's the, that was it THX, the, you remember, and the movies, the big bass, like you build up

39:54 the Sega, the dial up sound, it's a bunch of it all mashed together because that's what I think of when I think of technology. That's my ex-message sound is the, you've got mail. Oh yeah, yeah,

40:07 it was a, we don't get too far off, but like there's a new toy out that my daughter really wants and I don't think we're gonna do it, but like it's like canned phone, but it's basically it looks

40:15 like an old landline phone, but it uses Wi-Fi and has its own number. No, that's, I don't know, I'm actually on. But it was like sold out, like when my wife found it,

40:25 September, wherever it was like already sold out. I'm on this train. Oh, wow. So the reason is instead of getting kids a cell phone, you get them a fake land line, which is still a phone, but

40:35 it teaches them to like learn phone etiquette. And you come, like there's an entire generation that doesn't like talking on the phone. Well, they're afraid to answer the phone. They're afraid to

40:45 answer. They're afraid to call. My wife doesn't like calling to order food. I'm like, what, you used to do this. Like this was very normal. Like you used to answer the phone and had no color IE

40:54 Yeah, exactly. Like we didn't know who was calling. No less what was gonna happen. No, it's the, you know, whatever residence. Yep. And so anyway, the whole idea there is it's like, it

41:04 teaches them actually, you know, some kind of social skills that keeps them off of. They're having adulting classes for kids to learn how to interact or even adults to learn to interact with people.

41:15 Like I find that mind blowing. Wow, that is my phone. Like I just remember taking my niece and nephew. They've spent a lot of time with me and like when we would go to dinner that. making them

41:24 order, just that they learn to be comfortable talking, like I didn't realize how much that's like social skills that they need. But I just took that for granted, but I realized now there's classes

41:35 out there to help people with that. It's mind blowing. I started making my daughter if she wants a refill, if she wants to whatever, I'm like, okay, they would rather do without most the time

41:45 than to actually have to ask a person for a refill. Yeah, that's true. Yeah. Yeah, crazy And like Brazil, though, they don't even say, what's up is the thing outside of the US, by the way.

41:55 Yeah. Nobody calls like with a normal phone number anymore. But in Brazil, it's like only voice messaging. It's like, I'll send a message to

42:06 my brother-in-law and voice message. Yeah, it's just only voice messages. That's it. I don't know anyone used that text I never thought about that. Well, my wife will sit there for like two

42:20 minutes, like recording Oh my gosh, who has two minutes to listen to this? Right.

42:27 It's like the spot where the voicemails are apparently. Well, they've got, the most hilarious thing is like, they'll say it and they don't use like the talk to speech. No, it's just the voice

42:35 message. And then when they listen to it, they listen to it in 2X. So it's like, you know, it's almost like Alvin in the chipmunks, right?

42:44 And I'm just sitting, I'm like, you know, it actually is more efficient. And then I realized I was like, oh my gosh, I can understand Portuguese in 2X Yes, this is, yeah, it's even better,

42:54 right? Yeah, but no, it's, yeah, I don't know. I got dropped and I flew out, I think, 2015, I realized, took me three months to realize, oh my gosh, I can work remotely. This is my first

43:07 job. And it was a remote job, but it took three months for buttons too, and things to click. And I landed in Barcelona in the summer of 2016. And I didn't know Spanish. I didn't know. Could be

43:23 Spanish. I just, I literally got in the taxi and was like, I need to go to this address, Airbnb existed. So thank gosh, I showed up to this Airbnb. I didn't know anybody. And I made some, it

43:37 was like one of the best summers, I think ever, just a fantastic time. So it's hard to imagine that kids can't ask for a refill to like a soda, you know. It is crazy. And COVID kind of made it a

43:50 little bit worse Yeah, that actually too, that had, that's crazy to me, you know. Yeah, that

43:56 restriction or whatever. It's amazing how that just short little window, I mean, I just think from my perspective, how much after that, I hate being in really crowded places now. Oh, without,

44:10 yeah. Like COVID, I don't think it ever bothered me. Like, of course, I don't always like busy airports or things, but like after COVID, it was like, people. It's very true, it's very true.

44:23 That's also what I think it did to the kids. Yeah, yeah, no, it's true.

44:30 It's like, I don't have to talk to you. I think it's, yeah, expectation, but even just the technology. You can text people now, everything can be done, yeah, gonna order online, do all the

44:38 things. You know, half the time you can't get those little suckers that answer the phone, they don't think it's important. Oh, yeah. I call you answer, like, that's how this works, or you

44:48 don't have a phone.

44:51 That's why you keep this phone. That was a lot of good time. My wife will call me not to talk on the phone, she'll just call, just to be like, I sent you messages. So she

45:01 like calls, it's an alarm. Yes, it's an alarm for you. Yeah, and as you see it ringing, you're like, oh crap, it was like three messages. Yeah, she'll just - I'm literally going to

45:09 read on the phone. Yeah, but I was listening to this podcast on the other day, and they were talking a lot how there were all these promises of technology. And I think I hope I'm not misquoting

45:22 this, but the guy said, Peter Thiel said, they promised us flying cars and all we got was 140 characters. And it just resonated, especially with AI. I just feel there's all of these promises

45:36 coming out. AGI was supposed to be here about three months ago. I mean, and it just feels like we've gotten nothing. And I think that there are a lot of really good applications that have come

45:48 along. But - Everyone's figuring out what they can do. Yeah, which is, it's awesome. Like how can it bring value? Yeah, that's the thing. It's like, you have to be realistic with it. And

45:59 with every new technology, it goes through that hype cycle of like, they're gonna promise you the world because it's a big tech company and they have really good marketing teams and they really need

46:08 you to spend the money so that they can make their money back., And at like it's the end of the day, it's still business But yeah, I know the small companies are like, to show, we are solving a

46:18 problem. Yeah, but you get through that and then you go through a downturn of the trough of despair where everyone's just shitting all over whatever that technology may be for a while or talking

46:29 about how it doesn't work or having papers by IT come out and talk about how enterprise deployments aren't working and all this stuff. And then the reality comes out somewhere between those two, but

46:39 they're still valued to be extracted, right? Well, it's like blockchain. I, you know, in supply chains, blockchain was supposed to be revolutionary, I think. Because for traceability, that

46:49 was like one of the first things I ever heard about was that Walmart one, where Tristan, from South America, whatever. IBM went on and like, it was a big thing. Everybody was like, if you're

46:59 not doing blockchain, you're not doing traceability. And I'm sitting here going, I'm pretty sure like in the second paragraph of the original white paper, it says this is meant to solve a

47:09 double-spend problem And nobody, which there is a double-spend problem in food fraud, it's a massive problem. nobody was solving that. It was all something you could have solved with basically

47:21 just digital signatures. Yeah. I mean, it was just - I don't mean to share a database. John and I were, that was the first time I ever heard of that case study and we went to a, but we worked

47:30 together in a previous life. And a UT blockchain thing. UT blockchain thing. And they talked about that and then I was like, Jimmy, something or other, that Asian dude we're in a big cowboy hat

47:40 and he's like, it's a bleeping database with rules Like, it's not like, it's just, you know, everything old is new again in some ways. It's like, you just have to get that kind of, it's more

47:51 just turning inside out and giving that kind of consortium whatever. Yeah, it's just a nested if statement. Yeah. I was like, once Bitcoin hit, I think, gosh, 300, I think when it hit a

48:03 thousand dollars, that was when people were like, this is gonna be the best bubble ever. Like, and it has been, I mean, when I lived in Lisbon, I mean, it was just like left, right, and

48:14 center. You just saw all these guys and. they pull some out and they dump back into another coin or whatever, I don't understand it enough, but I'm just sitting here watching these guys and

48:23 they're just making out like bandits. And I'm like, so what are the, 'cause everybody promised these like, oh, it's gonna change voting and like, you know, legislation, like there were supply

48:35 chains and I'm just sitting here, I'm like, Identity, that was the one I bought. Yeah. Like, oh, you could have a distributed leisure for identity that the government doesn't have to have

48:43 direct access to, but can still verify your stuff. That makes a lot of sense. Nope. No, we've got a bunch of guys who got really rich and now are having a fantastic time in Lisbon. That's what

48:56 we're thinking. They're having a great time, I'll tell you that much. And Ibiza every weekend. But we'll, I think you're right. Like, it's gonna swing back. We're gonna find like where these

49:09 LLMs are going to land. What do we do? Like, what do we really get to feel that they do? Like, we have all these great ideas, but it's just another tool. Yeah, and that's, thank you so much

49:18 for saying that. As I say that, every time I talk to somebody about it, it's like, you can put, you know, nail in the wall with a drill, but a hammer is probably a better tool for that. Right.

49:31 They're all tools and they're all good at solving certain problems, but you have to use the right tool for the right problem to get the best results. Are people gonna be, oh, can we use AI for

49:39 this? Like, you can, but you could also use your GIS platform or you could, yeah. Like, do you wanna use what you already have? Yeah, we do a database join, you know, you've got the data

49:47 there, I mean, some of those - It's sometimes very simple, but we have all these big ideas of where we think we can go with it and you're like, really, we just needed a spreadsheet. Yeah, like,

49:59 I tracked my vitals and you got Apple Health and like 10, 000 apps and it was like, all of this is too hard, I just pulled up Excel. Yeah. There we go, that's easier You

50:13 know, that's when we think about when you're. anything you're doing, right? Why is a company gonna go with your product or service? It either has to be so painful what they're doing now that

50:25 they're gonna move, or they're gonna solve this problem that it's not enough pain, but there's something else that you have. Like that's it, that's gonna be what brings them in. So we don't

50:39 really wanna be the

50:42 way that I see it is like where we wanna come is, we sort of work kind of connecting to all the data sources, right? Connecting to Kingdom, Studio, Petra, all these things, right And? we're

50:53 running rules and every day we're working on new rules and new things that we can do to make sure that the data's high quality. That we're not moving bad data, yeah.

51:04 Yeah, but that's such, it's a sorry to interrupt you, but it is such an important thing to bring up because just like with machine learning, LLMs, in RAG and all of these other things, if you

51:15 have bad data to begin with, it's not gonna give you good results, right? Like we have a bunch of different projects that we've been working on and it's like, one of them is literally scanned

51:25 bankers boxes of files from going back to the 70s, it's handwritten stuff, it's all this crazy stuff. And like, I literally wrote a streamlet app just to review the PDF next to the markdown

51:38 extraction because every single page is scanned And most of it is not, you know, most of it scanned from the 90s to back to the 70s. So it's like, it wasn't even, half of it wasn't even digital

51:50 to begin with, right? Like it was just printed and then filled in by hand and you're just like, man, I can't put this in my rag database because it's pulling out just garbage, literal, just

52:01 strings of

52:04 letters that don't make sense. You have to have data quality that's checked. Like, is this good enough to be used? Like what's your sanity to do? Yes, you have to have. What is your sanity

52:09 check? Like, we're not trying at this point

52:13 AI is like, can we start making like the data AI ready and available? Like, can we do data quality checks? Could we, I don't know, make like a template and say, like, I'm looking for this kind

52:25 of data and it needs to pass these kind of quality roles to be able to consume, yeah, you know, over in a process. That's where MCP, so I followed that pretty closely. I first thought that,

52:37 okay, this was a, I don't know. Another password Yeah, because like you see like Facebook came out with like open social, I forget what the name of the standard was, but it was like their whole

52:47 thing of like how to make the social media networks like more interoperable. Um, and there's, I don't know, dozens of examples of this. And so I saw this from anthropic. I said, Oh, okay. And

52:60 then I saw, uh, open AI adopt it. And then I saw like tell Eric drop an MCP server, right? And then I started seeing cursors, like Azure MCP server context. So I said, Oh my gosh, I think

53:13 this is like one of those rare cases where this is really going to be something. And that way, it's just a protocol, just like HTTP or whatever, right? It's like a trans I was trying to

53:26 ask like how to explain it to someone he's much more technical in this. But you know, it's a translator, like it's going to be the thing that talks to the database Like instead of talking to the

53:37 database, you're talking to MCP server and he can tell you all the kind of requests that you can make and what kind of permissions that you need. So you're kind of giving them and you're serving it

53:46 up hopefully in a more consumable way. And that any LLM, it's um, what did you call it? Like LLM agnostic. Yeah. So any LLM can talk to it. So like things like that really matter that all of a

53:59 sudden can I expose your subsurface data through in your environment. And I think that's also where it comes important that you have the hardware on-prem, because - Nobody's gonna let their data out.

54:08 Especially on the sub-surface side, though, I mean, people have tried and failed a lot to put subsurface workflows in the cloud. And it's just like, whether it's the visual side of being able to

54:17 visualize things and the GPUs needed for that, or even just like the moving the data around, like the size that like, I have not heard of anyone having success utilizing the cloud for like geologic

54:27 type. I, you know, the only thing is it either all has to be there or all has to be on-prem. Like there's no, there's, you know, like, I forget a sitting with a friend and she's sitting there

54:39 and they were trying to design 'cause they were wanting to move seismic. I'm like, she goes, okay, she goes, you see how close and she goes, we're sitting really close. You see how close we are?

54:48 Like, that's how it needs to be closer. Patrol needs to be sitting like, or I don't think it was patrol's decision space. Like it needs to be sitting like right next to the seismic. It needs to

54:58 be on the same backbone. Yeah, I'm trying to explain that to people. I was even like talking to a private equity company, probably be back with it. private equity company and like geologists,

55:08 they're like,

55:10 he was having issues just with like the local file share, like latency between that. Like he, and so he just keeps it on his tower. Yeah. Like I mean, like his to a C drive, 'cause like it's

55:21 not otherwise pulling from like, and then they've got like ridiculous bandwidth internally. It's just like everything that the servers are sitting right down the hall from, you know, like, and

55:30 it's just, that's not performing enough. I mean, there's like a router checkpoint where it's like, you just have so much traffic going through there, and I don't know, maybe it's a 10 gigabit

55:40 line or something like that, or maybe you have something, you know, like fiber or something like that. But you've got 30 users out there, like how many of them are in there pulling up a model and

55:48 trying to spin it or do, like, there's companies, I can't remember the name of it, but basically they're doing where they bring their own, you know, rack servers and installing your data center,

55:59 but they're, you know, doing where they're like, okay, we're gonna manage the database and the application and the hardware. and give everyone VMs to this environment so that you can do kingdom

56:09 and patrol and whatever, but they're kind of trying to contain the environment, knowing every, 'cause they're trying to combine that, hardware, app support, and the users. I heard of that

56:20 company, I think. I can't think of the name of it. That's why I think it's gonna kind of come, and MCP is really interesting because it's not HTTP. I think it's GRPC. Yeah, yeah, I was

56:30 thinking it was similar. I was like, it was a protocol like HTTP. It's not HTTP. I think you could, I don't believe you could implement it with HTTP, but they use GRPC, and it's a two-way

56:43 connection. It's not like, it's almost like a request can come in, and the MCP server can actually ask back questions. Oh, yeah, that was the other call. Like, do you clarify this, or you can

56:54 even route? Are you asked for this, but I need to know if you want A or B, yeah, okay. You can actually route, so it's like, you can host the MCP server without actually having the LLM inside

57:05 of it, right? So you can actually say like, okay, well, I'm going to route back LLM request back to the U. And so you need to process this, give me back the response. So you can have it to

57:14 where you can almost like, I'm the LLM asking the MCP, but all the LLM requests are gonna be outsourced back to me. So you can almost even control like the LLM that's being used for the LLM parts

57:27 of the MCP servers For privacy, I guess. I mean, yeah, I mean, privacy, but also like. It's also like. To the best model for that use case too. Oh, yeah. Yeah. Yeah. Yeah. Yeah. For, I

57:39 mean, even just like, 'cause the way I describe it is it's basically an agent within an API. Yeah. It gives a language model the ability to call APIs without it really knowing anything about the

57:51 API. There you go, that's what I was. 'Cause the MCP server is what is essentially hosting all of that information. And so it's like you have functions within the MCP this or do that, delete this,

58:02 whatever, just like you would on an API documentation, and then, but it's like, oh, well, if I need to call this function, I'm going to need this variable and this variable and this variable,

58:16 and then it tries to go and find it. And if it doesn't, it will then turn around and ask you, okay, well, I've got these two things, but I need this last thing in order to go do the search.

58:25 It's like, right, it's kind of like a on Mother Duck, where when you start typing your SQL query and you screw it up and then it autocorrects it for you. It's very similar to that. It's honestly

58:35 kind of like that. You need a little bit more information. Yes, literally, but

58:41 it's wild because it really does open up the hole. I'm finding that really cool. The MCP server, he was kind of saying it and I'm like, I'm not getting the cool part, not getting it. And then he

58:50 would describe or I hear him talking to someone else. And then finally, it just started clicking that there were so many ways in which you could use. It's like, I can render it. You could ask it

58:59 for data. You could, you could And whatever you functionality, you give it like, yes. Okay, I have subsurface data. Like, how do you want it? Do you want this ERS, do you want it in this

59:08 format? Do you want that you could start saying like, what do I want? And how do I want it? I just didn't think about, because today we only know how part of what makes some of the projects fails

59:19 is, it's not enough that you want the data, but you want it a certain way. Right, yeah. And unfortunately, not everyone in data science, machine learning analytics, truly understand some of

59:29 that normalization on how to get to where they want. So if you can allow them to ask,

59:37 problem solved, you don't have to give them a table where the data spread in five places and they have to understand how to merge together, you can just say like, here's the well bore, here's the

59:45 log curves, that has its own value just to the other side of getting into whatever it is you're wanting to do downstream. That's just really exciting. Yeah, no, the most exciting part of it to me

1:00:00 the biggest problem with our industry especially from a data perspective historically has been that we just have so much siloed data and so a lot but we also need we need to keep that data siloed in

1:00:12 within the organization but we also need to augment it and enhance it with external data and how do we do that in a easy way MCP servers in my opinion with like a rag paired with it are like that's I

1:00:26 feel like that's like we're gonna see something really cool happen here in the next six months here with this space like yeah well I feel like as a program as programmers

1:00:37 we're spoiled like we got to play with

1:00:41 GBT codecs before Chad GBT was like online which to me at the time I was like what is

1:00:48 this but even then like we get all the coolest tools first because I guess like we're building them and we're - We want to solve our problems. We're the first ones that are chicken to like, try it

1:00:57 out. Yeah, we're the ones that get fed up. Beating our head against, and we're like, not sure. We're like, if this will save me an hour, I'm doing it. My life needs to be great before I make

1:01:05 other people's lives great. All right. That's, that's, yeah. Development center. But through it all like, just the agents that have been coming, like a watching cursor from the very beginning,

1:01:19 'cause I first played with it when Lex talked about it on his podcast, and I was like, okay, this is kind of cool. Like, I see where this is going, but guys, come on, we got a little bit to go.

1:01:30 And then with the latest LLMs, and obviously they have this

1:01:35 massive amount of, I don't know what to call it, because like you have the role model, but then in between what the user does and how the role model is being used, there's like a lot that's

1:01:46 happening, that people don't see. It's not the weights, It's almost like this.

1:01:53 I mean, they index all your history using all of the stuff that you've used in the past to make it smarter and better, to write code like you. Yeah, they're trying to customize. What do you call

1:02:03 that layer though? 'Cause I'm trying to, you could call it prompt engineering, but it's - It's way more than that. Yeah, it's, I mean, it's like, well, let me - It's like this like history.

1:02:12 Yeah, I don't know what to call that layer, but I know that like, you know, Grock and all these guys when they're serving a model, like they have that layer. You know, that's why like, if you

1:02:24 download llama and you just try it, you're just like, oh, this isn't - This is mad. This is, yeah. Well, it's 'cause you don't have that layer, right? That's - Haven't changed it. Again,

1:02:34 people think it's magic because they just see the output. They see the videos online or whatever. They don't realize that it's like, oh, yeah, there's a shit ton of money the time I went into

1:02:44 developing all the things under that, like you wrap it in this box, but you don't really have any idea what's going on under the hood from - average consumer perspective. And so they're just like,

1:02:53 oh, well, they all do this. And they're like, nah, it's not like that. Well, this is where like, I've been trying to look into automated prompt engineering. So like I saw Microsoft's prompt

1:03:04 wizard project. I've seen, gosh, I can't remember that. They have such a weird name. It's like DDYS or something like that. That is not the name. But they do some automatic prompt engineering

1:03:16 where it's like automatic few shot injection where they basically do rag and this thing to automatically inject few shot examples into the prompt and stuff like that. And there's a couple of other

1:03:30 articles I've seen on this, but I just, I thought to myself, I said, okay, I know for a fact that, well, I can't say for a fact, but I'm fairly certain that all of these big labs that are

1:03:45 servicing models, they're doing this. They have some sort of massive engineering layer between the user typing something in.

1:03:52 and actually talking to the raw model that is routing the user's request, classifying the request, expanding the prompt, like looking at previous examples of another similar prompt and seeing what

1:04:04 the example is like injecting that into the prompt, right? Like all of this is happening, but it doesn't seem like there is really a open source solution for this. It very much so is like their

1:04:16 secret sauce. Yes. I feel like that's where the true tech is So the differentiation in the tech is in all of that 'cause you're right, it's like, and then you start layering agents on, and it's

1:04:26 like, okay, well now each agent needs its own system prompt, and that could be modified in real time based off what you asked it. Like, yes, it's a crazy rabbit hole, but it's one of those

1:04:38 things though that I truly believe does differentiate these companies from everybody else because it's not just a GPT wrapper. You know, even GPT itself is doing all this other stuff under the hood,

1:04:49 right? This is where,

1:04:52 So one of the things that I have never actually talked about this, I guess like outside the company was,

1:05:01 I also found that like, so I've worked with Langchain all these other things and it's kind of hard to write the agents. I feel like traditional programming languages, I guess like line-based

1:05:13 serial execution languages to me just don't feel right, writing an agent. It just feels

1:05:20 feels cumbersome. It feels square peg, right? Yeah, like, so it feels like writing UI and code. It's like a really good way to do that. Yeah, instead of a declarative language and so like

1:05:33 something that we're kind of like working on internally is sort of like a declarative language for writing agents, but that will also provide that automated engineering layer. And it's not something

1:05:48 that we're not trying to, we kind of want to make it open source, like something that where it's like, this is just an easy way to build agents that automatically serves that layer. I don't know

1:06:00 what to call it still. I don't know either.

1:06:03 It's like, you know, it's the magic layer. It's in between the raw model and, you know - You heard it here first. We're calling this the magic layer. I truly like that because it's that spot on.

1:06:14 What is this fundamental to what y'all are doing? Y'all, I mean, that's what, that's the thing Like, you've got the prompt comes in before, it was just a generic rag, right? Where a prompt

1:06:23 went in and it went and searched and returned all this stuff. Now it's routing, it's expanding. We've got agents, we've got two calls. There's, and yeah, it just gets so much more complicated.

1:06:34 That's stacked that you don't know what to call it. Yes, that's exactly right. That memory of things has done. Yeah, memory is a whole nother one. I like that. I mean, that is your application.

1:06:41 I mean, that is what, yeah, generally, that's true. So we wanna just make it easier to build that layer. Just because I find like, I think LinkedIn is a fantastic framework, but I don't like

1:06:54 TypeScript.

1:06:56 I think that there's a couple of other frameworks and a Microsoft just dropped one. I know OpenAI is dropping them, but they're all dropping coding SDKs. I'm just sitting here, I'm like, this is

1:07:08 not fun to write these agents in or build these MCP servers. Because even the

1:07:16 MCP server, you can do some prompting, and you can do some chain of thought, or I forget what the latest term for this is, but it's

1:07:27 not a tree of thought, but it's basically where you ask the same question to five different models, and then you do a council of assembly, or it's like, Okay, three of them got this answer, so

1:07:38 I'll go with that one, or, I guess, democracy or something like that. Wow, these are more things I don't even think about. These are just different patterns people are using to try and, you

1:07:47 know - Then they'll throw a judge in there too. Yeah. Another LLM that acts as a judge. Yeah, there's a lot that people have tried. I mean, I like, I love the creativity. I mean, just things

1:07:57 that I've never thought. Just things I don't think about that are must be going on in the background to give you those answers. We actually, I mean, when we first started with our rank, it had,

1:08:05 it would ask the same question across three different models and then we would have another model that would try and figure out which model had the best answer and then serve that up Yeah. Which

1:08:14 sounds really good, a complete nightmare to manage super slow, super expensive. Yeah, because you're not really not using tokens on three different models. Right, you're tripling, actually

1:08:26 quadrupling because then you also have the judge on every single prompt. It's, you know, and there's code you're getting. There's small language models. EMO, I assume. No. No, it's so, we're

1:08:33 calling it like,

1:08:40 head name for I don't know if this is is just like a declarative agent market language. And it, you know, I built it in TypeScript, just because damol, yes,

1:08:51 we're going to go with

1:08:57 damol, it's a yaml, it works too well. But it looks like HTML. So it's like a tag based language. And, and then I wrote this sort of like engineer, this is the other thing I was like, well,

1:09:06 what if it could rewrite itself, just like you have the agents rewriting the code, like can it basically see how it's done and then rewrite itself to perform that better, right? And so I did test

1:09:18 that with like the mmlu, um, the moral questions part of like the original mmlu, where I could take like, I don't know, like gbt, four, oh, many and go from, I don't know, 50 to 86, right?

1:09:31 Um, by kind of doing chain of thought and these other things. But the idea was to just have something that was open source to the community that was easy, because there are so many smart domain

1:09:45 experts out there. They know like If you sit with them and they're like, well, I need something that does this and then this and then this and then this. They're already kind of data engineers.

1:09:54 If they look at them in these - and all of them, they're already kind of doing that. I mean, look at the volume of them doing Python. Yeah. Yes. Oh my gosh, what if you could hand them, yeah,

1:10:05 that's their competitive edge, but yeah. But even at this point, I mean, I even just argue like, you should be putting cursor or something like that in their hands right now, like, 'cause like,

1:10:14 they're dangerous and I'm trying to - They don't even know, they don't even know what they don't know, they don't even know what they're missing, yeah. And when they're not gonna, you know, push

1:10:20 something to a prod, sort of like, just let them solve their own problems Yeah, they're usually just working themselves, yeah. My laptop is full of ridiculous Python scripts that no one else

1:10:30 should ever see. And what is the problem for me at one point in time? But what do they do? They all translate it to like a spreadsheet or a PowerPoint anyways, right? Like they do everything over

1:10:39 here and then ultimately it's downtown to a couple pages. It's like, almost how can you sort of like, you know, if you look on Grock, you look on Kerson, now they have these like agent things

1:10:52 where you can almost have like a background agent. You can define this. fact, I've never used it, so I don't know. But based on the documentation, you can basically define a repeating task or

1:11:01 something that will continue to execute in the background. When it filter out all the junk in your email and that's real way good stuff. Yeah. Honestly, I just don't trust any of them. That's a

1:11:13 real thing. I don't think Brian Becker told me about a really good email. Super human. Might be. Might be super human. Yeah. I did hear about them. I mean, he, it's where I get ads. I'm just

1:11:25 so terrified it's going to remove something. It's not supposed to. Right. Like how often do I have to go? If it's important, send it twice. That should be like, does it get through on the

1:11:34 second time? Like, I don't think the algorithm should be. Well, because think about all this junk mail. I don't know. Yeah. It's been filled as someone who, you know, our company used to be

1:11:45 marketing media. And so we were sending out a ton of emails and that is not a fun thing to do in the state and age with especially with corporate spam filters and all the other stuff. They replied

1:11:56 to an email and we still can't get the email through the server. Oh, what's happening here? Or you're using like a very well-established server to send those things from. And it's still not

1:12:08 getting through. Oh, gosh. They're corporate, right?

1:12:13 All the fun stuff. This has been a good combo. Yeah. I told you I was going to go fast. We're an hour and 20 now. Oh, wow. Yeah. I think it started a little later than that. Well, maybe not.

1:12:23 We had about five minutes. Well, offset Okay, see. Yeah. I know. This is when you're geeking out over. Yeah. Yeah. I love it. You want to do quick little speed around anything? Sure. So we

1:12:35 like to finish with just quick random questions that just quick and short answers. Just to give you guys some more to let people understand you guys a little bit better. So the first one will be

1:12:50 what is your favorite

1:12:54 coding language?

1:12:56 code at all. I don't know. You start out in software engineering. They're right. Yeah, I don't code anymore. I mean, I probably I live in XML. Is that even a coding language? Data format

1:13:07 language. Data format. Since that's where I live, I would say it's XML. I mean, you have to understand XML. Do you know anything? You're writing Java or like if I'm in the code code, I'll do

1:13:19 I'll be in C sharp, but I don't really I'm was a C girl. So I don't even I can read all the code. I and deep and go through a debugger. I just yeah, I just don't know what I don't know to like

1:13:30 write things anymore. There's so many people in our industry that know C and C sharp and wild. Yeah, I mean, that's mine C sharp all day. I mean, I started three five and just never looked so we

1:13:42 didn't really get into it, but I'm assuming Panascape's written on like dot net. Yeah, yeah, yeah. We have some things that are outside of it, but yeah, mostly dot net. There's a little bit of

1:13:50 math libraries that live in I think a C world or something, but otherwise everything.

1:13:58 All right, so I guess one for each of you, favorite Portuguese dish, and then favorite restaurant in Houston.

1:14:07 Favorite restaurant? Uh, Fezioada. Okay. It's not Portuguese, so it's Brazilian, um, but Fezioada, which is, um, so it's like only served on Saturdays for lunch in Brazil. And it's

1:14:21 basically like a bean meats do that they cook for, you know, like 24 hours And then it's served with rice, orange slices, uh, cauliflower kale, like shredded kale

1:14:34 and stuff like that. I mean, it's just so good. It's awesome. It's usually made from like the parts of the animal that, you know, people, yeah. But no, nowadays it's not. That's where the

1:14:45 dish originates from was like, this is how they, you know, the pigs feet and the pigs here and all that stuff. But it's just like booting. Hey, you gotta use it in our gumbo or meat. Yeah, I

1:14:54 just don't tell me how it's made. No, no, it's like they don't, you can get it in different meats or whatever, but it's fantastic. I think here, I like always trying to find good mom and pop

1:15:05 places. And now, I don't think it's Cypress, maybe it's a Houston address, it's Soto's, I don't know, Soto's Mexican Cantina. Okay. And it's really good Tex-Mex out by us, like Grant and

1:15:18 Jones. Nice You're not from the Northwest, I don't know what part of town you're in. I'm from Old Katie, I'm saying. Oh yeah, so we're out 290, yeah. Like, so anyone that way, anytime people

1:15:29 are in that area, that's my concern. I used to be a Soto's out by us when I was in Sienna, actually, but I don't know if it's the same family or not. I don't know, it's kind of like Elheimadors,

1:15:38 like there's a lot of those. Yeah, it's cornered in Houston. You never know if it's a family owner, if it's a national chain, you just never know. They're all named the same. Or like Rosas,

1:15:49 right? Yeah, big. And then there's, I guess, a couple private ones and a couple that are still own, but it gets complicated. But guys, this has been awesome. We really appreciate it. Where

1:15:60 can people get in touch with you if they want to learn more? Go to LinkedIn and go to our website, Pandascapecom. Which is pretty active. You're putting videos out pretty often. I've been. It's

1:16:10 very nerve-wracking. I don't know how you guys do this. This is a lot easier. I think doing what you're doing would be more nerve-wracking for me. Like trying to, I'd probably take 20 cuts and,

1:16:19 you know, probably just scrap it all and not post it on LinkedIn. Sometimes I think it'd be easier if I did it with a tequila shot with me. A hundred percent. I can't tell you. I don't realize

1:16:29 that. You could do that and just create your own little way. But I can't do that at 10 in the morning when I'm trying to do that. That's me. I'm Bloody Maria. There you go. Do it on a Saturday.

1:16:38 I'm a mosa. Just throw out the orange juice. Just go straight to San Francisco. Take your, take Fridays off and then just dedicate Saturdays to, you know.

1:16:47 like kind of content. We got to start. We got to start with breakfast. Yes, there we go. Get you a bloody or a mamosa and go from there. There you go. Oh, that's a fun fact. I actually

1:16:56 learned. I think mamosa means Daisy or something like that. But I don't know. I always tell people like, you know, the mamosa I don't like all the sugary things. I'm like, yeah, I want to just

1:17:07 throw out the orange juice. Yeah, I had to keep the champagne

1:17:12 My wife is champagne and then literally. Like, is like an eyedropper? Like an eyedropper, which is the essence of orange juice is what I call it. Like someone asked me how to make a good

1:17:25 margarita. I'm like, you know, you take like the the lime mix and the concert and everything and you wouldn't just throw it out and then just put tequila

1:17:36 in the glass. Don't need anything. Just keep it neat. Hey, good tequila. That's that way to drink it. There's only one way like sipping like great. Awesome. Well, thanks guys, yeah, thanks

1:17:50 for having us.