Kyle (00:00)
All right, welcome back to another episode of Product by Design. am Kyle. And this week we are joined with another awesome guest, Kirk Marple. Kirk, welcome to the show.
Kirk Marple (00:11)
Yeah, thanks so much for having me
Kyle (00:13)
Kirk, 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 Graflit and also a former Microsoft General Motors and stats leader. Kirk, I'm very, very excited to talk to you today. 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
Kirk Marple (00:35)
Yeah, thanks. mean just been a career software developer had been at Microsoft after my masters and then was there a number of years in Microsoft research and Windows media and then started my own company for Gosh, that's about 11 12 years of my life and finally exited and now been working on this kind of unstructured data pipeline world for the past I mean six or seven years I mean first at some other companies like GM and stats now at at graph
Kyle (01:03)
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?
Kirk Marple (01:23)
I mean these days there's not enough time outside of the office, but 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, I don't get as much time to do it lately, it's always, mean everybody's happy when I do.
Kyle (01:40)
Very nice. Well, that sounds really exciting. Do you have like a signature dish or kind of a go -to dish that you like to
Kirk Marple (01:47)
Well, I always remember the first one that I did in high school or something was crepes 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. it's always crepes are always a good one.
Kyle (02:05)
Very nice. Well, that sounds delightful. All right. Well, Kirk, I want to dive into a number of questions and discussion that we have, but 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 journey that you've done and starting 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
Kirk Marple (02:34)
Yeah, I 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 computer, you go for computer science. mean, so it always kind of been around graphics, 3D, media software. my first, I mean, one of my first jobs at a college was basically building file parsers for
like JPEG or TIFF and even like fax compressed data. And it's kind of ironic how similar the current world is of like, mean, building parsers, building file -based workflows. So I've kind of always been around that space my whole career and did a, Microsoft worked on 3D virtual worlds, built like the first streaming audio control for Microsoft back when like real networks was, we were kind of competing with them.
It's always been this thread in the background, but nowadays, we call it unstructured data as a superset. Really, I've just gotten a lot of reps on building these 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 LM's. So yeah, it's kind of all strung together in retrospect, but wasn't a big plan upfront.
Kyle (04:02)
Yeah, 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 Graflit. You kind of mentioned it briefly, like what is it that you do? Like what is the focus and what does Graflit specifically focus itself
Kirk Marple (04:31)
I mean, the original idea was, I mean, making unstructured data is easy to work with as structured data. 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 a spark for unstructured data, a snowflake for unstructured data? And I was at General Motors, they were taking data that was like video, LiDAR data.
telemetry data, kind of IoT data. And we're having to build this all from scratch, this kind of pipeline to get the ML engineers to even do anything with that data. So 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 broadcasts. But a lot of those techniques just were not known in the kind of data engineering world.
That was why they brought me in at GM where they're like, look, you already know this video stuff. We need help figuring this stuff. Really the couple of companies I was at after that as CTO, were building pipelines for this stuff. Time and time again, it wasn't off the shelf. That's what I really wanted Graphlet to be. Originally, we were called unstructed and we were 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 then with LLMs, started to realize this fits really well. Like we are kind of a long -term memory for what now we call RAG, like Chat DPT Conversations. And so we really flipped 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
Kyle (06:30)
Okay. That's really interesting. And I want to touch on a couple of points that you made. Before we do that, let's 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 what is the difference between those. So high level structured data versus unstructured data.
Kirk Marple (06:51)
mean, the simplest way I look at it is, and usually say is structured data is something that looks like a spreadsheet. I mean, rows and columns, mean, the structure is essentially in that kind of tabular format. And that could live, 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 things like that, or you could consider IoT data 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 3D 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 -document.
area early and then documents were easy to plug in. mean, do more with documents and text extraction and stuff like that. For us, think we came at it from more of a media -centric side rather than a document -centric side.
Kyle (08:05)
interesting. And 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
And so what is it that as somebody's going through, how can we take that and make it into something useful? And then how can you make it a searchable thing? So as you're going through, I want to find all of X things in this. And so 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. it sounds like you kind of approached it from what would possibly be a more difficult, where 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 the, I guess, some of the learnings that you had going from that into more of like the documents and textual unstructured data?
Kirk Marple (09:33)
Yeah, I mean, 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 web page, 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
where somebody needs a new file format, we can map that to our existing structure. I've always been big on canonicalization of, 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. When you're talking to the LM, I don't really care if it came from a Word doc or a PDF or a web page.
track that hardly at all. And so the only special cases, we have tech 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. We don't care if it's a Notion doc, if it's a file off a Google Drive, or if it's a web page.
in the middle of our pipeline, it's just content. And then the text part of it, I think it's what's interesting. 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 move to more visual 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 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 guideable 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 are 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.
And that technique works really well. so, see, I think there's a really interesting transition from kind of old school to new school methods for
Kyle (12:34)
Yeah, that's really interesting. 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 were kind of in this transition phase,
How has that changed both what you've been doing and what do see that changing going forward with either things that you're working on or just how we handle unstructured data generally speaking?
Kirk Marple (13:17)
Yeah, I mean, 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 LMS, do all the parsing and gestion 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 us before. And I mean, especially like startups and folks that they want to work up here, but they got to build down here. And that, think, is really opening up the industry. 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.
Kyle (14:13)
I think that's really interesting. And I kind of want to touch on that a little bit more because you've brought up a point of 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, what have you seen 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?
Kirk Marple (15:09)
Yeah, I think when with I guess it's probably more like 20 months now, but since chat GPT came out, I think people started to map their data to that concept. The whole chat with your PDF, chat with your files. And I mean, a lot of it's just it's low hanging fruit. It's people understand how chat GPT works. And now where unstructured 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.
The people at 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, I mean, agent workflow on my data? And so I think we're seeing those, that's kind of the low -hang fruit of what the common applications are. but then, I mean, we also have this knowledge graph data model that we've had from day one.
and now it's becoming 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. mean, honestly, 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, mean, with some of the interesting ones, I we have folks in like the healthcare area, 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 can use it for rank conversations 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 knowledge graph concept, now looking for it as a customer, rather than it not clicking. And so I think we're in a new swell of capabilities now that we're primed, hopefully, target.
Kyle (17:18)
Yeah, that's great. And I wanted to touch on that idea of the Knowledge Graph. Tell us a little bit more. What is a Knowledge Graph and why is that such a powerful tool for someone to use?
Kirk Marple (17:30)
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. 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 can 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. And that becomes search query, really. But you have to index all that data properly. so it's more than just looking for words, it's looking for concepts. And so that's really been a big idea. I 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.
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. and that's, 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, OK, find me all the data that's in audio format in the last 90 days that's mentioned, LMS and real estate. And so that becomes a filter.
But then you can also give that knowledge graph data to the LLM. You can say, here's a list of products and here's all their metadata, like their SKU, their description, their whatever. It's not just a word per se, it's metadata, like a full entity. That is now what we call graph reg. That's actually something that 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.
Kyle (19:36)
Yeah. 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 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 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 being somebody who's worked on it for, like you were saying, six or seven years or more in this specific 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 has really kind of stood out to
Kirk Marple (20:46)
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 formats can we pull out. So I think that one's super exciting, just got started on that. But I think, 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? And that becomes a really interesting case because you're also pulling in data that's maybe not textually similar, but it might be entity similar. that could be something, and you can even add nodes in the graph after you've pulled in the text. So this might be, I mean, a human is 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 20 years almost. 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 API that you can integrate into your applications.
Kyle (22:12)
Yeah, yeah. And you bring up a really great point that I kind of wanted to touch on as well, that as you create these tools and make them available, how are you thinking about the user experience? Especially as you are a very developer -focused company, one, how important is the user experience? And then 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 aren't just something that gets the job done, but is in fact a very usable and easy thing to integrate into what people are
Kirk Marple (22:59)
Yeah, we've kind of come full circle. We were very end user focused. mean, somebody at a port or a railway was like our target customer for managing all that data. And then we flipped to just be like, OK, here's an API. And now we're kind of circling back a bit. I we built a developer portal. I mean, it looks like in SuperBase, 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, mean, how do you debug the problem? You're always going to run into something. And so we just released 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 what would I want as a developer to debug this application. 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 front -end and back -end development resources and all that 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 signed up, right now we can show them sample apps, can show them 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 15, 20 minutes. Upload a file, see what it output, ask it a question, like a chat DPT 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 people just want
have it click of, what does this do? How does it work? 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 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. And it was good teaching tool for us. So that's been a lot of our focuses.
closing that gap on usability for onboarding as well through the UX.
Kyle (25:33)
Yeah, that makes a lot of sense. And I'm interested, 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. 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?
Kirk Marple (26:02)
Yeah, no, mean, it's always something we're trying to get more feedback. I mean, I've tried sending out a survey, after, like we have a tickler email and there's a survey in one of them, we probably got like four responses. But like, mean, we get good, 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 we've seen by now people can sign up and just schedule a meeting with me and just on our website, grab 15 minutes, 30 minutes. And that has been awesome. Like we 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
And that is feeding back into, I mean, what are their hotspots? Like, what are the things they really care about? 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. Because understanding like, how did they find us? And all that, we can get a lot of details from all that. yeah, it's, I mean, it's so important. I mean, you can't do this in a vacuum, you know, so it's, I love taking those kind of calls.
Kyle (27:20)
Yeah.
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 then 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 think that's great feedback.
Kirk Marple (28:02)
Yeah.
And I think, 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 every day, I mean, I hear somebody being like, 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 now we need a production version of that. 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.
Kyle (28:51)
Yeah, 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, 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, we can win this business or something like that. And like you just mentioned, wanting to stay very horizontal while still supporting customers in 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.
Kirk Marple (29:36)
Yeah.
Now, it's an interesting point. And I think, I 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, I mean, ate up a huge amount of our time for like a year, but it was an incredible customer. I 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, 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 like if people, the diversity is in the data and the diversity is in the schema of the metadata that they want to pull out and the knowledge graph. And then maybe the diversity is
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 light, 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, always think of it as like this, really the tree trunks 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, we want to use Anthropic with our own API key and do some custom stuff, we'll be like, OK, cool. It's configuration. It's not rewriting your whole platform.
Kyle (31:29)
Yeah, yeah, definitely can understand that. 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 your current company?
Kirk Marple (31:52)
And 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. so, I I thought we were a little early when we got started and that the market would catch up and everybody would be 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 why 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 kinds 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, like we were using that term three and a half years ago and nobody got it. And so I think that's, I mean, a hard thing where you have to like sort of keep grinding, keep grinding, keep grinding. And for us, I mean, we want to be a foundational technology. Like it's something where we want to be the snowflake of unstructured data, like just commonplace that here's, mean, put your data in there, build on
and it just works. And so I think, I mean, that really was, I mean, even though we kind of took a foray into 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. 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.
Kyle (33:50)
Yeah. How important for you is 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 a long time. 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?
Kirk Marple (34:19)
I I would just say it's super important. I mean, it's probably the number one thing of, I mean, I got started doing this, I mean, on the side and then it turned into like a, I mean, from a side project to do 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 gonna be greater, I than 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. And I think the capabilities long -term are huge. 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, mean, dig it out of the ground first before you can really start doing anything with it. but I think yet for me, it's the vision is so key and I mean, 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, mean, future -proofing a little bit. Like 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
maybe to my detriment, but I think it's now been a positive. Build for scale. mean, 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, the technologies we're leveraging, I mean, we're not running into the same problems maybe some other folks are. I we have problems. I mean, obviously we have, there's always going to be issues, but...
some of those foundational ideas haven't really changed, I mean, in how the product was
Kyle (36:21)
That's really interesting. And I want to kind of touch on a couple of 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.
Kirk Marple (36:54)
Mm
Yeah, I on a good note, I mean, 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 we were selling to companies, 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 man. And that's a lot of what really drove this architecture and the concept.
of content is kind of, mean, content management, media management, they're all pretty similar. But I think also in another case, it's like, I mean, they're the trade off, like you said, I mean, we 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 get deployed 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. 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 and say we can. And so I think, think sticking to your guns on that and then finding a middle ground, like we've now said, like, okay, we'll.
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, mean, that carved off another big chunk of potential customers.
Kyle (38:41)
Yeah. For those customers or companies that maybe are looking at, like you said, just tons of untapped data or possibilities, how would you begin to approach something like that? Or what would you tell them? Where would they start or where should they start thinking about, we just have all this unstructured data and maybe we don't know what to do with it, or is there untapped potential here? What should they do?
Kirk Marple (39:10)
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? But isn't it? Is it an S3 bucket? Is it actually just your internal web pages? I mean, or something you want to scrape? Start understanding the locations, understand the volume. I mean, are we talking about 10 ,000 PDFs, a million PDFs? mean, is it video? Is it audio? Is it a mix? Start with that kind of stuff. And then really like, and then what do you want to do with
Do you just want to make it searchable? do you really want to have auto summarization of that data? Do you want to have rag conversations with it? Just put together all the pieces of what do you want to do with it? We always try and break it up to, there's that first step of the funnel of ingest from wherever it is, prepare it into a format that we can use downstream. 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, that's not a big deal. But if they have like 100 ,000 PDFs that they want to ingest over a week, then they're gonna have to think about scale and cost is really the other variable that a lot of people, they're like, 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. It's their open AI tokens end up being 80 % of their costs. And so we have to be really cognizant where they just can't throw like the kitchen sink at it and just do everything or they're going to eat like five grand in a day of cost. And 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
You can see your invoice, can see credits you use, getting better graphing capabilities and search on your logs and costs is something we're still working towards because it's critical for the larger customers.
Kyle (41:18)
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 advice would you have for somebody who's thinking about either of those things?
Kirk Marple (41:35)
Well, can say my son is just he kind of was a COVID high school graduate. So kind of took a couple of years of like figuring out what do you want to do? And now he's settled back into computer science, honestly. And so it's I mean, as somebody starting to start that, think it's a really interesting time because it's like people 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, it's a completely different
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 auto -complete in our IDE that helps us code. I mean, it's just going to be extension, continual extensions of that, that augment us.
Because I think there's going to be a push back 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 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 company. I think it's I mean, it's 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 of 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 to its huge amount of risk.
I mean, you're doing support at three in the morning. I mean, you're trying to make customers happy. And sometimes they just won't be happy. I mean, our round of customers now is great. they're all looking forward to the future, how they can integrate this stuff. But I mean, I've done, mean, where it's like wake up at four in the morning on a call back in my old company and try to do support over the phone, I mean, rolling out of bed. And it's, mean,
That's the life that you're signing up for too. So it's definitely both sides of it.
Kyle (44:07)
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, about anything else?
Kirk Marple (44:29)
Yeah, Graflit. We're just Graflit .com. I'm on Twitter as myself, just Kirk Marple, and Graflit'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 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. yeah, I mean, we're actually working to get more of a YouTube channel set up.
Get some of my talks and things like that and and then also get some just more tutorials and kind how to use the platform So hopefully over the next month or two, we'll have more of that stuff up on YouTube as
Kyle (45:07)
Okay, we will put all of the links in the show notes as well. It's a great site. You've got some great blog posts I've been reading through as well. And 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 keep learning more. So we'll put all the links in the show notes for anybody to check out. Well, Kirk, we've got a couple of wrap up questions that we want to wrap this up with 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 technical or business related in any way, but they can be if you
Kirk Marple (45:47)
Yeah, obviously I work a lot, but the only kind of real break that I get, my kids are all kind of grown now. 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 unrelated area and learn all about kind of how they're managing the wildlife and doing all that kind of stuff. So that's been my sort of secret, sort
before bed thing that I've been watching a lot of and definitely learning a lot about stuff that I knew nothing about beforehand.
Kyle (46:23)
Very nice. Awesome. And then final question. Are there any products that you've been using and enjoying? You can be digital products or physical products, anything like that.
Kirk Marple (46:35)
Yeah, I mean, it's I mean, we love linear for product management here. I think I mean, I've used your 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 it just the I we use Slack as well. But I think that just being able to be organized is key.
I think we're still underutilized knowledge management with LLMs. 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 you to kind of manage all of that data better. But I know there's some other vendors that are targeting those kinds of things as well. yeah, mean, knowledge management is always an untapped kind of unfinished business.
Kyle (47:20)
It is, yeah, that's a great call out. Okay, well, are some great shout outs to end with. But Kirk, this has been a great conversation. Appreciate all of the insights that you had and appreciate the conversation as well.
Kirk Marple (47:37)
Yeah, thanks so much, this was lot of fun.
Kyle (47:39)
It was, and I appreciate everyone for listening.
All right, welcome back to another episode of Product by Design. am Kyle. And this week we are joined with another awesome guest, Kirk Marple. Kirk, welcome to the show.
Kirk Marple (00:11)
Yeah, thanks so much for having me
Kyle (00:13)
Kirk, 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 Graflit and also a former Microsoft General Motors and stats leader. Kirk, I'm very, very excited to talk to you today. 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
Kirk Marple (00:35)
Yeah, thanks. mean just been a career software developer had been at Microsoft after my masters and then was there a number of years in Microsoft research and Windows media and then started my own company for Gosh, that's about 11 12 years of my life and finally exited and now been working on this kind of unstructured data pipeline world for the past I mean six or seven years I mean first at some other companies like GM and stats now at at graph
Kyle (01:03)
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?
Kirk Marple (01:23)
I mean these days there's not enough time outside of the office, but 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, I don't get as much time to do it lately, it's always, mean everybody's happy when I do.
Kyle (01:40)
Very nice. Well, that sounds really exciting. Do you have like a signature dish or kind of a go -to dish that you like to
Kirk Marple (01:47)
Well, I always remember the first one that I did in high school or something was crepes 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. it's always crepes are always a good one.
Kyle (02:05)
Very nice. Well, that sounds delightful. All right. Well, Kirk, I want to dive into a number of questions and discussion that we have, but 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 journey that you've done and starting 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
Kirk Marple (02:34)
Yeah, I 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 computer, you go for computer science. mean, so it always kind of been around graphics, 3D, media software. my first, I mean, one of my first jobs at a college was basically building file parsers for
like JPEG or TIFF and even like fax compressed data. And it's kind of ironic how similar the current world is of like, mean, building parsers, building file -based workflows. So I've kind of always been around that space my whole career and did a, Microsoft worked on 3D virtual worlds, built like the first streaming audio control for Microsoft back when like real networks was, we were kind of competing with them.
It's always been this thread in the background, but nowadays, we call it unstructured data as a superset. Really, I've just gotten a lot of reps on building these 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 LM's. So yeah, it's kind of all strung together in retrospect, but wasn't a big plan upfront.
Kyle (04:02)
Yeah, 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 Graflit. You kind of mentioned it briefly, like what is it that you do? Like what is the focus and what does Graflit specifically focus itself
Kirk Marple (04:31)
I mean, the original idea was, I mean, making unstructured data is easy to work with as structured data. 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 a spark for unstructured data, a snowflake for unstructured data? And I was at General Motors, they were taking data that was like video, LiDAR data.
telemetry data, kind of IoT data. And we're having to build this all from scratch, this kind of pipeline to get the ML engineers to even do anything with that data. So 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 broadcasts. But a lot of those techniques just were not known in the kind of data engineering world.
That was why they brought me in at GM where they're like, look, you already know this video stuff. We need help figuring this stuff. Really the couple of companies I was at after that as CTO, were building pipelines for this stuff. Time and time again, it wasn't off the shelf. That's what I really wanted Graphlet to be. Originally, we were called unstructed and we were 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 then with LLMs, started to realize this fits really well. Like we are kind of a long -term memory for what now we call RAG, like Chat DPT Conversations. And so we really flipped 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
Kyle (06:30)
Okay. That's really interesting. And I want to touch on a couple of points that you made. Before we do that, let's 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 what is the difference between those. So high level structured data versus unstructured data.
Kirk Marple (06:51)
mean, the simplest way I look at it is, and usually say is structured data is something that looks like a spreadsheet. I mean, rows and columns, mean, the structure is essentially in that kind of tabular format. And that could live, 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 things like that, or you could consider IoT data 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 3D 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 -document.
area early and then documents were easy to plug in. mean, do more with documents and text extraction and stuff like that. For us, think we came at it from more of a media -centric side rather than a document -centric side.
Kyle (08:05)
interesting. And 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
And so what is it that as somebody's going through, how can we take that and make it into something useful? And then how can you make it a searchable thing? So as you're going through, I want to find all of X things in this. And so 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. it sounds like you kind of approached it from what would possibly be a more difficult, where 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 the, I guess, some of the learnings that you had going from that into more of like the documents and textual unstructured data?
Kirk Marple (09:33)
Yeah, I mean, 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 web page, 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
where somebody needs a new file format, we can map that to our existing structure. I've always been big on canonicalization of, 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. When you're talking to the LM, I don't really care if it came from a Word doc or a PDF or a web page.
track that hardly at all. And so the only special cases, we have tech 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. We don't care if it's a Notion doc, if it's a file off a Google Drive, or if it's a web page.
in the middle of our pipeline, it's just content. And then the text part of it, I think it's what's interesting. 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 move to more visual 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 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 guideable 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 are 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.
And that technique works really well. so, see, I think there's a really interesting transition from kind of old school to new school methods for
Kyle (12:34)
Yeah, that's really interesting. 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 were kind of in this transition phase,
How has that changed both what you've been doing and what do see that changing going forward with either things that you're working on or just how we handle unstructured data generally speaking?
Kirk Marple (13:17)
Yeah, I mean, 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 LMS, do all the parsing and gestion 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 us before. And I mean, especially like startups and folks that they want to work up here, but they got to build down here. And that, think, is really opening up the industry. 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.
Kyle (14:13)
I think that's really interesting. And I kind of want to touch on that a little bit more because you've brought up a point of 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, what have you seen 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?
Kirk Marple (15:09)
Yeah, I think when with I guess it's probably more like 20 months now, but since chat GPT came out, I think people started to map their data to that concept. The whole chat with your PDF, chat with your files. And I mean, a lot of it's just it's low hanging fruit. It's people understand how chat GPT works. And now where unstructured 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.
The people at 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, I mean, agent workflow on my data? And so I think we're seeing those, that's kind of the low -hang fruit of what the common applications are. but then, I mean, we also have this knowledge graph data model that we've had from day one.
and now it's becoming 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. mean, honestly, 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, mean, with some of the interesting ones, I we have folks in like the healthcare area, 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 can use it for rank conversations 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 knowledge graph concept, now looking for it as a customer, rather than it not clicking. And so I think we're in a new swell of capabilities now that we're primed, hopefully, target.
Kyle (17:18)
Yeah, that's great. And I wanted to touch on that idea of the Knowledge Graph. Tell us a little bit more. What is a Knowledge Graph and why is that such a powerful tool for someone to use?
Kirk Marple (17:30)
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. 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 can 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. And that becomes search query, really. But you have to index all that data properly. so it's more than just looking for words, it's looking for concepts. And so that's really been a big idea. I 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.
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. and that's, 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, OK, find me all the data that's in audio format in the last 90 days that's mentioned, LMS and real estate. And so that becomes a filter.
But then you can also give that knowledge graph data to the LLM. You can say, here's a list of products and here's all their metadata, like their SKU, their description, their whatever. It's not just a word per se, it's metadata, like a full entity. That is now what we call graph reg. That's actually something that 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.
Kyle (19:36)
Yeah. 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 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 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 being somebody who's worked on it for, like you were saying, six or seven years or more in this specific 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 has really kind of stood out to
Kirk Marple (20:46)
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 formats can we pull out. So I think that one's super exciting, just got started on that. But I think, 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? And that becomes a really interesting case because you're also pulling in data that's maybe not textually similar, but it might be entity similar. that could be something, and you can even add nodes in the graph after you've pulled in the text. So this might be, I mean, a human is 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 20 years almost. 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 API that you can integrate into your applications.
Kyle (22:12)
Yeah, yeah. And you bring up a really great point that I kind of wanted to touch on as well, that as you create these tools and make them available, how are you thinking about the user experience? Especially as you are a very developer -focused company, one, how important is the user experience? And then 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 aren't just something that gets the job done, but is in fact a very usable and easy thing to integrate into what people are
Kirk Marple (22:59)
Yeah, we've kind of come full circle. We were very end user focused. mean, somebody at a port or a railway was like our target customer for managing all that data. And then we flipped to just be like, OK, here's an API. And now we're kind of circling back a bit. I we built a developer portal. I mean, it looks like in SuperBase, 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, mean, how do you debug the problem? You're always going to run into something. And so we just released 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 what would I want as a developer to debug this application. 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 front -end and back -end development resources and all that 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 signed up, right now we can show them sample apps, can show them 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 15, 20 minutes. Upload a file, see what it output, ask it a question, like a chat DPT 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 people just want
have it click of, what does this do? How does it work? 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 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. And it was good teaching tool for us. So that's been a lot of our focuses.
closing that gap on usability for onboarding as well through the UX.
Kyle (25:33)
Yeah, that makes a lot of sense. And I'm interested, 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. 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?
Kirk Marple (26:02)
Yeah, no, mean, it's always something we're trying to get more feedback. I mean, I've tried sending out a survey, after, like we have a tickler email and there's a survey in one of them, we probably got like four responses. But like, mean, we get good, 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 we've seen by now people can sign up and just schedule a meeting with me and just on our website, grab 15 minutes, 30 minutes. And that has been awesome. Like we 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
And that is feeding back into, I mean, what are their hotspots? Like, what are the things they really care about? 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. Because understanding like, how did they find us? And all that, we can get a lot of details from all that. yeah, it's, I mean, it's so important. I mean, you can't do this in a vacuum, you know, so it's, I love taking those kind of calls.
Kyle (27:20)
Yeah.
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 then 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 think that's great feedback.
Kirk Marple (28:02)
Yeah.
And I think, 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 every day, I mean, I hear somebody being like, 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 now we need a production version of that. 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.
Kyle (28:51)
Yeah, 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, 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, we can win this business or something like that. And like you just mentioned, wanting to stay very horizontal while still supporting customers in 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.
Kirk Marple (29:36)
Yeah.
Now, it's an interesting point. And I think, I 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, I mean, ate up a huge amount of our time for like a year, but it was an incredible customer. I 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, 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 like if people, the diversity is in the data and the diversity is in the schema of the metadata that they want to pull out and the knowledge graph. And then maybe the diversity is
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 light, 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, always think of it as like this, really the tree trunks 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, we want to use Anthropic with our own API key and do some custom stuff, we'll be like, OK, cool. It's configuration. It's not rewriting your whole platform.
Kyle (31:29)
Yeah, yeah, definitely can understand that. 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 your current company?
Kirk Marple (31:52)
And 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. so, I I thought we were a little early when we got started and that the market would catch up and everybody would be 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 why 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 kinds 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, like we were using that term three and a half years ago and nobody got it. And so I think that's, I mean, a hard thing where you have to like sort of keep grinding, keep grinding, keep grinding. And for us, I mean, we want to be a foundational technology. Like it's something where we want to be the snowflake of unstructured data, like just commonplace that here's, mean, put your data in there, build on
and it just works. And so I think, I mean, that really was, I mean, even though we kind of took a foray into 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. 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.
Kyle (33:50)
Yeah. How important for you is 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 a long time. 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?
Kirk Marple (34:19)
I I would just say it's super important. I mean, it's probably the number one thing of, I mean, I got started doing this, I mean, on the side and then it turned into like a, I mean, from a side project to do 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 gonna be greater, I than 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. And I think the capabilities long -term are huge. 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, mean, dig it out of the ground first before you can really start doing anything with it. but I think yet for me, it's the vision is so key and I mean, 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, mean, future -proofing a little bit. Like 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
maybe to my detriment, but I think it's now been a positive. Build for scale. mean, 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, the technologies we're leveraging, I mean, we're not running into the same problems maybe some other folks are. I we have problems. I mean, obviously we have, there's always going to be issues, but...
some of those foundational ideas haven't really changed, I mean, in how the product was
Kyle (36:21)
That's really interesting. And I want to kind of touch on a couple of 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.
Kirk Marple (36:54)
Mm
Yeah, I on a good note, I mean, 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 we were selling to companies, 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 man. And that's a lot of what really drove this architecture and the concept.
of content is kind of, mean, content management, media management, they're all pretty similar. But I think also in another case, it's like, I mean, they're the trade off, like you said, I mean, we 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 get deployed 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. 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 and say we can. And so I think, think sticking to your guns on that and then finding a middle ground, like we've now said, like, okay, we'll.
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, mean, that carved off another big chunk of potential customers.
Kyle (38:41)
Yeah. For those customers or companies that maybe are looking at, like you said, just tons of untapped data or possibilities, how would you begin to approach something like that? Or what would you tell them? Where would they start or where should they start thinking about, we just have all this unstructured data and maybe we don't know what to do with it, or is there untapped potential here? What should they do?
Kirk Marple (39:10)
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? But isn't it? Is it an S3 bucket? Is it actually just your internal web pages? I mean, or something you want to scrape? Start understanding the locations, understand the volume. I mean, are we talking about 10 ,000 PDFs, a million PDFs? mean, is it video? Is it audio? Is it a mix? Start with that kind of stuff. And then really like, and then what do you want to do with
Do you just want to make it searchable? do you really want to have auto summarization of that data? Do you want to have rag conversations with it? Just put together all the pieces of what do you want to do with it? We always try and break it up to, there's that first step of the funnel of ingest from wherever it is, prepare it into a format that we can use downstream. 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, that's not a big deal. But if they have like 100 ,000 PDFs that they want to ingest over a week, then they're gonna have to think about scale and cost is really the other variable that a lot of people, they're like, 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. It's their open AI tokens end up being 80 % of their costs. And so we have to be really cognizant where they just can't throw like the kitchen sink at it and just do everything or they're going to eat like five grand in a day of cost. And 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
You can see your invoice, can see credits you use, getting better graphing capabilities and search on your logs and costs is something we're still working towards because it's critical for the larger customers.
Kyle (41:18)
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 advice would you have for somebody who's thinking about either of those things?
Kirk Marple (41:35)
Well, can say my son is just he kind of was a COVID high school graduate. So kind of took a couple of years of like figuring out what do you want to do? And now he's settled back into computer science, honestly. And so it's I mean, as somebody starting to start that, think it's a really interesting time because it's like people 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, it's a completely different
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 auto -complete in our IDE that helps us code. I mean, it's just going to be extension, continual extensions of that, that augment us.
Because I think there's going to be a push back 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 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 company. I think it's I mean, it's 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 of 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 to its huge amount of risk.
I mean, you're doing support at three in the morning. I mean, you're trying to make customers happy. And sometimes they just won't be happy. I mean, our round of customers now is great. they're all looking forward to the future, how they can integrate this stuff. But I mean, I've done, mean, where it's like wake up at four in the morning on a call back in my old company and try to do support over the phone, I mean, rolling out of bed. And it's, mean,
That's the life that you're signing up for too. So it's definitely both sides of it.
Kyle (44:07)
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, about anything else?
Kirk Marple (44:29)
Yeah, Graflit. We're just Graflit .com. I'm on Twitter as myself, just Kirk Marple, and Graflit'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 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. yeah, I mean, we're actually working to get more of a YouTube channel set up.
Get some of my talks and things like that and and then also get some just more tutorials and kind how to use the platform So hopefully over the next month or two, we'll have more of that stuff up on YouTube as
Kyle (45:07)
Okay, we will put all of the links in the show notes as well. It's a great site. You've got some great blog posts I've been reading through as well. And 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 keep learning more. So we'll put all the links in the show notes for anybody to check out. Well, Kirk, we've got a couple of wrap up questions that we want to wrap this up with 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 technical or business related in any way, but they can be if you
Kirk Marple (45:47)
Yeah, obviously I work a lot, but the only kind of real break that I get, my kids are all kind of grown now. 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 unrelated area and learn all about kind of how they're managing the wildlife and doing all that kind of stuff. So that's been my sort of secret, sort
before bed thing that I've been watching a lot of and definitely learning a lot about stuff that I knew nothing about beforehand.
Kyle (46:23)
Very nice. Awesome. And then final question. Are there any products that you've been using and enjoying? You can be digital products or physical products, anything like that.
Kirk Marple (46:35)
Yeah, I mean, it's I mean, we love linear for product management here. I think I mean, I've used your 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 it just the I we use Slack as well. But I think that just being able to be organized is key.
I think we're still underutilized knowledge management with LLMs. 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 you to kind of manage all of that data better. But I know there's some other vendors that are targeting those kinds of things as well. yeah, mean, knowledge management is always an untapped kind of unfinished business.
Kyle (47:20)
It is, yeah, that's a great call out. Okay, well, are some great shout outs to end with. But Kirk, this has been a great conversation. Appreciate all of the insights that you had and appreciate the conversation as well.
Kirk Marple (47:37)
Yeah, thanks so much, this was lot of fun.
Kyle (47:39)
It was, and I appreciate everyone for listening.
