Structured Thinking for Unstructured Data: A Conversation with Founder Kirk Marple

Speaker A: Foreign.

Speaker B: This is Product by Design, a podcast by Prodity where we explore technology, artificial intelligence, user experience, product management, and the philosophy of building products and companies. All right, welcome back to another episode of Product by Design. I am Kyle, and welcome to. And this week we are joined with another awesome guest, Kirk Marple. Kirk, welcome to the show.

Speaker A: Yeah, thanks so much for having me here, Kirk.

Speaker B: Uh, it's great to have you. And let me do a brief introduction for you and then you can tell us a little bit more about yourself. But Kirk is the CEO and founder of Graphlet and also a former Microsoft General Motors and Stats leader. Kirk, I'm very, very excited to talk to you today. Um, we've got a number of things that we're going to talk about, but before we do that, why don't you tell us a little bit more about yourself.

Speaker A: Yeah, thanks. I mean, just been a career software developer, had been at Microsoft after my master's and then, um, was there a number of years in Microsoft Research and Windows Media and then started my own company for, uh, gosh, that was about 1112 years of my life and finally exited and now been working on this kind of unstructured data pipeline world for the past six or seven years. First at some other companies like GM and Stats, and now at Graphlet.

Speaker B: Awesome. Well, I'm excited to talk more about it because I think this is a very interesting subject to talk about, and obviously you probably do too, having worked in IT for so long now. But before we do that, why don't you tell us what are some of the things that you like to do outside of the office when you're not working in unstructured data and unstructured data pipelines?

Speaker A: I mean, these days there's not enough time outside of the office. But, uh, I've always liked to cook. I mean, I actually thought I was going to go to culinary school before finding computers. And so that's kind of the, uh, I don't get as much time to do it lately. But it's always, I mean, everybody's happy what I do.

Speaker B: Very nice. Well, that, that, uh, sounds really exciting. Do you have like a signature dish or kind of a go to dish that you like to do?

Speaker A: Well, the. I always remember the first one, uh, that I did in, in like high school or something was crepe suzettes. And they actually let me bring alcohol into the high school or junior high, whatever it was, and flame it. These days I probably get arrested for doing that, but it was always. Crepes are always a good One very nice.

Speaker B: Well that sounds, that sounds delightful. All right, well, great. Kirk, I want to dive into a number of questions and discussion that we have. But uh, as we do that, why don't you tell us a little bit more about your journey. You kind of touched on it a little bit. Some of the software development, uh, journey that you've done and starting uh, your own company. But tell us a little bit more about that. What brought you into what you're doing right now and what has kind of been your journey to that?

Speaker A: Yeah, I mean I kind of fell into media software per se and had a big interest in computer graphics and that was what I went to undergrad and it was like, was actually looking to see if hey, does anybody have a computer graphics degree? And then realized it's compute, you go for computer science. I mean, so it always kind of been around graphics, 3D media software. And my first, I mean one of my first jobs out of college was basically building file parsers for like JPEG or TIFF and um, even like fax compressed data. And it's kind of ironic how similar the current world is of like, I mean, building parsers, building file based workflows. So I've kind of always been around that space my whole career and did uh, at Microsoft, worked on 3D virtual worlds, built the first streaming audio control for Microsoft back when real networks were kind of competing with them. It's always been this thread in the background, but nowadays we call it unstructured data as kind of a superset really. I've just gotten a lot of reps on building these kind of platforms dealing with a lot of different media data metadata that goes with them, like the title, the author, stuff like that. These days we're integrating that with computer vision, natural language and obviously LLM. Um, so yeah, it's kind of all strung together in retrospect, but wasn't a big plan up front.

Speaker B: That feels like a career progression that I think many of us can relate to where when you look back on it, it kind of all works out, but not necessarily what you would have thought of when you first got into it. So tell us a little bit more about Graphlet. You kind of mentioned it briefly, like what is it that you do? Like what is the focus?

Speaker A: Yeah, I mean the original idea was, I mean making unstructured data as easy to work with as uh, structured data. I mean, so we've all used databases for years and people using Postgres or using Snowflake and I guess about seven years ago or So I had this idea of like, I mean, why isn't there spark for unstructured data, A snowflake for unstructured data. And I was at General Motors, um, they were taking data that was like video lidar data, telemetry data, like kind of IoT data. And we were having to build this all from scratch, this kind of pipeline to get the ML engineers, engineers to even do anything with that data. So it was a lot of ingestion and parsing and I had already had a lot of experience in the broadcast world selling to like ESPN and folks like that, building for broadcast. But a lot of those techniques just were not known in the kind of data engineering world. That was kind of why they brought me in at GM where they're like, look, I mean you already know this video stuff, we need help figuring this stuff out. And really the couple companies I was at after that as like cto, we were building pipelines for this stuff. Excuse me. And time and time again it just wasn't off the shelf. And that's what I really wanted Graphlet to be. And originally we were called unstruct data and we were kind of this, um, an unstructured data warehouse you could call it. We actually started from the application layer down selling a data cataloging tool that you could put your data into, search, visualize. And um, then with LLMs started to realize this fits really well. Like we are kind of the long term memory for what now we call rag, like chatgpt conversations. And so we really flip back to my original idea of just selling the platform. And that's kind of where we got today, where we said, look, we already have this great pipeline, now let's get it in the hands of more developers. And that's been like the last 18 months or so.

Speaker B: Okay, that's really interesting and I want to touch on a couple points that you made. Before we do that, let's like high level talk about the difference between structured and unstructured data. Because I know that some people who are listening right now may be wondering like, what is the difference between those? So high level structured data versus unstructured data.

Speaker A: I mean the simplest way I look at it is, and usually say is structured data is something that looks like a spreadsheet and I mean rows and columns, I mean the structure is essentially in that kind of tabular format. And that could lit. I mean you could call that a database, you call it a spreadsheet, you could call it a, I mean whatever data warehouse. And unstructured data is Kind of everything else, I mean and we've even done a good bit of work early on with geospatial data. And so it's unstructured in a way where I mean it's geospatial regions, it's uh, it's things like that or you could consider IoT data, um, to kind of. It sort of falls into both categories a little bit. But unstructured today, I mean it's kind of that classic media. So it's images, video, documents, even CAD drawings. But I mean documents and anything textual is now with LLMs are kind of forefront. So we actually cut our teeth more on the non documentary area early and then documents kind of were easy to plug in. I mean do more with documents and text extraction and stuff like that. So for us I think we kind of came at it from more of a media centric side rather than a document centric side.

Speaker B: Interesting. I'm kind of interested in that just because that's probably where I've gone back and forth in a lot of the work that I've done. So I spent a lot of time maybe in some of the same areas that you have where we were working with a lot of captured images. So from things like drones, things like 3D walkthroughs of buildings and things like that and taking these types of things and then having to take that and make sense of it. And so like what is it that as somebody's going through, how can we uh, either how can we take that and make it into something useful and then how can you make it like a searchable thing? So like as you're going through, like I want to find all of X things in this and so like those are the types of things that become a very, very interesting problem to solve, but also very difficult one. And then probably some of the things that are more familiar to people like taking unstructured text and documents and things like that. So it sounds like you kind of approached it from what would possibly be a more difficult. Where you're, you're taking a lot of like captures and media and things like that. How has that approach in your experience, like has that been a helpful thing and what has been I guess some of the learnings that you had going from that into more of like the documents and textual unstructured data.

Speaker A: Yeah, I mean I think the biggest value I see is kind of being more of an agnostic approach. Like we've abstracted to essentially what we call content. And so a lot of times I say we're kind of like a content management system with LLMs built in. And by content it could be anything. It could be a video, it could be an image, it could be a webpage, um, it could even just be raw text. And so I think by keeping that abstraction, it makes it really easy for us to plug things in where, I mean, somebody needs a new file format, we can map that to our existing structure. And I've always been big on canonicalization of, uh, let's find an intermediate format that we can map everything to that is consistent so all of your downstream processes just deals with that canonical format in the middle. You don't have to like, I mean, when you're talking to the LLM, I don't really care if it came from a Word doc or a PDF or a webpage track that. Hardly at all. I mean, and so the only special cases, I mean, we have like text splitters that maybe are different, like if it's code or not code, things like that. But generally it's really agnostic. And that actually I think for us is a lot of the power where we can have this massive funnel of data sources and data formats too. I mean, we don't care if it's a notion doc, if it's a file off a Google Drive, or if it's a webpage in the middle of our pipeline. It's just content and then the text part of it. I think it's what's interesting. I mean, once you get into the weeds with document extraction, it gets really complex. I mean, how do you do tables really well. And I think now what we're seeing is an integration or a, uh, move to more visualization analysis, not just ocr, because there's a lot of cases. I just had a case this week, the document had sort of a table of radio buttons and so like one was selected, but none of the typical ocr, um, really do a good job at that. But taking a snapshot, a screenshot of that page and putting it to Claude sonnet, the latest 3.5 model, actually did a really good job. And you could prompt it to extract text. And I think we're really close to a phase shift maybe from some of the older OCR kind of classic models to just using visual models for this kind of operation. And they're also more guidable that you can be like, okay. I mean, the one thing that's interesting is they don't 100% do a good job the first time. But if you give them feedback and say, look, I think you screwed up, that Radio button detection. Have another look. They really, I mean we've added something called a revision strategy where you can now ask it, prompt it over and over again to go be like, okay, here's what you just did, go back and look again. Um, and that technique works really well. And so yeah, I think we're in a really interesting transition from kind of old school to new school methods for that.

Speaker B: As you've kind of seen going along those lines, as you've kind of seen AI really taking off and coming into this probably new phase that we've seen over the recent few years, what have been some of the shifts that you've seen and what do you see going forward? You mentioned something now just where, you know, we're kind of in this transition phase. How has that changed, uh, both, you know, what you've been doing and what do you see that changing going forward with either things that you're working on or just how we handle unstructured, um, data, generally speaking.

Speaker A: Yeah, I mean I think for us what we've seen from customers is at this point in time they kind of have to build two products. They're trying to build their application for their customers and then they're having to build an infrastructure product internally to actually get their data flowing, kind of get it available to LLMs, do all the parsing, ingestion and all that kind of stuff. And what we're trying to say is look, that layer, let us do that, like we can do it better. We can scale, you pay by usage and it opens up the capabilities for companies that could have never dealt with this before. Um, and I mean especially like startups and folks that they want to work up here, but they got to build down here. And that I think is really opening up the industry. And that's really been my thought of the more we can provide tools and at that kind of lower level, it just lets people innovate and come up with new applications they couldn't before.

Speaker B: I think that's really interesting and I kind of want to touch on that a little bit more because you've, you've brought up a point of, you know, the fact that this sort of pipeline and the ability to have this ability really frees up a lot of customers and a lot of businesses to do other types of things. As you've built up this company, like what have you seen, uh, that are some of the most common use cases of employing this technology of what you're doing that really have freed up companies to do other things. And what have been Maybe some of the most common or maybe some of the most interesting things that you've seen customers doing that maybe they wouldn't have been able to do before.

Speaker A: Yeah, I think with, I guess it's probably more like 20 months now. But since ChatGPT came out I think people started to map their data to that concept. The whole chat with your PDF, chat with your files. I mean a lot of it's just, it's low hanging fruit. It's people understand how ChatGPT works and now where unstructured um, data fits into their workflows makes sense more. And I think we had been talking about it 18 months before that and nobody got it. Like very few people got it, um, the people on the computer vision side got it. But it was really something where I mean it wasn't clicking at first and now we're seeing where it's just, it's another checkbox of like hey, do you have an AI copilot or a, I mean agent workflow on my data? And so I think we're seeing those, that's kind of the low hanging fruit of what the common applications are. But then I mean we also have this knowledge graph, um, data model that we've had from day one and now it's becoming more, I mean kind of more accessible where you can build those knowledge graphs using LLMs and gather more than just terms. You can gather all the metadata for a company or all the metadata for a person like their email, their first name, last name. That was difficult. I mean honestly. I mean all the, they call them named entity recognition models were pretty much just word focused, term focused. And so that's really opened up a lot of capabilities now and now, I mean with some of the interesting ones, I mean we have folks in like the healthcare area, um, real estate construction looking at building knowledge graphs based on their domain knowledge and they can use it just for search and to be like another kind of type of database or they could use it for rag conversations, um, or both. And so I think that's some of the, I mean even just the last month I've heard people are really getting this kind of knowledge graph concept looking now, looking for it as a customer rather than not clicking. And so I think we're in a new swell of capabilities now that we're primed hopefully to target.

Speaker B: Yeah, that's great. And I wanted to touch on that idea of um, the knowledge graph, like tell us a little bit more like what is a knowledge graph and why is that such a powerful tool for someone to use.

Speaker A: I mean, really the way I look at it is, I mean, it's all about the knowledge embedded in the media. I mean, and I started looking at this from a podcasting angle originally. Like, when listening to podcasts, there's, I mean, topics discussed, there's people discussed, there's places, organizations. And how do you navigate that? Like, you could read a transcript, but can you navigate the interconnections and be like, follow that thread of, okay, we just discussed whatever real estate, Go find me other podcasts that mentioned real estate in the context of LLMs. Uh, and that becomes a search query, really. But you have to index all that data properly. And so it's. And it's more than just looking for words, it's looking for concepts. And so that's really been a big idea. I mean, I had this idea six or seven years ago of building a podcast discovery platform and kind of looking at it from, like, the music and entertainment side of it as well, of, like, bands and venues and things like that, and creating a knowledge graph around that. And that same pattern now you can apply to any data set. Um, and that's, I think, where we're really seeing the power of it. It's about the interconnection and the relationships partly as a search mechanism, as a filtering, to be like, okay, find me all the data that's in audio format in the last 90 days that's mentioned LLMs and real estate, and so that becomes a filter. But then you can also give that knowledge graph data to the LLM, and you could say, here's a list of products and here's all their metadata, like their sku, their description, their whatever. And it's not just a word per se, it's metadata, like a full entity that is now what we call graph rag. That's actually something that, I mean, we were almost doing from day one, and now it's become like a term. And that's something we're leaning into because it's really. You can give so much more color to the conversation by pulling in the kind of next layer of information around the graph.

Speaker B: I think that that's a really fascinating side of all of this and really being able to take it from kind of like you were saying, it's not just, uh, a keyword search or something like that, but it's taking this whole entity or concept and mapping out, like, all of the relationships, like, how does this tie to everything else? And. And what are the other potential entities or things that are tied to that, and what are the Relationships to that and how far can that potentially go? It's a fascinating thing that the potential for it just feels like, I don't know, mapping all of the knowledge that there is really, but within specific domains. And you've probably seen that even more, um, being somebody who's worked on it for like you were saying, six or seven years or more, um, in this specific uh, domain are there specific use cases that you've seen that have just been like, wow, that's a really cool thing that maybe somebody's done or a company's done that uh, has really kind of stood out to you.

Speaker A: I mean we just started this one in the biopharma healthcare space and this is a really exciting one because it's really going to stretch the bounds of our capabilities of what kind of data formats or, um, sorry, what kind of entity formats can we pull out? So I think that one's super exciting. Got started on that. But I think, I mean just in terms of we have other people that are pulling out product information from documents and looking at certification documents and trying to correlate, okay, is this product certified by this set of requirements? Um, and that becomes a really interesting case because you're also pulling in data that's maybe not textually similar, but might be entity similar. That, I mean that could be something. And you could even add nodes in the graph after you've pulled in the text. So this might be, I mean a human sort of like tagging that document to a specific product per se. And that's really where the value comes in of like you're creating that web of interconnections. I mean it's not new per se. This is like the Semantic Web. And I mean it's been talked about for gosh, I mean 20 years almost, but making it into production that you can actually use this through a nice API. I mean it lowers the bar for access and it's not a research project per se, it's really just another, another API that you can integrate into your applications.

Speaker B: Yeah, yeah. And you bring up a really, um, great point that I kind of wanted to touch on as well. That as you create these tools and, and make them available, you know, how, how are you thinking about the, the user experience? Uh, especially as you are, you know, a very developer focused company. One, how important is the user experience? And then how, how do you focus on that and how do you create something that really easy to use, whether that's creating these knowledge graphs or creating other tools that uh, aren't, you know, aren't just something that is, gets the job done, but is in fact like a very usable and easy thing to integrate into what people are doing.

Speaker A: Yeah, I mean we've kind of come full circle. I mean we were very end user focused. Like, I mean somebody at a port or railway was like our target customer for managing all that data. And then we flip to just be like, okay, here's an API. And now we're kind of circling back a bit. I mean we built a developer portal, I mean it looks like I know Supabase or somebody like a database company because it's like, hey, get your API key. We have billing, just the real basics we started with. But now we're adding back some of the features that we had in our original application. Like, hey, I can now see my content that I've ingested or observability. And logging is huge because as a developer tool, I mean how do you debug the problem? You're always going to run into something. And so we just released, uh, a month or two ago a logs page that, I mean you can see essentially every API call that you're making. So I think from a UX standpoint that developer experience and like what would I want as a developer to debug this application? And we got some early feedback on that and it was just something where we're like, okay, we're going to have to get this in there. But it's also trying to balance, I mean front end and backend development resources and all that kind of stuff. But yeah, we've been focused more lately on observability, seeing your data. But what we'd like to do too is it helps with onboarding, 20 minutes that someone's um, signed up right now. We can show them sample apps, we can show them, I mean the code for it, we can show them documentation. But what we're finding is people just want to kick the tires in that first like 15, 20 minutes, like upload a file, see what it output, ask it a question, like a chatgpt kind of experience within our developer portal. Originally I wasn't really thinking that maybe that would be a requirement, but I think now the more we see, um, people just want to have it click of uh, what does this do? How does it work? Um, and then people can start writing code against it. And we saw a big uptick in usage when we released our sample apps. We just used Streamlit, created a bunch of sample apps that talked to our API, um, and our SDK and that was a big win. We Got a lot more usage just by people actually seeing it in the wild, kind of like actually using it. Um, and it was a good teaching tool for us. So that's been a lot of our focus is closing that gap on kind of usability for onboarding as well. Through the ux.

Speaker B: How do you gather that type of feedback and maybe prioritize what are some of the most important things? Because obviously those are the questions that I know so many of us are constantly faced with.

Speaker A: Uh,

Speaker B: how do we get the feedback? And then how do we think about what is the most important things for us to be focused on? So how do you think about that within your company?

Speaker A: Yeah, no, I mean, it's always something. I've tried to get more feedback. I mean, I've tried, um, sending out a survey after we have a tickler email and there's a survey in one of them, we probably got four responses. But we get good. I mean, some of the feedback's good. I mean, it's right in line with, hey, better, quick start or something like that. Just get up and running in that first 15 minutes. But I think also, I mean, we have a discord that people can ask questions on. We can do, I mean, just gather information from that. And I think just, I mean, the good that we've, we've seen by, um, now people can sign up, just, uh, schedule a meeting with me and, and just on our website, grab 15 minutes, 30 minutes. And that has been awesome. Like, I just had one today. They just signed up, they wanted to learn a little more. Zero friction, set up a call the next morning. We talk it through, they understand it, and I'd be happy to take those calls all day. And so that ends up feeding back into, I mean, what are their hotspots like, what are the things they really care about? Um, they can ask questions about pricing, they can get a lot more detail. And then we can also. I mean, it drives marketing as it also drives product. Um, because understanding, like, how did they find us? And all that, we can get a lot of details from all that. So it's. I mean, it's so important. I mean, you can't do this in a vacuum, as you know. So it's. I love taking those kind of calls.

Speaker B: Yeah, I think that that's great feedback and being continually close to the actual users and people who are both using the product, new customers and things like that. It can be so easy sometimes to become separated. And I know that that's a common thing that I see and that we talk about all the time is not becoming too distant from the people who are actually using the product. And how do we close the gap if that's become the case? So I, uh, think that's, that's great

Speaker A: feedback and I think, I mean this is, there's such a diverse set of things people could do with the platform. Where my old company, it was, it was video transcoding. I mean it's basically like much more like everybody kind of does the same thing. Like we were helping all the companies, there were some, a little bit of differences, maybe like what hardware they're using. But it was a much more obvious like solution that everybody needed. But here, I mean every day I hear somebody being like, oh, I want to use it for this, I want to use it for that or I already built this as an mvp. But we realize how we now we need like a production version of that. Um, so the diversity can be good and bad, but I think for us we want to stay more of a horizontal platform but still try to figure out those places where we can help people in their vertical apps as well.

Speaker B: Yeah, uh, that's great feedback as well. And how do you think about that? Because I'm interested. Because I feel like often it can be very easy to get feedback from customers or users and say like, okay, um, we're hearing this and now we're going to become much more focused in a specific area or just for this customer. Anybody listening has probably had this exact experience where it's like, hey, this customer wants something very, very specific that if we develop, uh, we can win this business or something like that. And like you just mentioned wanting to stay like very horizontal while still supporting customers in like what they need vertically. How are you balancing that? Not getting necessarily too focused on specific customer verticals while still being able to support a wide variety of use cases?

Speaker A: No, it's an interesting point and I think, I mean I dealt with this in my last company too where we had one customer who ended up just being like a whale of a customer. And it was like something that, ah, I mean, ate up a huge amount of our time for like a year. But it was an incredible customer. I mean it was very lucrative. But it can really, I mean that was the kind of downside where you end up being almost like a one customer company for a while because there's just so much they need. And even like they were supposed to do their own internal tech support but that never really happened. And so we ended up doing that, which ate into the time. But now, I mean I think it's really, I mean, for me we want to be data model first. Like, if we can have a data model that works for any vertical, then we're halfway there at least. And that's really what I try and focus on, is if people, the diversity is in the data and the diversity is in the schema of the metadata that they want to pull out in the knowledge graph. And then maybe the diversity in kind of how they want to access the API, like can we support Python as well as Typescript or whatever. But I mean really the only other diversity there is the models themselves. That if somebody wants to fine tune a model or they want to use a model on a different provider, we want to make that available to them. And so we kind of. I always think of it as like this, really. The tree trunk's very solid, but there's a lot of branches off of it and you can follow the tree trunk and not even get into any of the edge cases, the configuration or whatever, and it just works. But then if they want to be like, oh, we want to use anthropic with our own API key and do some m custom stuff, be like, okay, cool. It's configuration, it's not rewriting your whole platform.

Speaker B: I'm interested in. As you've built this company, as you've kind of gone through this experience, what have been some of the things that have surprised you most, especially as a founder and leading, uh, your current company?

Speaker A: I mean, I think the thing you always think you understand the timing of the market and it's very rare to really know it unless you're looking back in retrospect. And so I thought we were a little early when we got started and that the market would catch up and everybody would be like clicking in on unstructured data. And it really took like probably almost two and a half years for it to become commonplace. Probably like a year and a half longer than I thought. And I mean, and that's what I mean. We got started right, I guess, right after Covid, I guess so. I mean, it could have. I mean, the market changed a little bit, but I think that's a thing where, I mean, I've never really wavered in the vision for the value here. It's really just okay, is the market ready for it? Are the customers available? Can we make it cost effective? And those kind of things. And so, I mean, there's been market maps and VC kind of things that have come out in the last two or three months that look almost exactly like our original Pitch deck of the value proposition of unstructured data. Unstructured data platforms. We were using that term three and a half years ago and nobody got it. And so I think that's a hard thing where you have to sort of keep grinding, keep grinding, keep grinding. And for us, I mean we want to be a foundational technology. It's something where we want to be the snowflake of unstructured data. Like just commonplace. That uh, here's, I mean, put your data in there, build on it and it just works. And so I think, I mean that really was, I mean even though we kind of took a foray into uh, application development more than platform, we really circle back to the original vision, but I think we may start offering more applications on the platform as well. Um, I didn't want to cannibalize the opportunity for developers to build on it, but I think there's still value in extending our developer platform for some of the features that we had before for data visualization.

Speaker B: How important for you is that, that vision and really being able to have it and articulate it and stick to it? Because obviously it was something that you have seen and you've been able to like articulate and have, for, for a long time. What, what was that? What was the importance of that early on and how important has it been for you going forward and not just now, but like going into the future as well?

Speaker A: I mean, I would just say it's super important. I mean it's probably the number one thing of, I mean, I, this, I mean, on the side. And then it turned into like a, I mean from a side project into a funded company. But you also have to realize not everybody else has the same vision you do. I mean you have to build a team. And I mean my obsession with the problem is going to be greater and then the team, the team around you and you try and warm everybody up to it. But it's, I mean for me it's such a passion because I just see it's such an untapped thing. Um, and I think the capabilities long term are huge. I mean there's, we've talked to companies that have like 20 years of data just sitting unused and I mean if we could get the data velocity up and actually leverage that through unstructured data processing through language models, I mean the ability to go back and ask questions or summarize old data, see trends, I think, I mean we're still untapped in so many ways, but it's always kind of have to like dig The, I mean dig it out of the ground first before you can really start, start doing anything with it. But I think yeah, for me it's, it's the vision is so key and I mean, but also having a solid vision for how you build it, I mean it's not just about what you're trying to build. It's, I mean, future proofing a little bit. Like I've, I've never been one to kind of like just knock out a crappy MVP and move and try and build from there. I've kind of, maybe to my detriment, but I think it's, it's now been a positive build for scale. I mean at least build for conceptual scale. Like start thinking about architecture or start thinking about how could this scale even if you don't do everything day one. And I think that's benefited us a lot where I mean the way we're built on Azure technologies, we're leveraging. I mean we're not running into the same problems. Maybe some other folks are. I mean we have problems, uh, I mean obviously there's always going to be issues, but some of those foundational ideas haven't really changed, I mean in how the product was built.

Speaker B: That's really interesting and I want to kind of touch on a couple things, but I'd love your take on some of the thought you had in building for scale or building maybe thoughtfully from the start. Obviously that probably comes with a trade off of maybe building a little bit slower. What was your thought process with that in wanting to maybe build a little bit more quality or build maybe more thoughtfully or how would you think about that versus building an MVP that maybe isn't quite as good but probably could have been done maybe a little more quickly?

Speaker A: Yeah, I mean on a good note, I mean I had several years of just noodling on the concept on the side, but I knew that for my old company it was all on premise. I mean we were selling to companies, they had literal servers in their data center that you could touch. And I wanted to build a cloud native media management tool. And that was my original goal is take what I'd learned from my previous company and build a cloud native media. And that's a lot of what really drove this architecture. The concept of content is kind of, I mean content management, media management, they're all pretty similar. But I think also in another case it's like, I mean the trade off like you said, I mean we don't support on premise deployments today. Like we only support cloud deployment we are getting interest now in private cloud where we could like run, we could deploy into your Azure subscription so you wouldn't have to pay by usage, you just pay for the metal. But I mean there's companies that are like no, we only want it on prem and um, at this point I just have to be like, sorry, that's not us, which sucks. I mean you want to have every customer, but that's a trade off you really have to make is, I mean to do what we do well, there's going to be customers we can't touch and we're not going to, we're not, I'm just going to try and fudge it, say we can. And so I think sticking to your guns on that and then finding a middle ground like we've now said, like, okay, by Q1 next year we'll have private cloud deployments and this now, I mean it will be in the Azure marketplace and all that. And that, I mean that carved off another big chunk of potential customers.

Speaker B: Yeah, for those uh, customers or companies that maybe are looking at, like you said, just tons of untapped data or possibilities like, like how would you begin to approach something like that? Or what would you tell them? Like where would they start? Or where should they start thinking about like we just have all this unstructured data and maybe we don't know what to do with it or is there untapped potential here? Like what should they do?

Speaker A: I mean I think that I always tell people to start, like understand where your data is. I mean, what are your sources? Is it on Google Drive? Is it in an S3 bucket? Is it actually just your internal web pages? I mean, or something you want to scrape? Start, start understanding the locations, understand the volume. I mean are we talking about 10,000 PDFs, a million PDFs? I mean is it video, is it audio, Is it a mix? Um, start with that kind of stuff and then really like, and then what do you want to do with it? Do you just want to make it searchable? Do you really want to have like, I mean auto summarization of that data? Uh, do you want to have rag conversations with it? Just kind of put together all the pieces of what do you want to do with it? We always try and break it up to, there's kind of that first step of the funnel of ingest from wherever it is, prepare it into a format that we can use downstream and so we focus on, let's not even worry about how you want to query it, what you want to do with it, where is it? How do we even get our hands on it? Do we need, like, is it authenticated? Like just all those questions, that's kind of where we start. And then once we understand their volume, we can be like, okay, I mean, that's 100 PDFs, it's not a big deal. But if they have like 100,000 PDFs that they want to ingest over a week, then they're going to have to think about scale. And cost is really the other variable that, um, a lot of people, they know they want the outcome, but it's not even our costs that end up being the big thing. It's like the LLM, their OpenAI tokens end up being 80% of their cost. So we have to be really cognizant or they just can't throw, take the kitchen sink at it and just do everything or they're going to eat like five grand in a day. Um, of cost that I think is an awareness and it goes back to the developer experience of making sure you have good visibility to costs. And that's something we're still working on, is like you can see your invoice, you can see credits used, but getting better graphing capabilities and search on your logs and costs and is something we're still working towards because it's critical for the larger customers.

Speaker B: Yeah, definitely. For somebody who is thinking about getting into either engineering or potentially even starting their own company, like wants to found a company, what, uh, advice would you have for somebody who's thinking about either of those things?

Speaker A: Well, it's interesting. I mean, I can say my son is just, he kind of was a Covid high school graduate, so kind of took a couple years of like figuring out what he wanted to do and now he's settled back into computer science. Um, honestly. And so it's, I mean, as somebody starting to start that, I think it's a really interesting time because it's like kids that got cell phones and their life was, they never knew life before an iPhone. And now I think with AI and the ability to code with AI, uh, it's a completely different ballgame. And so I'm actually really excited for him to see like, okay, what does it mean to learn computer science as an AI native developer? And so I'm not one of those people that thinks that AI is going to take our jobs. I'm more of a old school where it's like, it's just going to lift everybody up. Like there's going to be stuff we don't do and there's going to be stuff like, I mean it's like we have autocomplete in our IDE that helps us code. I mean it's just going to be continual extensions of that that augment us. Um, because I think, and I think there's going to be a pushback towards architecture. Like how do I plan this? And I think product management, it will even take a higher value because AI is not going to know how to build the product. AI might know how to do the syntax and the semantics but uh, it's not going to know how to think about the user and UX and all this kind of stuff. So I'm actually really interested. I think it'll be a really interesting generation. I think for the other half of that is like starting a company. I think it's, I mean it's something that there is really nothing like, I mean the ability to have the touch with customers and to know that you can solve their problems by just sitting down and writing code for a couple hours or I mean you can actually make money and like do this and you're just so much more tied into the day to day. But there's a huge downside too. It's huge amount of risk. I mean you're doing support at three in the morning. Um, I mean you're trying to, I mean make customers happy and sometimes they just won't be happy. But I mean our round of customers now is great. I mean they're all, I mean, looking forward to the future, how they can integrate this stuff. But I mean I've done, I mean where it's like wake up at four in the morning, a call back at my old company and try and do support over the phone. I mean roll it a bit and it's, I mean that's the life that you're signing up for too. So it's uh, definitely both sides of it.

Speaker B: Yeah, that's definitely true. Well Kirk, this has been an amazing conversation. I think we've touched on a whole bunch of different things from the technical side all the way to the product and UX side, to the founding and maintaining a company side. Where can people find out more about you, about the things that you're working on, uh, about anything else?

Speaker A: Yeah, Graphlet. We're just Graphlet.com. um, I'm on Twitter as myself just Kirkmarple and Graphlet's there as well. On our website we have a bunch of blogs and kind of use case stuff that you can check out and it's free to sign up. I um, don't know if I said that before, but we have a free tier that, I mean pretty much any hobby plan would fit into that people can just kick the tires and try it out. Um, and yeah, I mean we're actually working to get um, more of a YouTube channel set up. Get some of my talks and things like that. Um, and then also get some, just more tutorials and kind of how to use the platform. So hopefully over the next month or two we'll have more of that stuff up on, on YouTube as well. Okay.

Speaker B: We will put all of the links in the show notes as well. Uh, it's a great site. You've got some great blog posts I've been reading through as well. And you, you do a lot of talking about this which is very, very educational. So I will look Forward to the YouTube channel as well to learn, to keep learning more. So we'll put all the links in the show notes for anybody to check out. Um, well, Kirk, we've got a couple of uh, wrap up questions that we want to, to wrap this up with. Um, as we finish up here. Have you read or watched or listened to anything recently that you would like to share? Now these of course do not have to be product or techno technical or business related in any way, but they can be if you want.

Speaker A: Yeah, I mean, um, obviously I work a lot, but the only kind of real break that I get, my kids are all kind of grown now. So the main break on TV I've been watching is all the game warden shows. There's like all the, like Lone Star Law and Northwest Law and all the ones about the game wardens. And it's so funny just to have a new diverse related area and learn all about kind of how they're, they're managing the wildlife and doing all that kind of stuff. So that's been my sort of secret sort of before bed thing that I've been watching a lot of and definitely learning a lot about stuff that uh, that I knew nothing about beforehand.

Speaker B: Very nice. Awesome. And then final question. Are there any products that you've been using and enjoying? They can be digital products or physical products, anything like that.

Speaker A: Yeah, I mean it's, I mean we love LINEAR for product management here, I think. I mean I've used JIRA and other tools before, but I think linear has become such a key part of our structure. Even started using it just for operations, like business operations and just the, I mean we use Slack as well. But I think that just being able to be organized is key. I think there's still underutilized knowledge management with LLMs. Um, I mean if I had a month I would try and build a product like that on top of our platform that I keep wanting to get back to to kind of manage all of that data better. But I know there's some other vendors that are, that are targeting those kind of things as well. But uh, yeah, I mean knowledge management's always an untapped kind of unfinished business.

Speaker B: It is, yeah. That's a great call out. Um, okay, well those are some great shout outs for end with. But Kirk, this has been a great, great conversation. Appreciate all of the insights that you had and appreciate the conversation as well.

Speaker A: Yeah, thanks so much. This is a lot of fun.

Speaker B: It was. And appreciate everyone for listening. Thanks again for listening. If you like the show, be sure to follow or subscribe on your favorite podcast app. You can follow the show on TikTok at Proddity Co and on Twitter Roddity Co. You can also follow me on both of those platforms. Kyolarryevans. If you want more product conversation, check out my my newsletter, Proddity Proddity Co. You can also follow me on Medium at kylerievans or check out my Medium publication, Prodity. Of course you can check out all these links in the show notes.