In this episode, the build of SLC Chat (Serious Land Capital’s proprietary underwriting AI) gets dissected, from architecture decisions to API economics. The product is positioned against off-the-shelf LLMs that miscompare land deals because they lack proprietary data and nuanced pricing formulas. Total build cost came in around $60 and half a week of solo work.
Key Takeaways:
- Build Cost Collapsed 99.8% SLC Chat shipped for ~$60 in API credits and half a week solo, versus $25K to $30K and 4-5 engineers required for a similar product just over a year ago.
- API Economics Drove the Pricing Tiers Each question costs ~$1.50 on the backend (optimizable to $0.25-$0.30), leading to $20/month for 5 questions and $200/month for 100 questions.
- Plan in Cowork, Execute in Replit All product planning runs through Claude Cowork with Opus 4.7 while Replit handles execution-only tasks, avoiding redundant AI subscription spend.
- Off-the-Shelf LLMs Produce Slop for Land Comps Generic ChatGPT summaries miscompare properties badly because they lack proprietary deal data, anonymized comp sets, and pricing formulas with proper hedging.
- Feedback-to-Fix Cycle Hit Five Minutes Seth Williams’ bug report on broken conversation history was parsed through Cowork, pushed to Replit, and shipped in five minutes.
Listen to the full episode for the unfiltered build economics and the pricing logic behind a tiered AI product that has to break even on API costs.
(Podcast transcript below)
Welcome to Get Serious. We’re at Serious Land Capital. We have successfully funded over $6.25 million worth of vacant land deals over the past several years with industry leading 41 % operating margins. Just a note as I’m recording this, it feels like the market is just a little bit better actually. The spring is not quite as slow as was kind of expecting here.
Yeah, real estate, hyperlocal, still really rough buyers market all the way around. But I do just want to mention that, you know, kind of my anecdotal data that we have and a little bit of, you know, spider sense from just being in the industry a long time here. So more to report on that soon. Now, as is usual, often just giving some updates and some really technical frameworks when it comes to.
utilizing AI and the current AI capabilities here, specifically as it relates to our continued development of SLC chat, Serious Land Capital chat. And the goal really is to be my brain, my team’s underwriting brain in chat pod form, utilizing all of the chunked content that we’ve produced over the years, all of our internal
anonymized data related to the thousands of deals that we’ve reviewed and funded and all the proprietary mathematical formulas that we would utilize in order to price properties with the appropriate amount of nuance and hedging that is required within the land investing space. And the off-the-shelf AI models really do not do the best job.
know that because I see a number of folks who will just send us chat GPT summaries for deals they want funded and it is off by it’s like the worst company possible. Again, AI like I’m as pro AI is pretty much anybody, but you have to provide the proper context if you don’t give the AI the proper directions that you’re going to get slop back. You can’t just ask something simple and.
expect that you’re going to get a great answer again, especially in something so complex and nuanced as land investing and when you know, a lot of imagery and so forth needs to be taken into account. So really from the technical side of things and the development side of things, like I just continue to be blown away at the capability of these tools here. And I was remarking to my
business partner and a couple of other friends of mine in the space because we’ve been building a whole bunch of tools and various businesses over the past 10 plus years. And I almost have to chuckle in recollection on how painstakingly long and expensive it took just to get things done. The most basic type of features like setting up ability to take credit cards, for instance. Like 10 years ago, was a
you know, at least for us, like a four to five day task. when all the back and forth you had to have with the support teams and approvals and documentation, like just a big time sync, for something that you would think should be so much, so much simpler. Whereas nowadays, like it’s a simple stripe web hook, that can be done through, you know, these agentic builders like replet.
which I’ll really get into here because, and, and, and that can be done like within a day. Again, just absolutely crazy. The, capabilities that, that are at your fingertips for so cheap, nowadays. So, you know, what happened over really the last week, just over a week, is that, you know, we had finalized the internal build for SLC chat. I was satisfied with the voice, the accuracy of the responses coming back.
to us and so now it was like, let’s transfer this over to the public facing version of this. And so, you know, I
We, we originally intended this to be a custom GPT and then we were going to set up a paywall around it. you know, utilizing another third party provider. I ended up going against that direction, more because we wanted, to utilize, Claude’s models as well as just the tools have gotten good enough. They’re like, okay, the learning curve for setting up our own website and basically web app.
is not that significant compared to having to use a third party provider that’s going to be taking a potential larger chunk of possible margin from us to run it on their system. So I thought, OK, let’s just conquer this learning curve as well here too. So what we ended up doing is.
Again, this could be a much longer discussion here, but I just want to show you the overall thought process. so, you you’ve heard me talk about Cloud Cowork all the time. Again, for those of you who aren’t familiar, that’s a desktop version of Cloud that also has local file manipulation and, you know, reading capabilities. And can also browse your internet and do all the other normal features that that Cloud, the Cloud Chat version can do as well.
and also has some ability to write code and so forth as well. Not quite as powerful as Cloud Code, but kind of sits in between Cloud Chat and Cloud Code. And so, because I could utilize the most powerful models on Cloud, again, we have a higher price plan. I was like, okay, I wanna do all my product planning on Cloud Cowork side. And it already had all the additional context and could read all the files that we had created.
And then I’m going to set up another account on Replet. And Replet is, you might be familiar, one of the more vibe coding programs where you can build web apps or mobile apps or just landing pages effectively and have effectively this AI agent who can help build you almost whatever you want out of code. And it can also help you plan everything out as well too. But again, using…
AI is going to be expensive. if you buy a whole bunch of different plans just to do the same thing when you’re already paying for one, like it didn’t really make sense. So what I did is just handled all the planning on cowork side. And so then I could just have the anticipated output and the directives to Replet throw that in there. And so I could still have one of the cheaper plans there, but not have to go as much back and forth with it. was just assigning it things rather than an active dialogue. So I just handled that.
with, with cowork. Now you could do that in Replet, but again, it’s just, you know, managing costs. Um, plus, you know, the Replet it’s unclear what like their backend, um, model is. I don’t think it’s quite like as quote unquote thoughtful as, uh, uh, Opus 4.7 for instance. So I’d rather have like the best thinking model possible and then just giving it for, you know, doing tasks into Replet. And so, um,
You know, we were able to design this, whole feature set and like, okay, how do we, you know, trying to keep this as basic as possible. Like let’s get an MVP out here. I want to get, you know, some early beta users to test this out and make sure that like our internal model we built on Claude works on the public facing one. So that’s like step number one. and so I could just send all those directions over. could produce a landing page and then, you know, pay wall there links up to Stripe. some of this, had to handle a bit on, on the backend here. Like there’s.
As simple as utilizing some of these AI agents is getting, there’s, there’s definitely still learning curve to it. Like I, you know, and I would consider myself pretty advanced within the AI space. this is like not something you can just like pick up with, with a snap of your fingers. like it took me probably like a day or two to really understand the full learning curve on, how to utilize replete and back and forth with coworker properly. And there was like, there was more.
bugs that got generated then that then expected, especially if you’re trying to like patch things. Oh, like one bug gets fixed and you know, another comes up because you’re fixed actually introduced another bug. And so like that there’s just a lot of testing involved. And even though Replet has built in quality control and QA, it’s not going to catch everything. And so then you still have to go back and like make sure, does this user flow actually make sense? Can I sign up with
you know, credit card with a promo and it’s like logging my user properly on the backend and replic. That’s why, you know, I’m in this is another aside here, but like for again, as simple as these AI tools are getting, like there’s almost always going to be room for people who are just want to be like that the extra step above from an AI architect perspective, because there’s still going to be so many people who don’t want to connect.
all these various systems in place and they’d rather have like, you know, somebody else handle it on the backend, even if it is considerably easier than it was even a year ago. Like even the Stripe side, like Replit can’t handle all that, like it itself, Stripe’s its own backend. You have to set up, you know, fairly technical elements and know how to, you know, follow a co-work’s directions. And like, it doesn’t always know the,
UI on the backend as well too. So I would have to screenshot a whole bunch. Like I can’t figure this out right now. And so like there’s a lot of troubleshooting is kind of the takeaway here where again, these AI models are going to continue to get better month after month, year after year. But still it’ll pay to have a more technical brain when you’re trying to architect these things out. Or again, just hire somebody else.
who can help you out with kind of piecing some of these things together. But it was enough for me to grasp and continue going on. But yeah, I had to teach myself quite a bit. And you have to like connect a claude API account because I wanted the claude models on the backend. And so that requires setting up a whole new claude account, like not within my personal. And I have to buy tokens, which are a lot more expensive compared to.
you know, utilizing the monthly subscription claude plans. But if you’re just paying like on a per use basis, hosting your own website, like it is quite pricey. And with how much, you know, information we have within our knowledge base for the SLC chat, like the, again, not to get too technical here, but like the first question a user asks and you know, assuming they don’t.
ask follow up questions within five minutes of each other. Like each question is going to cost me, like our business, like roughly a dollar 50, like pretty, you know, as you can see, pretty significant spend. Now we could adjust some of the backend infrastructure to make that probably closer to like 25 to 30 cents. Nevertheless, it’s not like a penny or less than a penny. Like that there’s, you know, quite a bit of expense that goes into
hosting and producing AI answers. so, yeah, a lot of these companies are heavily subsidized with private funding and so forth to try to gain user growth instead of being like purely profitable, whereas like, yeah, we’re trying to be bootstrapped with it. So like I have to limit the amount of questions that can be asked per pricing tier.
to ensure that it’s not going to be a loss later for us or that we won’t at least break even on the service we’re trying to provide. So that’s just education on behalf of the end user here. And still, the end user could say, where’s the value prop here? Who cares if it’s expensive on the backend for you guys? I need to get the answers that I want here. So like…
You know, the higher price, you, you, you make something like less people are going to, you know, want to be in the mix to, to purchase it. So yeah, your value prop has to be excellent. So that’s what I’m trying to figure out as well too. It’s like, I know for sure there’s nothing else like this on the market that does like, as I went over at the top, but how to justify this, this price here. And so
Like I was willing to set up, one question free trial, you know, it’s going to cost me a buck 50 just for somebody to try it out. And like I could include, you know, protections from people, uh, know, spamming random fake email addresses. Like again, I could just instruct Claude cowork to hate, tell the replet to do all this. I don’t need to know how to do any of that on the backend. Like, is just remarkable. Um, where I can just have the idea.
Hey, I want these protections in place. Or I could ask claude cowork. Hey, like what else I’m missing here from a software design perspective? Like what’s, what are other, you know, um, uh, typical, um, or, or, you know, best, uh, uh, best in class, uh, user design, uh, or backend design, whatever, um, that can help protect some of these, you know, edge cases or people abusing the system and so forth. And like that can just get built up. Um, now again, sometimes it created some errors where it, uh,
made it harder to test or like we ran into some other bugs. like, again, this stuff’s not perfect, but it took a lot more iterations than I thought. Nevertheless, you could get a lot of this done quickly and just keep instructing Replit and Cowork to keep going back and forth to fix things accordingly here. And so yeah, free trial that’ll and you know, collecting an email to even get that. So then you can start marketing and then anticipating $20 as they really start our package for five questions month.
So that really ensures, again, if each question’s costing like roughly a buck 50 at most, that will ensure there’s at least enough margin on the backend. And I’m just being transparent about this because I’m also like want to help people how to build things like this and what you should be thinking about when it comes to distributing a product like this as I’m going through it in real time.
But like, also believe in the value of my product that people will be willing to pay for this because of what we’ve produced in the data we have on the backend. And then the premium tier. And again, I’m just like more focused on the premium set up for a lot of products nowadays and services and just like keep going after, you know, the higher earners or the folks who are like really care about what, you’re offering would be $200 a month for a hundred questions. so again, that
protect some of the margin on the chance that people are really heavy users and taking lengthy amounts of time to ask questions not within a short sequence, but more spread out time between each other. So all that was considered as I built this out.
And again, like trying to strategize between the expensive Claude API tokens and just trying to see, is there really a product market fit? We don’t like know that for sure until we start getting some early users and get some additional beta testing and so forth. So like, want to do this as cost consciously as possible.
But like all throughout this bill, as I was getting like the learning curves just knocked out of the way and teaching myself this, like these were a couple, like even setting up the claude API and the, you know, replet, this was like 50 bucks, 60 bucks. Um, plus our, our existing claude coworker subscription, um, and just me working on it for roughly half a week. Um, and I was able to get like,
you know, again, just just a tremendous, tremendous amount done like a working product could could get actual beta test, you know, use uses out of it, I could use it. And it was like, you know, still giving gold standard responses on the back end. All of that was able to be sorted out again, within a handful of days and like just over a year ago when I was building out landpricer, similar product. But you know, more complicated in some some areas, certainly. But
You know, we didn’t have a lot of these agentic tools there and like to get this amount of work done again, like just over a year ago, like took, you know, a few months, maybe a little bit less than that, but more engineers working like at least four or five different engineers and the speed from, you know, feedback coming in to iterating and fixing a bug and implementing was just way longer. Whereas me with cowork and replet, I can do something, get a fix in within 10 minutes.
Um, and probably cost, you know, again, like a little over a year ago to do something similar, probably cost me out of pocket, like somewhere between like 25 to $30,000, like just stunning, stunning difference, 60 bucks compared to 30 grand, um, to, uh, to build something relatively similar, um, which again, just give, gives the idea of the power of these systems and how much has changed in such a, uh, a short period of time. Um, and so like,
That’s to provide you where we’re at. There’s ongoing beta testing right now. I actually just got some excellent feedback back from my friend, Seth Williams. You all know him from Ari Tipster, the guy who’s just, you can trust him to provide outstanding feedback and especially from the tech perspective, like love diving into this stuff similar to me. And so…
you know, he, definitely pointed out some, some key issues like right off the bat. like even some things that I was waking up in the middle of the night thinking about, like, yeah, this still needs to be easier to use. Like we should be able to attach, like PDFs, like people do their own DD and throw it in. and, so it’s just like, you know, easier to use compared to like just chatting and typing away for like what type of comps might,
might be relevant to the subject property. Cause again, like we’re in the age of cowork. So everybody wants speed and just ease of use here. And Seth was pointing out like ideally this should be more agentic. I’ve tried to build stuff like that before, like not saying it can’t be done or it certainly can be done like having Claude or perplexity browse on the internet and pull comps and so forth. Entirely possible for a future version of this product and possibly integrating into
you know, those larger LLM model APIs and so forth. Like we could certainly explore that. But as a medium term solution or something I could build like right off the bat is okay, yeah, being able to upload PDFs or your own spreadsheet reviews of certain properties. And so then you don’t have to go back and forth with the bot and like save yourself some, you know, cost on the questions here. And it’s like, that was an improvement that, okay, I have a game plan, you know, going back to cowork.
throw it in a replet and we can get that thing set up without issue. Or like there was a key bug where, you know, I thought that the conversation history on my testing was working. so you could look at, know, whatever property you reviewed, or whatever question you asked the SLC chat for Seth, it wasn’t working. And so it’s like his questions were disappearing. So again, cowork like pointed that out, you know, took the transcript from the feedback he gave me. Parse it all through. Okay. Here’s the fix that needs to go in the replet. I throw it in there within five minutes. It’s fixed.
Um, like again, the, the speed from feedback to fixes is just nuts. Um, and then also figuring out, okay, like, you know, branding wise, we probably need to adjust how this thing is going to be marketed. Um, how you’re going to handle this, like upcoming masterclass, um, for demoing the product, like what will be the most effective way to do that? You know, looking at Alex Hermosys, a hundred million dollar offers, how do we best, um, you know, uh,
set ourselves off to like, you know, check as many boxes as possible to, to make this activity again. Like I can run all that within one cowork prompt. It can spit back out at me and we can just iterate and push forward, take that feedback and keep improving the product. again, like this, I just cannot get over the speed and the low cost nature it takes to do development for real, real products. it’s effectively a one man show, behind these things. so.
wanted to share with you all where we’re at on that. And again, just be transparent about our process. And I’m hoping there’ll be at least a handful of you. There already have been that have been asking about this and looking forward to sharing more with you all shortly.
Um, and you can see, uh, below that I’m going to be hosting the masterclass. you’re listening to this, um, prior to May 19th, but, May 19th at, 3pm central, I’m doing a free hour long masterclass again, really demoing this thing, you know, the fixed version of SLC chat, um, showing you how, why I built this thing. Um, uh, and you know, showing why it’s superior to a lot of the, uh, out of the box, LLM systems or some of the other products within the, uh, the land space that.
do different things better, but like not from really an underwriting perspective, which is, you know, the differentiating piece that I’m trying to develop. So if you’re interested in hearing more about that, seeing it live, would love to have you there. You can find the link to sign up within the show notes here, or you can just DM me, hey, interested in the masterclass and I will send you the Cal invite directly here.
And for any of you that are in Kellyn’s Accelerate to Automate program, I’m also, you know, came on as a AI advisor consultant for her and her team. They’re tracking for $40 million this year, just absolutely wild. So thrilled to be a part of her team as well here too, and continue to develop my own AI skills and deliver more to you all. So excited for that, subscribe and share everybody. Take care now and bye.
