WEBVTT 1 00:00:00.000 --> 00:00:45.000 Foreign. Welcome to the GeoMob podcast where we discuss geo innovation in any and all forms, whether for fun or profit. Hi everyone. Welcome back to another episode of the GeoMob podcast. Today it is my pleasure to be joined by Lucas Martinelli, who is now the founder and CEO of Reprompt and previously he worked for quite some time at Mapbox. So very extensive geospatial experience and we have a lot to talk about because both his current business, but also geocoding in general. So very excited for this conversation. Lukas, welcome to the show. Introduce yourself, please. 2 00:00:45.000 --> 00:01:21.000 Thanks Ed for having me. Well, I'd say I'm an OpenStreetMap nerd that turned that beautiful hobby into a profession and then is turning it into a company now running my own startup as well. And this is really how it all started is with OSM when I was a student and I got into that because the professor told me, hey, OpenStreetMap is really hard to use because it doesn't use this new thing called vector tiles. And that's how I got started both mapping and then creating infrastructure on top of OSM. And lo and behold, 10 years later, I have a whole company around this. 3 00:01:21.000 --> 00:01:28.000 Fantastic. So let's dive right in. What exactly is reprompt? What does your business do? 4 00:01:28.000 --> 00:02:13.000 We are building deep research for the physical world. So what that means is we power the databases of PIs and addresses of most mapping companies in the world. For example, grab TomTom here, TripAdvisor, Foursquare, a couple I can mention, they all use us for when they have businesses or addresses that they use for ride sharing or their mapping services and they give it to our company. And similar to like a perplexity or ChatGPT does deep research. We do deep research that is very location specific and we help them fix up their data. Very similar to mapping, armchair mapping essentially in OSM or if you contribute to a foursquare or whatnot. 5 00:02:13.000 --> 00:03:06.000 Okay, so this is quite interesting because over the years at GLMob we have had many, many different people, hobbyists, companies, organizations, talking about the challenge of POIs and getting a comprehensive data set. And of course the challenge there is a to collect the data, but then to keep it up to date, right? Because the world is constantly changing and we've seen so many different approaches here from I've seen companies that literally pay people to kind of walk down the street and, and record what business is where. Obviously crowdsourcing with things like OpenStreetMap or whatever, but people driving around the Googles of this world, driving around with their vehicles, collecting things from. So how are you doing it? What is your approach and how is it different? 6 00:03:06.000 --> 00:04:41.000 Absolutely. So we believe the only way to really create a global coverage is if you do like full on web scraping. That's its drawbacks and benefits. Right. When you go and do ground truth survey, absolute excellent data, some of the best you can do. Problem is it's always out of date. And whenever people have tried it with cryptocurrency and whatnot, people game the system and it's flawed in that. The other one of the best approaches is if you, if people just naturally give you the data because they benefit from it. But we're trying to cover the entire world and we can't rely on that. So you can think of us as having indexed all the websites that take about pois, having a search on top of it. And then when you ask a question, hey, is the Isa Ramen restaurant open? What are the opening hours? We go, we find all the websites that are talking about this, we visit each one of them and. And then we cross reference whether they're sharing the same facts or not. Just like a human analyst would. Or what if you did that when you were filling something in OSM and you were only armchair mapping? And we then submit that to folks like Foursquare, to folks like a TripAdvisor. And we measure ourselves constantly against what a human would do if they had a computer and a mouse. And we're roughly at 97% accuracy of that. So you can think of us as the world's best AI armchair mapper. 7 00:04:41.000 --> 00:04:46.000 So how many POs are we talking about here in your database? 8 00:04:46.000 --> 00:05:33.000 We have a database of 210 million POIs. And so we've run that armchair mapper 210 million times. And so this is something I'm very excited about. Right. In the past, when Apple Maps was coming online with OpenStreetMap or Amazon, they had these ginormous teams of folks that were doing mapping, right? Both tracing roads and buildings. And it's really cool because now with large language models and vision language models, you can replicate that same work. It costs you money, right, because you have to spend it on the tokens. But it's so cool that now I can build a company that essentially replicates that same work of deploying tens of thousands of virtual humans doing that sort of laborious work. 9 00:05:33.000 --> 00:05:47.000 If you don't mind, let's go into this into some detail. So you are crawling the web yourself or you're getting data from someone else's crawl because I imagine just the act of crawling the web is of course, a huge technical challenge in its own right. 10 00:05:47.000 --> 00:06:59.000 Oh, this is very interesting. So what we do is we use different search engines at first, when you tell us about the place, first step, similar to geocoding that you're familiar with, is like figuring out what you're even talking about. So we ground it. So we do a bunch of searches, we figure out which websites talk about it, and we call that grounding and entity resolution. And so as we search about the place or a business or an address or a building, we take that point and we associate all the websites that talk about it. So that could be a store locator site, that could be a public Facebook profile, that could be a Uber Eats page, and we check that it's indeed talking about that entity, and we connect those websites and so now they're permanently connected to that place. And if we ever touch this website or read it again, we can go back and update that Python. If you ever want to redo something about that PUI or search it, we immediately know which websites to go visit. And so we have an index that's not fully the web, but very focused on these spatial entities. That's an approach that's much more manageable. 11 00:06:59.000 --> 00:07:26.000 But immediately, having spent some time on geocoding, immediately, I'm horrified at the challenges here. For example, if someone is like, you know, the McDonald's, McDonald's in such and such town, but, you know, there might be five or six to McDonald's, so how the deduplication and the understanding which, which actual branch is meant or, you know, how do you manage that? 12 00:07:26.000 --> 00:08:44.000 Oh, that's great. So, right, if you Type into a ChatGPT, Deep Research or whatnot, it will just go off and collect all of that. So the very first, we make sure we collect all the possible places you could talk about. And if it's very clear, there's only one, you talk about. So Apple Store, San Francisco, there's only one. But like, you know, possible there's multiple. We now know we have high conviction, just like a human race, you're talking about that one. If there's multiple, it's a toss. Unless, you know, you've given us a little bit of information about a road that it's close by or a neighborhood, and if we can't resolve it, we stop, right? Because any details we would talk about it would be wrong. And, and so because our users have other agents use us or upload like CSV records or whatnot, we then stop working so it first step for us is really know exactly which place we're talking about. Just like if in OSM you clicked on one PI, you're editing that, you have lat long, you have some information to ground it. Okay, how there's actually a really great use case that we find where. Have you ever looked at the credit card transaction statements that you're getting? 13 00:08:44.000 --> 00:08:50.000 Yes, yes, of course, of course. We've had speakers at Dual Maps. Exactly. Trying to geocode those. 14 00:08:50.000 --> 00:08:56.000 That's right. I think those are great examples because those often don't have the road, they just have the store and the city. And so that's. 15 00:08:56.000 --> 00:08:59.000 Yeah. Or they just have some kind of weird cryptic code. 16 00:08:59.000 --> 00:09:30.000 That's right. Yeah, that's right. And so this is a good example of. Oh often it could be many, many different merchants to match to, but sometimes it's very clear it could just be one possible. And so that's very important. And LLMs by default do a really poor job at that because they're just trying to like oh, make it work and go down a rabbit hole. And so this is this first layer where you both need a geocoder. You need to know all about the neighborhoods, you need to be very certain about the entity resolution and only then proceed further with researching. 17 00:09:30.000 --> 00:10:01.000 Okay, so my next question is what you described this kind of multi step approach, I mean it sounds quite comprehensive but thus also sounds slow. How do you do this fast enough? So people often ask me, can't I just use an LLM for geocoding? And I'm like yes, if you want to get your answer in 15 seconds or do you want it in 15 nanoseconds? I mean like, you know, it makes a difference. So. 18 00:10:01.000 --> 00:11:01.000 Oh, you're absolutely right. So I Mapbox. At Mapbox I used to work on a geocoding service, right. That always has under 10 millisecond latency. And that's a very different problem. You need to pre bake the index, make sure it's really fast, make sure it's memory. And you could say we've built one of the world's slowest geocoder. It's literally like deep research. Right. That can take minutes and minutes and then we do tons of tricks to make it cheap as well. Right. Like instead of LLMs, real time inference, we use like batch and we retry because we're doing this millions and millions of times. So we have a very large volume slow but highly accurate researcher geocoder as you can say. And once you've produced that result it can be used by a normal write geocoding in your database later on, but we think that's quite valuable. So whenever the normal geocoder fails you can fall back to us and we go do like the deepest possible way. 19 00:11:01.000 --> 00:11:22.000 You're doing this on your own hardware or are you using kind of public various other big LLMs like the OpenAI and Claude or whatever services you're using or you're running your own kind of models and things to this 20 00:11:22.000 --> 00:13:10.000 excellent question. So three years ago when we started it was like all the debate should you run your own hardware or your own models Fine tune. We chose not to fine tune our own models like GPT3.5 at that point was coming out. But we always use the existing models and had layers and we're sort of glad so far that we always use existing models because it is not possible to hit the same economics. So we have avoided the last two years essentially fine tuning or custom models and whatnot and use the off the shelf models with other inference providers that don't share your data and use that now models have gotten ridiculously better and it's much more about the prompt. However we found that the most important thing now is that you have the right tools for the context. So example, give it your own geocoding tool and then having when you have a lot of these millions and millions of oh hey, break out this address into components, then you can take an open model, take one of the large models, take our annotations of humans and create a very specialized fine tuned model that you just run for address extraction. And because we don't iterate on that job a lot and then you can still run that on someone else's inference and that works really well. So what we've learned is use the best possible models, use them as teachers and whenever you have good subtasks that you can really optimize yourself, those you can use a qin and fine tune it yourself. 21 00:13:10.000 --> 00:13:23.000 Okay, Sounds like you've put a lot of work in here. Give us a brief history of the company because it's only been going for a few years, hasn't it? 22 00:13:23.000 --> 00:13:57.000 That's right. So maybe let me share how I even got into. Yeah, because it definitely keeps you captive. Once you're in, you can't leave. And I just love it so much. So I did this project called Open Map Tiles as my bachelor thesis, if you've heard of that. This was essentially combining all of the OSM data and creating vector tiles and making it free for everyone to use. And then I created this map editor called maputnik for editing map libre styles. 23 00:13:57.000 --> 00:14:01.000 And this was while you were a student? 24 00:14:01.000 --> 00:15:44.000 That was a student. And that's essentially mapbox got wind of that and was like, hey, you can't really reverse engineer all our protocols. And I was, it's open source, that's the whole point of it. And we got into a huge kerfuffle. But what ended up happening is mapbox said, why don't you just work with us? And I was quite motivated that that point because they were doing open data and open source. And so I got the chance to join their team and it was later on in their executive team. And my job, very funny enough, was to move it from open data to proprietary data. And so I spent huge amount of time essentially helping buy 150 data sources, combining them, like conflating them not just PoIs, but also addresses, roads, buildings, and then creating these data layers that power geocoding, search, directions, services. And so when LLMs came out three years ago now, I was like, oh wow. Everything I used to do with my mapping teams and the teams combining, I think if we break the problem down, what could a smart human armchair mapper do? And break it down to that primitive, I would redo it completely. And so then three years ago I, me and my co founder who lives right next to me, we're literally neighbors. Which is pretty funny. It can only happen in San Francisco. We got into iCombinator and so for the last three years we kind of took everything I did at Mapbox and we turned into a product that other companies can use. And so it's been a, it's been a three year journey. 25 00:15:44.000 --> 00:15:55.000 Congratulations. How difficult was it on the commercial side to kind of convince companies of this new approach or, I mean, if you could share. 26 00:15:55.000 --> 00:17:24.000 Yeah, I would say one thing that I'm really glad I learned having run a data team is doing evals and evaluations. So everything in mapping, when you sell to a company, you first of all have your name and your integrity that you're doing a good job. And then two, there's no way around not eviling your data or telling how good you are by actually showing samples. And that takes time. That can just take two months where our process was always go send us some data from your database. The first one that you have, give it to us, we'll run our deep research agent over it with fixes, we'll give you those fixes back. We'll have a human team on our side annotate to see how much we agree you'll do the same. Then you end up at a metric like, oh, we believe it's 97% accurate agreement with our humans cross sampled. And that's how you prove that an AI can actually be as accurate as, as someone you pay for doing this. As an armchair mapper, and especially early on, Ed, it was kind of hard to convince people because LLMs were hallucinating a lot. And I think what's shifted is now people are like much more just, oh, yeah, of course it can do that. While three years ago you had to do a lot of evaluations to prove it's accurate. 27 00:17:24.000 --> 00:17:33.000 But if I may ask, would you say the technology has actually improved or just people's attitudes have changed? 28 00:17:33.000 --> 00:17:54.000 I would say that, you know what, we use technology, maybe it went from 85% out of the box to 88. So not that huge. Definitely. Everyone is much more open to exploring it and going and just funding a project compared to before. 29 00:17:54.000 --> 00:17:55.000 Okay. 30 00:17:55.000 --> 00:18:53.000 And I think that's really because everyone at this point has asked ChatGPT a spatial question. Hey, is this place open? Hey, could you help me draw a geofence? And so knows much more what's possible. And so I've really appreciated that. At the same time, things have gotten better in aerial and imagery understanding. So if you looked at a Gemini two years ago and you asked the questions about Arial or you ask a question about street side, it's very clear it wasn't trained on that data yet. So you couldn't count how many doors it has or spatially. Reason. Well now if you do it, it's clear that much more street view data, for example, in Gemini has gone into it and it reasons much better on it. And you can still see that Cloud and OpenAI really struggle at anything aerial or street side. So some models are getting way better, some aren't. 31 00:18:53.000 --> 00:19:44.000 One question on that. I mean, one of the big issues around AI is the legalities of this data. So in terms of like, so if Gemini, for example, is using, if I use Google's Geocoding API, I have to sign up for some sort of key or token or whatever. And in doing that I agree to their terms and conditions and thus they retain control of their data. And presumably it's the same when I'm using Gemini and things like that. Like I can't just keep their data. I have to admit I haven't used Gemini or at least I haven't studied it in this use case. But what is the issue around that if I'm using these API these AIs to correct my data. But whose data is it then? I mean this is a big open question, I think. 32 00:19:44.000 --> 00:20:42.000 Yeah, if you ask an LLM without any web searches and whatnot, what's the right address, it will just kind of pull from its weight and potentially hallucinate. So we never do that. We feed it all the evidence that we found across all the websites and make a judgment on that. So what happens is essentially it is public data that you're fixing your data up. And so we do a couple things to make sure it's no problem if you're using. So one thing we do is we do not visit websites that have like a robots txt or so that avoids it. The second thing is if you think about facts of the world, such as an address, you can't really copyright that. But it gets tricky, right, if you think about facts of like is it wheelchair accessible or not? If only one website talks about that. So what we do is. 33 00:20:42.000 --> 00:20:58.000 But if I may interject, I mean, for example, it is the fact that this coordinate is this address and yet if I do that geocoding via any commercial proprietary geocoder, I have to agree their terms and conditions, right? 34 00:20:58.000 --> 00:21:00.000 That's right. 35 00:21:00.000 --> 00:21:07.000 So I don't think it's as simple as saying this is a fact and thus it's free to use. 36 00:21:07.000 --> 00:21:38.000 That's right. If you use someone's geocoder that does not work. But if you purely are on the web and visited a website just like a human would, and you see it like either an aggregator, self declared site and then what we do on top is like if three websites agree on a fact, it can be considered as clearly public data that can be used versus just one website talks about a fact. And so I think that's one thing we've done really well on figuring out. There's lots of AI companies out there which just like literally. 37 00:21:38.000 --> 00:22:08.000 Well that's just it. Yeah, I mean they're just, it's very unclear a where did they get their data? Did they have the rights to that data? Who owns the data that they spit out? Is it, you know, the company or is it the user? Is it, you know, obviously there's not yet case law for this and it may vary differently from jurisdiction to jurisdiction. But I think this is a big question mark that everyone, geospatial, but also other industries will of course have to deal with. 38 00:22:08.000 --> 00:22:57.000 Oh absolutely. And that's why as someone who has a geocoding business, you should be very careful if you use a geocoder what it applies to downstream data. And so that's why we built our own geocoder. Right. If you train on aerial, it's very important. You have rights of extraction for that aerial. Same with street side. That's where we then work with Mapillary or a Carta View. And so that's one of the cool things about websites is that you can go and you can read them and the facts that you find can be considered public data and you can contribute that. You could even contribute that to projects like Foursquare, Places or Overture and whatnot, as long as you don't violate any toss. And so that's kind of a really hard thing that we had to solve that you wouldn't think about is how to make sure the data is clean. 39 00:22:57.000 --> 00:23:01.000 Well, exactly. That is a hard thing. It's much more difficult than people might suspect. 40 00:23:01.000 --> 00:23:05.000 Oh yeah. 41 00:23:05.000 --> 00:23:20.000 Are you using OSM Overture? I presume not if you're saying you're just kind of getting this from the web. But maybe what's your opinion of those projects and how does it all interplay? 42 00:23:20.000 --> 00:23:21.000 I'm a huge. 43 00:23:21.000 --> 00:23:24.000 Obviously you said you're an osmo. 44 00:23:24.000 --> 00:25:14.000 We, you know, I work a lot on it, especially the road network at mapbox. And I think it's like what happened there, you know, across Amazon, Facebook and everyone is that the core road network and topology is shared and then everyone has their own attributes about it. I think similar things have happened with addresses where like the boundaries in there are shared and everyone uses them. And we don't use any OSM because I. The ODBL is tricky to use with commercial customers. We use Overture often as a base because we again, I'm a fan of these projects, like helping Overture Foursquare have the full PI coverage if possible. That's great. I think that helps everyone. And then the really valuable information that always changes, I can still make a business on top of that. And so. So I think it is important to not have something with ODBL that you're interested when you build a commercial product. As much as I love OSM and so we build our database on top of Overture, this is like 35% of it overture. And then the rest we fill in ourselves and we try to key or match against all the popular projects. So like Foursquare, like Overture. So we have those IDs associated. Because when I was buying data, I would have loved to have the data just already keyed to open data sets out there because everyone is using them now. I think a tricky thing Puis Always hard in OSM because it is not possible to ground truth and get all of them yet to have some of the best ground truth information. But it's not really possible for an LLM agent to contribute in an ecosystem. I think that's still quite hard. But I am quite excited for platforms 45 00:25:14.000 --> 00:25:16.000 where 46 00:25:16.000 --> 00:25:30.000 if we all spend money on tokens, it would be nice if they could share in wikidata the information that they've gathered. And I wonder what that project will be in the end. If I'll have some extra time, I'll launch something like that. 47 00:25:30.000 --> 00:25:58.000 Okay, well, that sounds interesting. Yeah, I mean, you're right. Pois in osmo. It does feel like we've kind of hit the limit there, especially in that we, you know, as you said, There are some POs are very easy for contributors to identify and contribute to, but many are not and it's not clear that we're getting all of them and that how do we scale it up and how do we. 48 00:25:58.000 --> 00:26:23.000 I can tell you some facts which POS are really good that I'm like, for example, schools, railways, even embassies, OSM has one of the world's best data sets. They don't obviously have all the embassies, but a lot of things where a passionate person can find all the locations we find OSM just outperforms anyone. But anything that is like restaurants, it's not. 49 00:26:23.000 --> 00:26:29.000 It's just so hard because of the pace of change and the exactly so. 50 00:26:29.000 --> 00:27:04.000 But military fields, I think those are some of the most interesting where people for open source intelligence are really into it. Or if you're talking about transit, it's one of the world's best data sets and this is where humans are taking real or really digging into making sure it's each quality. So I hope maybe there we'll see people doing more OSM imports, using AI agents to clean up and then standing still behind it with their own personal brand. And then as it gets better, we will maybe allow AI agents to contribute, but I think that will be a while. 51 00:27:04.000 --> 00:27:47.000 Yeah, I think what's needed is also a cultural shift in the community which we're not quite there with yet. This is something we could talk about. We could do a multi hour podcast about this, I guess in the interest of time, maybe I do have a few more topics that I wanted to get through with you. So. So first of all, you know, you, like many of our listeners here on the podcast, are from Europe, but then you, you made the move to the us. Was that just because of the job in Mapbox or because you did that when you started this business or talk about maybe what, what took you there and what your impressions are and any, any comments you have on that. 52 00:27:47.000 --> 00:28:46.000 Oh, yeah. I wanted to really work on everything geospatial and with the smartest possible people. And so there was just a density at Mapbox back then of people who worked both on a routing engine like OSRM and Valhalla and a geocoder like Carmen and a map renderer. And he wanted to be around all of those. And now when I found my wife here, I have a son, so I'm of course now also invested in the. The Council. Thank you. But for launching a company, San Francisco has just made it kind of ridiculously easy compared to Europe in how easy you can get access to really things like spending money on tokens. Right. I think it would be pretty hard to spend hundreds of thousands of dollars on tokens to create better map data and getting that funding. So that is still really real 53 00:28:46.000 --> 00:28:47.000 how 54 00:28:46.000 --> 00:30:04.000 easy it is to get access to capital as one reason to be here. And then the other thing I think that has really helped to be here is we got access to agentic coding like Cursor. And everyone talks about Claude code like nearly a year early. Remember in 2024, early on, you know, starting. Everyone starting to use cursor. So those little hints of what's coming are much easier to pick up here and it gives you a little bit of an edge. And I think that's one thing I really like. One thing I really love about Europe and Misto is people just dedicating a career to being really, really great at one thing as opposed to hopping around. And so even when we were at Mapbox, there was REM team in Berlin was just like, so killer. And I love that about the geospatial expertise now, including myself, where you spend 10 years on something and you just become an expert at it. There just seems to be more people like that in Europe that stay at that versus shifting from, oh, let's do spatial, let's do crypto, let's do AI and the next. 55 00:30:04.000 --> 00:30:30.000 Yeah, yes, we do have many of those people in Europe and I'm pleased to say that many of them are geophysics mob regulars. So hopefully you'll get the chance to come to a geomob in Berlin or London or one of the other cities at some point. We would love to have you, you and or your colleagues speak about what you're working on. I think it'd be very interesting. So good. 56 00:30:30.000 --> 00:31:07.000 You Know, another thing I'm exciting excited about is there's all these geospatial services. You work in a geocoder and there's like many geocoders out there that are like in Rust or C or something or a routing engine. And it used to be very hard for someone that knows a ton about geo to contribute without also being C wizard. So I'm quite excited now with Claude code and all of that that people both creating data is one thing, but also contributing to projects that are historically really hard, like routing and then adapting them for your own purpose. That's something I'm quite excited about. 57 00:31:07.000 --> 00:31:39.000 Well actually we've had many talks over the last year or two, particularly on this issue of routing rout for the use case of delivery is routing for cyclists, routing for pedestrians, routing for wheelchair. All these kinds of different use cases. Yeah. So there's a lot happening there. What are the things that someone who's been in the geospatial industry now for some time and at some of the, you know, at the cutting edge, let's say, what are the things that get you excited for what's coming next? 58 00:31:39.000 --> 00:33:11.000 I still think that data sets are still becoming more open over time. So that's always been a cool trend in addresses, but now and you needed historically someone that speaks the language and looks through Prox or Zurich's open data portal and get it in and ingest it. So now you can have AI go catalog all of that. So I think it's actually kind of a cool trend as governments make data more open and then AI and all these different weird portals can fully navigate it with a browser and catalog it. I think it will sort of create that flywheel even more for open data. And then I'm also really excited that you know, there used to be no way how to beat the Google Maps really. And I say this because you could only beat it with like having better ground truth collection where you couldn't quite do what DeepMind does without a lot of money. You couldn't hire tens of thousands of people. And because of LLMs you can go and play at the same level. And that's super exciting in geospatial even with search. We talked about search. LLMs are kind of slow for search, but you can now create search re ranking training data sets that you needed a lot of people to do or Google only could do because that traffic and now you can build much better search by having LLMs create the training data. So the playing field is just much more open. And so I'M super excited about that. 59 00:33:11.000 --> 00:33:18.000 Yes, but I mean. So I agree. Nevertheless, it's also more open for your competition as well, so. 60 00:33:18.000 --> 00:33:19.000 Oh yeah, yeah. 61 00:33:19.000 --> 00:34:09.000 So it's kind of double edged in that regard. So it remains to be seen who will the rising tide lift all boats and we're all just get a bit better and thus, you know, nothing really changes or will someone find a better way to implement it all and thus, you know, create true. A true differentiator. We shall see. So anyway, let's wrap it up there. Lukas, in all seriousness, I do think you should make the trip over to Europe and get to a GeoMob event. We would love to have you. It sounds like you're really doing very cutting edge stuff and I think a lot of people would like to learn about it. So I hope we can make that happen in the near future. 62 00:34:09.000 --> 00:34:14.000 I saw friends in Berlin. Yeah. Did you? 63 00:34:14.000 --> 00:34:19.000 Well, we have quite a few map boxers in. In Berlin as well, so. Yeah. 64 00:34:19.000 --> 00:34:28.000 Yeah. Anyway, I used to go a lot to that office in Kreuzbergen. 65 00:34:28.000 --> 00:34:38.000 If. If this conversation has raised questions for people or they would like to learn more about. About you, about reprompt, what is the best way for them to get in touch? 66 00:34:38.000 --> 00:34:59.000 If you go to the website Reprompt IO, you can actually very quickly upload some spatial data and play with our agent. Like ask it, hey, what are the opening hours of this place? And otherwise, just send me an email at Lucas with a Kreeprompt IO. 67 00:34:59.000 --> 00:35:10.000 Okay. Can't get easier than that. So thank you very much, Lucas, for taking the time to come on the show and look forward to tracking your progress in the coming months and years. 68 00:35:10.000 --> 00:35:17.000 Thank you. Ed, 69 00:35:17.000 --> 00:35:56.000 Thanks for joining us today on the GeoMob podcast. We hope you enjoyed the discussion. You can get the show notes over on the website, which is@thegeomob.com and while you're there, you can sign up for our monthly newsletter and find links to our accounts on Mastodon and LinkedIn so you can stay up to date on all our upcoming events. We hope to see you at an event soon.