Latest / 4kd.ai / Ep. 2 | Agentic AI in the Wild: Real-World Builds and Business Use Cases

Transcript

Ep. 2 | Agentic AI in the Wild: Real-World Builds and Business Use Cases · 4kd.ai

• Craig (00:23) Hey, back everybody to Forked.ai, the podcast for better AIers. That's my new tagline, Jack. Worked on that all night. So guys, it's great to see Mr. Dave Moses again and Jack BeVier. Guys, welcome back. It's episode two. Here we go. Dave, you're wearing the sports coat today. So you get to kick us off with just give us briefly what you want to talk about today and what you're working on. And then we'll move to Jack. David Moses (00:55) There's a disincentive for me to wear sport coats in the future. ⁓ But anyway, I... Craig (00:58) No. I want to see you come in a tux on the next one. David Moses (01:03) that would be awesome. So yeah, I've been working on some pretty cool stuff. I did my homework from last time. I was able to create the replacement for Mile IQ, which was tracking miles for our people. I think I've got it working pretty cool. And it has some other cool things I've been working on that. And some people in my office have taken more of an interest in, hey, this looks, what if we do this, what if we do that? So that's nice. Jack BeVier (01:26) So how did you, Craig (01:27) So let's, Jack BeVier (01:28) I'll dig right into it. How did you build that? What did you use? What's it do? What tools did you plug into? Craig (01:28) let's. David Moses (01:37) So I used OpenClaw and with Opus 4, I actually have a waterfall of models. I have like three consumer tokens and it just, when it rate limits or overloads one, it just goes to the next one. So it can keep going. And so I basically just told what I wanted to build. I already have an EC2 server, so that's kind of where I'm hosting it. I got a free piece of software called TrackCart, T-R-A-C-C-A-R, loaded in everybody's phones, had them set up the server on our side. Completely free software. Yep, yep, that's what tells me where they're at. Craig (02:12) Does that take care of location services? And then how accurate is it? David Moses (02:21) It's within... It really depends on their setting on their phone, but as long as they've got like precise locations set up, I mean, it's accurate. I can tell where people are inside my building, like what part of the building they're in. It's TrackCar, T-R-A-C-C-A-R. Craig (02:32) Can you give us the name of that software again? David Moses (02:41) And it's a free app, and then you just set up the server. You can have them host it, or you can set up a server yourself. I just told my OpenClaw to do it, and it did it. It set up the server for me. Craig (02:51) So for those that are new, well, everybody's new to the show, ⁓ tell us what the use case is. Tell people again briefly what it is that you do and what this app is meant to do and how it solves a problem for David Moses (02:54) You Yeah, so we manage plus or minus 400 single and small multifamily rentals, and we also manage about 4,500 units of condo, so HOAs and condo associations. And we have various people out in the field, both our managers and our technicians, and we need to know where they are. We need to optimize routing them. We need to track their time and track their miles. So those who are using personal cars, We can reimburse them. Those who aren't using personal cars, can track, you know, where those cars are going, how many miles they've got on them, that kind of thing. And so... Craig (03:42) And how would you be doing that prior to OpenClaw and? David Moses (03:45) Before what they were doing is they would check into a service ticket and they'd check in at the beginning of the day, and they'd out at the end of the day in the payroll software, and then over the course of the day they would check in and check out of individual work orders. And that would, we'd get the time from the work orders, we'd reconcile it against the total time for the day, we'd figure out what applied to a work order, what was just general time. like meetings and whatever, going and getting the truck fixed, whatever they were doing. And then we would track their miles on a software called Mile IQ, which was 10 bucks a user a month, and it just gave you a nice log. They would go in there and they would swipe left if it was business, swipe right if it was personal, and then it would give us a log at the end of the day or the end of the week, and here's how much they're owed. and we would just add that to their check to reimburse them if they were using their own personal vehicle and if they're using a company vehicle. We just had the log for compliance, IRS compliance. And all that was looking fine except when I went to start to apply those, I wanted to create an automation that would apply those miles and those hours automatically into the payroll software and automatically in QuickBooks assigning the cost to the job. And MyLiQ has an API, but they don't really have an API. It doesn't really do anything. So there are no endpoints that are useful to me, and you have to integrate it with this other piece of software. I was just like, no. So I was like, maybe there's another way. And so I found this track car. Actually, I can share my screen if you want to see kind of what it does. Craig (05:25) Kyle, we set up for screen sharing? I love it. Let's do it. David Moses (05:30) All right. Can you guys see? Now you can see a lot of you. ⁓ So basically, there's two things. One, this is like mission control. You guys can see this. So this is basically a custom overlay over Google Maps. I use the Google API to get the map. Craig (05:31) Eheheheheeee Is this every one of your employees that are out on the road right now? David Moses (05:49) Yeah, so the employees, we only have a few techs. And so you'll see those folks, right? Those are the techs. All of the non-highlighted houses, those are houses we manage. All of the red ones are houses we manage that have open work orders. And you can see if you click on one, it'll actually, you can click on here, it'll open up the work order for you. Yep, yeah, it's. Craig (05:55) in You built this? Us. So can you, can you just like, just give us like the what, so you, got, you, you download a track car. You figured out how to make your endpoints work to track mileage and sort of where people are. But this is way beyond just that though, Dave. This is, I haven't got a full interface of like where all of my properties are. Like this is not just a track car build here. This was a multi-phased project. David Moses (06:44) But I needed to load in, in order to make it work, I needed to load in all the places that are relevant. It also has all these little tiny ones. Any place we've bought supplies in the last 15 months, I had it go and find it, find all the address, the GPS coordinates of all those places, and add them to the map. So it knows when we're going to pick up supplies. And then you basically come here, you throw in your phone number, and it's gated, so it has to be you know, from our office with one of our phone numbers. And it'll, you know, send you a code so that way you don't have like anybody and everybody just going and looking at this. Right, and then you pop in here and then these are all the pending days. So I click on one and you can see this is how many hours they worked that day, how many miles they drove, this was the cost to the company. ⁓ This is the reimbursement they've got coming to them. I can switch this from personal vehicle to any of the vehicles we own. ⁓ And it... ⁓ It's got like, you see each one, each drive and each stop got allocated to a job. And it looks like here he drove home for lunch. Craig (08:05) Does it know it knows that it's allocated to the job because of the location service? David Moses (08:11) So I basically, it just gives it a bunch of context. Here are the open work orders. Here are all the places that we manage. Here are all the supply places. Figure out, and here's where they went. So then it just, figure out where does it apply. Like where, were they going to a place? Were they going to, does this drive make sense? Does this stop make sense for any of the open melds? And if it does, assign it to that meld. If not, like you'll see he went home for lunch, so it allocated those as non-work. But it actually went through, I added this yesterday, and it basically said, well, yeah, he went home for lunch, but he would have gotten the miles and the time to go from one job to the next if he hadn't gone home. So it estimates that, just creates a drive for him. he recaptures that three miles ⁓ and that five minutes, meaningful to him, probably not so much to us. And then these are our divisions, so it allows us to change what we allocate. Craig (09:07) So. David Moses (09:10) it to, we can override the miles and give them more or less. And then we just click Submit here at the bottom. And once you click Submit, once you click Submit Approval, goes into our payroll system. then it goes into, ⁓ when we enter payroll, then that goes into QuickBooks. So that's that. Craig (09:32) That's fantastic. Jack, what are you thinking? Jack BeVier (09:36) Dude, so you had to connect. So you have it connected to property mill to end up folio, right? To give it the context to identify the open work orders. David Moses (09:42) Well, no, I only connected it to Mongo. Yeah, I don't have it connecting directly to the APIs. I have the API, all the data in the APIs, all the data in PropertyMeld, all the data in AppFolio, all data in QuickBooks, all comes into a Mongo database, like once a minute, depending on what it allows. So it's all in Mongo, and it just grabs it from... because it's lighter, it's faster, it doesn't, you know, when there's an agent involved, I don't want it to be able to go back and write something incorrectly. So yeah. Craig (10:20) So you basically have a heartbeat that pulls in data from multiple sources, like a Koran. We'll get into what AI has been costing Craig Fuhr and Jack Baviera over the past couple of weeks in a minute. However, what I'd love to know is how are you keeping your token costs down if you're pulling, because it's pulling a lot of data to update that. that app, how are you keeping your token cost down? David Moses (10:48) So actually, it's a really, really minuscule amount of money. I mean, it's costing me pennies a day to run that. It's not a lot of context that goes in. It might be 10,000 tokens that goes in, and it's like maybe 1,000 tokens out. Craig (11:09) because it's sort of updating Mongo with new information, not pulling everything. David Moses (11:14) ⁓ the update to Mongo is not agentic. From AppFolio and property mail, that's not agentic. That's just a simple API call, pull in the data, dump it into Mongo. ⁓ Really, the agentic part is just, here's all the addresses with open work orders. Here's the open work orders. And here's the drives. Here's all the GPS points that they hit throughout the day. Craig (11:22) Got it. Got it. Okay. David Moses (11:36) figure it out and then right now I'm using I think chat 5.4 to do that reasoning ⁓ I could probably switch that to grok I think I'd probably get just as good of a result for half the price so yeah Craig (11:42) Yeah. So ⁓ interesting use case. Love that. So the map is insane. I was thinking of different use cases for location services the other day for me. And it was more of a, so I love the idea of my bot knowing where I am to keep my shit right. Like. Meaning, Hey, you said you wanted to go to the gym three times a week. And I know where the gym's located, but you're not there. And it can call me and go, Hey, it's six, it's six AM wake up because your car, your phone is in your, in your house right now, not at the gym. Fun use case. And the reason why I bring it up is because I think David Moses (12:22) you're at. Craig (12:39) for the creative curious person, they look at what you did Dave as an insane business enterprise use case, which is beautiful. But it's like, hey, how do I start with this? How do I take these concepts that these guys are talking about and sort of make something a little bit more simple and useful? yours is an amazing case. Mine could be built in about 30 seconds. So that's awesome, man. is that? Is that like rev one for you? Are you still working on making it better? Like what would you, how would you tell, how would you make that whole app better date right now if you had to like sort of go add some more features to it? Just curious. David Moses (13:20) So my vision for that is to have an agent doing dispatch, an agent doing, so basically the agent knows where everyone is and it knows when someone calls in a work order and it says, hey, you're five minutes away, let's text the next. work order and tell them, we're going to be half hour late or an hour late. This guy's going to go and fix whatever is going to be fixed. And boom, now I don't have to have somebody going 30 minutes this way, 30 minutes back. That's the idea, I think. I mean, it's all about efficiency. It's getting to people quickly. Craig (13:47) I love when Jack leans back in his chair and gives it the old. That's, that's, that's Jack of y'all. That's a good one. That's a good one there. David Moses (14:01) and prioritizing it. If that person that we're going to delay has got an active leak, we're not going to delay that. Craig (14:10) Does it does is the supposition though that like that like all of your people they all have different skills. I would assume that like you know all the handymen have sort of similar skills or whoever's going out to the projects. And so do you rank them by skills and then sort of like how fast they work or do you just like hey that guy's closer. He's the better guy right now because he's closer. David Moses (14:31) I mean, great. I mean, those are great questions. And I think I'll never be able to do that. But the AI will. It'll be able to look through these are all the work orders that this guy has done. Here's the ones he got five stars on or stars on here, the ones he got one star on. So it's not going to send them, you know, send him to go do something that he sucks at. We have the data we've been collecting, you know, reviews from tenants for a long time. We we have the Craig (14:39) Yeah. David Moses (14:57) detailed notes of the issue and the fix for all these work orders for going back years. So it shouldn't be difficult to do. It's just dig through it, figure out what their skill set really is, and does this make sense? And sometimes it's just, hey, you're a warm body. We want to show this important customer that we're on top of things. So just go there and be a warm body. You're not going to fix this, but you'll put some eyes on it. Give us some meaningful data and when we do get out there, maybe we save some time fixing it, but it could just be that, right? But the other thing is just up-skilling them. It's like if we just say to them, look, we got a lot of this stuff going on and you're missing out on work orders, they're incentivized to do more work orders. They're incentivized to do more things when they go to one work order. So it's hey, this, you know, smoke detector is beeping, so I don't have to come out here again to do that. Let's just swap out the battery while we're here. That kind of thing. it's really, I'm a really big believer in figure out a workflow, figure out what you want it to do, and get it to the point where all you're tweaking is the context and the prompt. So what am I, what, what, what, Fuel, am I giving the AI to make a decision? And what instructions am I giving it to make that decision? And if you can just tweak those two things, even if it's not gonna get it right now, the model may be smart enough in three months where it will. Craig (16:35) Because you've laid a great foundation for it with the workflow, gave it the right tools, stood up like a Rev 1, and now you can just tweak. Is that what you're saying? David Moses (16:46) Yeah, and you get feedback. The other thing is, how are they getting the recursive self-improvement? How are you giving feedback after the agents made a decision? If that's just me as the programmer doing it, then I've failed. But if I have a way where the users of that workflow can give meaningful feedback and that there's another workflow... that analyzes that feedback against the existing production flow and says, hey, if you tweak this prompt in this way, or if we go and grab this data from here, it'll make this better. So they may be bug fixes, it might be feature requests, whatever it is that the ultimate users have. It allows the system to suggest its own improvements and eventually just improve itself. Craig (17:39) Go ahead, Jack. Jack BeVier (17:39) Okay, so I want to rewind like a lot. How did you connect? And forgive me if we talked about this in a previous episode, maybe I just need to hear things a couple of times. But how did you connect your various systems to Mongo? Like literally, is it, am I like super overthinking this thing and I just need to log into Mongo, create an account and then point it to my Appfolio API and... and just tell it to ingest everything. And then I just start querying the Mongo and then I get the Mongo API and give that to my open claw and I can start querying the Mongo database. Like it's all just in there and organized enough because it's labeled that it is searchable. Or do I have to like do something to turn my AppFolio API into a database, into something useful? Craig (18:22) set up. Yeah, so can I so do you so Mongo is basically a database. Correct. I've not used it. I use super base and the question the question I would add to that is do you have to set up the schema in Mongo or can it can it just take the the API and figure out what schema would be set up in Mongo? David Moses (18:28) Yeah, go. Craig (18:51) for you to query. David Moses (18:53) Yeah, so as a caveat, there are probably better ways of setting up the schema than even the way the APIs have it, the way the actual software companies set up their APIs. But I don't really account for that. What I do is I want it mirrored. want it exactly in the same schema setup that the existing Craig (19:15) interesting. David Moses (19:16) you API has. For you, Jack, the answer to your question is just, you know, because Jack uses AppFolio and David uses AppFolio, it's really just, David, send me the workflow that pulls all this into Mongo, and then you can just put it on your N8n and use it. All you'll have to do is set up a Mongo account, and, you know, it will literally create all of the collections for you when you, the first time it writes. you know, first time it writes to a collection, right? If it created a collection for tenants, right, as the endpoint, then, you know, the first time it does it, it's gonna automatically create that collection. doesn't exist, so Mongo, you know, you, just inherently, when you write to Mongo to a collection that doesn't exist, creates the collection. So that's really all you'll do. It's a very simple-looking flow, to be honest. It's really just a... a code node that has a list of a bunch of endpoints, right? And then a HTTP node that it will run through, loop through each one of those endpoints, grab everything, all the records that have updated in a recent amount of time, right? Whatever cadence you have it on. And then writes whatever that data is to the Mongo database. The step two. is adding the web hooks because AppFolio has web hooks. So just have the thing trigger for that endpoint whenever something changes, whenever a record changes in that endpoint. And then the final one is to create a separate collection that tracks changes. okay, so if this record got mirrored into the system, if it's different than it was before, track that change. Here's the field that changed, here was the value before, here's the value now. And if you can do all those things, then you really have a good sense for, mean, that opens up a world of... of information. Jack BeVier (21:02) I feel like that's like a like what you're describing right there is like a crucial skill, right? Like, and it can be applied to any database, but like, but like, I just want to make sure that like, I don't know that I that I get it right in my brain. And also I think it's like, it's like one of those like stupid little logistical things that like is a humongous unlock to it. So like, it's not like there's a single API endpoint that you can connect a database to in AppFolio, you have to identify the architecture of endpoints. Craig (21:18) so important. Jack BeVier (21:32) that exists. Okay. And okay, so right. like you have and you have to ping all of them individually, like each of those endpoints is the key to getting data out. So you have to create an end flow that goes to each of those endpoints and then triggers a flow to pull the data into Mongo and then monitor for is I'm going to script the language here. I'm going to script the vocabulary here and then David Moses (21:32) There's like a hundred and of them. Jack BeVier (22:00) Separately, it's a webhook that you have to monitor, or you just set up a flow that each of the webhooks in, I'm screwing this all up, in AppFolio trigger, then it pushes data through those endpoints. Clean that up, clean that up for me. David Moses (22:18) Yeah, so essentially. Craig (22:19) There is nothing more fun than watching Jack Maver learn on the fly. It's the best. David Moses (22:22) No, no, it's essentially what you said, right? Unfortunately with AppFolio, and this isn't the case with all software, but AppFolio literally has all of their webhooks going to the same URL. You give it a URL, every webhook that fires, it's gonna give you the information for what endpoint changed and the record that changed when it fires, but it's essentially gonna just send everything to the same ones. You have one workflow and a giant, you know, a giant switch or you don't have to have a switch once you get to the point where you don't need one, but basically it just says, this is the endpoint. Okay, so I'm going to this leg of the flow, right? If you set up the original flow correctly, which I didn't the first time because I didn't know how to do it, but the second time I did, I got it right, it'll just fire the webhook. It'll say, because this is the endpoint that had a change in it, this is the endpoint that I'm going to... use in my call. And you had to know how to do this stuff even probably six months ago. But Opus 4.6 is smart enough where you can literally give it access to the API documents and it will build this workflow for you. It knows. Yeah, that's it, right? Craig (23:28) Yes. that's. So let me let me give you Dave Jack BeVier (23:29) Just tell it what you want it to do. Craig (23:33) let me let me give you a crazy use case that I think is really great here at Dominion. Jack so we take calls all day long phone rings off the hook because our marketing department is insane and we get calls on multiple different voice over IP numbers through a program called CallRail. Currently CallRail. David Moses (23:52) Choreo? Yep. Craig (23:55) currently we're looking at other options I believe so in any event, call rail does have an API and I'm I'm not really the kind of guy that's going to go in there and look at every single a endpoint on an API. I just said and so I built this app called wreck me and basically what I want is I want to take a call from a customer. Have that call transcribed. Summarized and any key points. that were in the call name email address phone number FICO experience state that they're in stuff like that it should automatically see that put it into like a nice report for me and because I'm not yet connected to our CRM which is Salesforce I would just then copy and paste that into the customer notes for the lead right. And so what's the reason for that? Well, the reason is now I can have like my own sort of dashboard that shows me all of the calls I took on Monday and the ones that were important and the ones that I needed to get back to and the follow ups that I had from them. And it's brilliant. It's working wonderfully. to my point, the CallRail API is just okay. And I didn't really, and I don't know what all the end points are. So I was in, yesterday and I was like, hey, we've got this great app. It's like 98 % there. It works. It doesn't work great. And I was like, here's the API docs. Here's all the open call docs. Here's exactly, you know, here's my entire open call workflow. Make it better. And so it's, it's actually working on that as we speak. But you're, but, to your point, Dave, I think it's a, yeah, you don't need to like, I'm surprised to hear people still talking about N8n. I believe in it. think it's really, it's an interesting way to do, you know, to graph out your workflow. But for me, it was too kludgy and too geeky and too, you know, I'm not a tinker like that. And so I just, I just give it all to open call. I just give it all to cloud code. And I'm just like, you know, figure it all out. And it does perfectly. Jack BeVier (25:53) What's it building? What's it, what's its replacement for? What's it, what's, what's it, what is the open claw version of N eight N bunch of Python scripts. Craig (25:54) to Jack's question. David Moses (25:54) slow. Craig (25:57) Pardon. David Moses (26:04) Yeah, a bunch of markdowns and bunch of Python scripts. Craig (26:08) So to your point though, Dave, so let's say the API that you were pulling from AppFolio, Was that it? So I think you said it had over 100 endpoints that it can query, correct? And so let's suppose that I love the idea of just saying, yeah, pull them all in. Pull them all into the database now. And then we'll figure out later which ones that we want to generate reports on, correct? David Moses (26:19) Yes. Craig (26:38) That's much better way doing it rather than saying I just want this this this this just pull it all David Moses (26:40) It is. Just put it all, yeah, it's not, mean, storage is cheap. What I would say, the two counterpoints to not using a software like NANet are, there's really just two that I can think of. Number one, it silos your credentials, so your credentials aren't living somewhere in OpenClaw that anybody who could hack it could figure out a way to get into. NADN gates those, even the workflow, even the agent in the workflow can't see the credential. ⁓ Craig (27:13) about the what about if you use like a one password CLI and just have a vault then for your open club. David Moses (27:18) Yeah, that would also work, yeah, for sure. And the other reason to use N8n, or something like it, is to visualize the workflow on an execution. So you can actually, it gives you that step by step, this is what fires, you can click and look at the data that went in. For non-geeks, people who aren't like us, that probably... Craig (27:29) Yeah. David Moses (27:42) isn't important. Most business owners or people who work in businesses are trying to make them better. They probably don't care. As long as it works, it works. And I'm actually at the point now where I don't even look at the Mongo collections. I just say, just organize this data. The AI will just organize it, set up the collections how you think they should be set up, organize them so that some future agent can look through the collections and intuitively understand where this information probably lives, find it, and use it. And so I'm not even paying that much attention to how it's organizing the data that's not just a mirror from, all I really care about is like, it's got a mirror from the API the way it is. It's also a cool backup, you know, because things change. People do things and they screw things up. And none of these APIs do a really good job of audit logging. Now I can create my own audit log and I can see, yeah, you changed this record from this to this on this date. You're the user that did it. And, you know, this is why I'm creating an agentic workflow that replaces you from, so you don't have to make that mistake anymore. So, which is, you know, that's what you want. You want to be able to track where are people making mistakes? Where are they wasting their time? Craig (28:32) Yeah. David Moses (28:57) You build workflows ideally around the things that cost your business money. If it's costing a lot of time, costing a of money, or there's a lot of error that happens in that particular process, that's where you want to target, okay, this is where an automation should happen. I'm thinking about it from that standpoint. Probably the entrepreneur in me would criticize myself and say, I should probably spend a lot more time working on how I can get more business, how I can grow, you know. But that's... Craig (29:29) I would take the contrarian point on that, like, hey, look, if you believe, like say, if you're making loans to real estate investors and that AI is going to exponentiate the amount of business that we could. take in, then you better make your people and your processes way more efficient to handle that, that flow. And I think that there's a part of me that like stays up till four in the morning, talking to my bot building stuff. And I'm like, this is stupid. Like, what are you doing, dude? Like go to bed. But there's this, there's this also, there's a part of me that says, no, wait a minute. You better learn these tools, man. Because you know, when Jack, when Jack like ramps up, production, you better be ready to handle it in a way that like, you know, nothing falls through the cracks. And I think that's kind of what, what I'm focused on. Jack, Jack, let's, can we transition? Jack purchased, question, question. Jack BeVier (30:29) Wait, I have one more question for Dave. I only get an hour with Dave a week, so I have to get as much out of this selfishly for me. David Moses (30:35) You can call my phone rings Like I tell people it's only one way like I can if I don't call you back Don't get mad at me because my phone really only does work one, but call me Jack BeVier (30:48) So what's, how are you doing? Or are you turning these, are you turning this into Quickbook or into journal entries? Like, you bookkeeping that? Are you still bookkeeping or like, are you, have you taken it to the accounting side of things yet? David Moses (31:02) We are, we are currently, like all of my team, they do bookings. If I have to do a journal entry, if I, if it's gotten to the point, like I sold a property, I do post, I want to see it. So there are certain things where I'll do the entry. And when it's that, or I bought something, so I'm going to have to. you know, tell somebody what it is. So instead of telling them what it is, I have an agent where I'll say, here's what I bought, go journal it or go at it as an expense into QuickBooks and it will just do it. And it knows the problem is, it'll just do it, yep, I just tell it to do it. I can even like, I can literally take a picture of a receipt at a restaurant. Jack BeVier (31:33) You do have that set up and it'll just do it. David Moses (31:42) And if I write which division it's for at the top, will just, hey, do want to enter this in the QuickBooks? Yeah, go ahead and enter it in QuickBooks. Boom, just enters it. Jack BeVier (31:50) And all of this is running through OpenClaw. David Moses (31:53) That's running through open cloud, yeah. Jack BeVier (31:55) That's running through open claw. Craig (31:55) So you take a picture on your phone of a receipt, you write a couple notes on it, and you feed that into Telegram, I'm assuming, or Discord, or however it is that you talk to your bot. then, yeah, how do you do it? Jack BeVier (32:07) How do you do it? How do you do it? David Moses (32:09) both. do Telegram when it's not something I'm gonna do a lot of, like when it's not repetitive. When I'm working on anything that's like a longer term project, I'll use Discord because I can actually allocate literally the token cost by the workflow. And I'll show you my other, that very geeky mission control that literally tells me this is how much this particular build costs in tokens. But it also silos it for memory. That's really the most important thing. It's very cool. I'll show it to you. But the purpose wasn't to do that. I just thought, this is cool. It can do that, so why not? But the real... Craig (32:38) ⁓ I am so not sleeping tonight. Jack BeVier (32:42) That's such a good idea. Craig (32:50) So how many Discord channels do you have set up? Curious. David Moses (32:55) Probably about, well, I'm not the only one. There's like three of us working in them. it's probably 25. No, it's probably like 10 or 12 channels. And it's probably like 25 to 30 threads. Because in each channel, you can have a bunch of threads. And you can actually get, you can cost these things down to the thread in terms of how much time. But again, the idea with using Discord or. Craig (33:18) Jack, I just got giddy. just had chills. David Moses (33:21) But I can't stress this enough, the whole purpose, the most important, if you jump into OpenClaw, you'll find that the most frustrating part of OpenClaw is that it uses compaction just like any LLMs. That means as you talk to it in a particular conversation, it's going to forget what you were talking about. So if you silo those threads, now you can just say, go back through the thread, get yourself up to speed, and it's not reading a bunch of nonsense about where you should have dinner on Thursday. It's only looking through the context that's relevant to that thread so it can get itself back up to speed very quickly. It's really more for memory than is for cost allocation, which is kind cool. Craig (34:02) Can you set it up, stupid question, can you set it up to speak only to a particular sub agent or skill? Like so, I've got a UI skill, right? And I wanna take some project that I'm working on and I say, hey, UI skill, go take a look at, know, Fort Day iNews and, you know, use all your skills to make it better. Can you do that? David Moses (34:13) Mm-hmm. Craig (34:27) Or are you always still talking to your main agent and the main agent is the orchestrator to handle the rest? David Moses (34:33) So I've tried to do it the first way, and I'm sure it can be done. I failed. So the way I would say, the way I do it is I only talk to one agent. He's in a bunch of different threads. I have more than one agent I talk to, but in terms of building workflows, I only talk to one agent, and each of my other two people who build, they have their own agent. But then I ask them always, and this is like, this has worked out way better. I don't let them build anything. I make sure they understand they're just the orchestrator. They are to spawn an agent or two or three, depending on what they're building. And those agents are to do the work. And you as the orchestrator are just to make sure they keep going. Make sure they keep doing what they're supposed to do and deliver a product. And when they say they're done, don't trust them. Test it and tell them. that they're not finished yet, if they're not finished. And I found that that kind of hierarchy where I don't speak directly to the agents actually doing the work, I'm speaking to the orchestrator agent, works better, especially because Claude gets overloaded, Codex, run out of, it'll just API limit you when you get to a certain point. you need, when those things happen, the flows tend to stop. I can't go any further. And then if you don't have a cron job set up for him to continue to check on it or check himself on it, then it winds up the whole thing just you wake up the next morning and he'll have stopped at 10 PM and nothing will have happened. But if you have the orchestrator agent continually checking in on them and the orchestrator agent is on a different token or on a different, you you haven't using API tokens now that he can keep them going. if they, if they get, he'll literally tell me, Craig (36:00) Yeah, yeah. Been there. David Moses (36:17) they got overloaded at 2 a.m. I waited 30 minutes for a cool down and then I had them resume. And it'll tell me that that's what Craig (36:29) Yeah, let's move on a couple of a couple of things here so Get you know staying on topic. I think when people fire up open claw For the first time and they and you just start playing with it is it's very powerful powerful right out of the box I think when you hook it, you know, you hook up some LLM to it for a lot of people It'll probably be clawed or anthropic You could really hook it up to anything But you'll get to a point where, and you guys chime in here, you get to a point where you're just working on a lot of stuff, a lot of ideas, you've projects that are half built, some that are not, some that are real close to the finish line, and you're human, so you kind of forget what it is that you're working on at times. But more importantly, the token costs, if you're just banging on this thing, everything anthropic is very, very expensive. You cannot have a... $20 a month plan and expect to do any significant agentic work that might be overnight while you're sleeping. I thought that I had a really beautiful organization, Dave, where I had my main bot, Rocky, and then I had all of these bots underneath that specialize in certain things. And when I told Rocky to do something, he was supposed to tell the other guys that are now working. on a much less expensive token. In this case, was I'm using Minimax and KimiK2. both of those. so, I mean, you're talking exponentially less expensive than an anthropic token, right? I was on a way to a doctor's appointment, phone rings, it's Jack Bevere. I always pick up when it's Jack. And he said, hey man. David Moses (37:56) Two five. Craig (38:17) Hey, hey, we're cutting you off. And I was like, from from Dominion or like, do I still have a job? So he's like, no, man, you're on track to spend $10,000 this month in tokens. You've already spent over $4,000 just in the last few days. Jack BeVier (38:23) Yes David Moses (38:35) You Craig (38:36) I damn near had to pull the car over. mean, like, I know there's a lot of people who work for someone and they're like, yeah, I'll just fuck it. Yeah, I'll just bang on that guy's shit all day long. And like, I was shocked. mean, literally shocked. Went home that night and tore it all apart. I mean, broke my bot so bad, Dave, it would do nothing because I'm like, this is, this is untenable. I can't, like, I felt horrible. And so what I discovered is, that that orchestration feature that everybody talks about, that's that's really kind of the holy grail of like this whole technology, right? How do I get multiple employees, digital employees doing my doing my work, but at the least cost, right? Like I want to talk to my my smart guy, Rocky. I want to talk to him and Opus because he's smart and Opus has great attitude when it comes to the persona effect of Opus is so much better than I think than like any of the other tools. Jack BeVier (39:16) you Craig (39:32) It just gives you back like this, this, you know, the responses that you're looking for. That said, it was not orchestrating. And I really had to dig deep with open call docs and not that deep. I just basically had to go back to it and say, Hey, here's all the open call docs. Here's where you totally F'ed up. You are now building yourself as the orchestrator. You're smart, but you're orchestrating. You're delegating everything. And here's all of this. And I had to rebuild all of my skills. so that they were sort of like, by the way, if anybody is building a skill, if you're not using the Claude Creator skill, the Claude Skill Creator skill, it's awesome. And so you just tell it like, hey, I've got the skill, it is my copywriter, let's say. And it's copywriting for LinkedIn. And so it knows the LinkedIn algorithm, it knows the hooks and all that stuff. And if I say, I'm gonna write something for LinkedIn, it goes to the copywriter. It has some... It has drastically changed, I think, my token count, because now they're all auditing themselves for tokens as well, and I get a report. And so it's working really well. So for folks that are trying to set up multiple agents under sort of like a, you know, you're making a digital organization, a digital business structure. You really have to be careful because you'll limit out very quickly on tokens and you'll wake up one morning with a bill from Anthropic. that you were not expecting. So just be very careful when you're doing that. Dave, Jack, you guys can jump in anywhere. David Moses (41:02) That's spot on. That's spot on. Jack BeVier (41:05) Yeah, so like, how do you guys have like, I've heard eggs. So again, I'm like the like super detailed, like exactly, exactly. How do you do what you just said there? Like, cause it feels like a good, it feels like it goes, goes over my head. Like, all right, I can get the concept of like, Hey, I'm going to sub out to a bunch of other models. Lesser tasks can use lesser models. Fine. Get that. But how do you set that? How do you set that architecture up? Like you literally say like, Hey, I want you to download a suite or I want you to get access to a suite of other models and like, let me know when you need me to put the credit card into the Kimi website and the whatever website, like, and then get those API tokens and then you're good. Like, that what is that exactly what you do? Like, how do you, how do you get all these other models in the mix? ⁓ mechanically. Craig (41:51) Dave, so Dave, I'd love your input on this. So here's what I did. I basically so I've ever since I started working with OvenClaw, I just talked to it through telegraph and I and I, you know, I've made a lot of stuff. So it's obviously working. I mean, I've made a lot of interesting projects with it. That said, it was not doing what it was. was flat out lying to me is what it was doing. And so, Dave, I basically, I don't like using Telegram as my IDE. I'd much rather be working in Cursor and just give it full access to my OpenClaw workspace and say, hey, let's use Codex to build out the following. Help me take a look at the entire workspace. So I basically opened up Codex, the app, on my desktop. I gave it access to, yeah, it's just, it's like Claude Cowork, right? Where you could just basically give it a folder to work in. And I said, take a look at this folder and tell me everything that's wrong with it. And man, did it ever. It was like, yeah, this is not working the way it's, it basically confirmed everything that I thought. And then I said, okay, now do a full audit of all of the skills that I've already built, the ones that are active, the ones that are not. Jack BeVier (42:43) is an IDE. Craig (43:09) tell me which LLMs that they're working with and tell me if Rocky's actually using them or not. Nope, he's not. And so then I said, okay, here's all the open call docs. Here's all of my APIs in one password. There's a session token that can only use this vault, which is all of my APIs and take what you know and make Rocky an orchestrator. And it did. And so I don't know that I love because when you're working in codex, it's local, right? I can't do that from my phone. So you have to kind of screen share, which is really kludgy. I need to, Dave, need to do, I need to have a way where I can work on my Mac mini that's here at Dominion running open claw. without having to like screen share into it, which I do with using tail scale. ⁓ It's fine. It's fine. It works, you know, but like I'd much rather have something more seamless. Anyway, so that's what I did. I gave, basically gave the entire, the entire directory plus all the docs to Codex. And I said, let's start working on one PRD at a time. Jack BeVier (44:05) Yeah, tell scale. Yeah, that's what I do too. Craig (44:23) like, you know, let's work on the skills, let's work on each skill, let's work on Rocky. And then, you know, basically said, you know, every subagent skill now is just gonna use Minimax. Minimax 2.7 is really, it's pretty sharp. And it appears to be where... Jack BeVier (44:38) You just told it in your instructions. You just told it in your instructions you wanted to optimize for cost efficiency. Craig (44:44) Exactly. David Moses (44:46) So. Craig (44:46) And then it has a fallback. So let's say Minimax gets API'd out. I've banged on it too hard. It'll go to Kimmy K at that point. Go ahead, Dave. David Moses (44:56) So I first of all I love that idea. I have not done that and it sounds really useful. Craig (45:02) let me, let me, let me give you one last thing. So in codecs, which, which by the way, I really, I, I was doing them side by side. would do the same thing in Claude code. And then I gave Claude code the same thing. I gave it the whole access to my, to my workflow. and then I said, I gave it, I said, build me a PRD in a four phase approach with all of the tasks and tell Rocky that he, you know, to delegate. And his report back to you has to show that the phase was complete. Everything was recorded in obsidian, super base backup. And so he comes back with a response. I put that into codex or cloud code, you know, COVID on the desktop. So I basically copy and paste that over. And then it says, that was a pass. Let's move on to phase two or Hey, that was not a pass. Here's what he needs to change. Give that back to Rocky. He goes out, finishes it. then we move. So everything is tested and validated through Codex to go back to Rocky. And it's been really great. David Moses (46:04) That's really. I have to listen to this podcast on like point eight. Jack BeVier (46:11) I'm going be doing that. Yeah. I'm going to be doing that. I'll like, wait, say that Yeah. Craig (46:14) What is it that, but like you've done far more, I think far more difficult things with your bot. Like what is it that you don't get? Open up freaking cowork, point it to point cowork or codex to your open claw folder and then just give it all the context that it needs. It writes the PRD for you and everything. Then I don't even. Jack BeVier (46:37) What's up, PRD? Craig (46:39) a product requirement document. So basically it's like, hey, this is a specification that we're working like, here's what we're working on. Here's the specification. Here's like, take a screenshot of where we are now. And the mission for me is this is what the end product looks like. Right. This is our end goal. Now, I don't want to do that in one shot. I wanted to do it in a three phased approach or for, you know, as many phases as you need to get to that fricking bottom line mission. Right. And at each interval, I need you to test Rocky to make sure that he did exactly what he was supposed to do. And dude, it has been insane. Like I watch him spawning up agents on the fly. And then I ask it, well, like we have agents. We have skills. Like, why aren't you using those? Well, you know, like, it's just easier to like, you know. So I don't have it working perfectly yet because I figured like, why take all the time to build these skills out if it's just going to spawn up its own, which I find sort of interesting. But yeah, it's been a real unlock for me, Dave, to just give the whole folder to Codex and then just have it write up like this beautiful, unbelievable PRD with tasks that are well-defined. And then to be able to validate and test with Rocky, because he brings back this perfect report to say whether, no, and it gets it every time. He didn't get it right. No, there's two other things that he had. David Moses (48:01) That idea is going to save me an hour a day. Just that, just basically making it like, here's a markdown file that is your north star for this thing that you're building. whenever you get stuck, whenever you think you're done, whenever you think you've reached a certain step, go back, look at this document, and see if you're on the right track. mean, that's going save me a lot of time. Craig (48:06) What? Right. But don't you, what I find Kludge is I'm working in Codex, which is not inside of my, you you could ease, I don't know why I need to do that. You know, I could have Codex working inside of OpenClaw. I don't know that it's so efficient. And then have my conversations with Codex rather than having to go outside in the Codex app. do all of this and then copy and paste it into Rocky. Does that make sense? David Moses (48:56) Yeah, that does make, and maybe I wouldn't do it in Codex or Cloud Code, but I would still have that hierarchy where you have, you know, an agent, I have one agent that exists on an API token from Anthropic. If everybody else goes down, this is going to revive them. But I could see the idea of one agent that is, you know, that runs on Opus 4.6 that just makes sure that, you know, your rocky mind's auto and forge. Auto and forge, forge builds things and auto is like my personal, you know. But I can see that main agent that's just, you know, keep everybody in the same, you know, keep everybody heading in the right direction, rowing in the right direction. Craig (49:42) What's so funny, Jack? What do you think of Jack? Jack BeVier (49:45) I'm laughing on the inside and maybe it's coming out because we're like, dude, we are zoomed out like so freaking abstract. And I'm like, you know what I did in the past two weeks? I built a web scraper. I'm pretty excited about it. It's like pretty cool. And you guys are talking about like... Craig (49:53) Who left? Jack BeVier (50:02) training the architecture behind your bot to make it more efficient. Like you're just like zoomed out, like three levels up. I'm just like, can I scrape some stuff off the internet? Craig (50:08) Well, when you get a call from the owner of the company, when you get a call from the owner of the company, tell them you've spent like $4,000 in the last 10 days, you probably got to go back to the drawing board at some point, figure out a better mousetrap if you're going to use these skills. David Moses (50:11) and Well, just Jack BeVier (50:23) think it's great. We're just figuring this shit out, you know? David Moses (50:23) a suggestion. KimiK 2.5 and Minimax 2.5, not 2.7, but 2.5, they're both open-way models. And Minimax, I would wager, don't know what kind of, do have a Max Studio? Okay, so Jack, if the underwriter of said tokens were to consider the cost benefit of Jack BeVier (50:45) Mm-hmm. David Moses (50:47) obtaining a couple of Mac Studios or one Mac Studio at about 512 gigabytes, those machines can share RAM and you could actually run KimiK 2.5 and Mini Macs locally for free for as long as... Jack BeVier (51:05) I, so speaking of which I, I got, I got all riled up last weekend and I went to, and I was reading, I was down all kinds of rabbit holes and I got all fired up. And so I went to, I went to micro center up in Parkville, Maryland, where they had, where they have, yeah, I was right up your way. And, where they had 21 Nvidia GTX sparks in, in stock. David Moses (51:19) Yeah, we have a Micro Center too. Craig (51:21) You didn't call me? You're up. I like Micro Center. Come on, let's go. Here we go. David Moses (51:26) you Jack BeVier (51:33) And I bought two because it was limit two per household. And so I have two sparks sitting here that I'm like, I could be running local models on this shit and given, and then I'm looking at my token usage and everybody's token usage. And I'm just like, maybe we should be running. Maybe this was actually an excellent investment and I should actually, I should be doing this. So yeah, I'm, I'm down that path. Yeah. David Moses (51:54) Yeah, the reason I would suggest Mac Studios is because they can share memory. the next model comes out and it's a trillion parameters, whatever it is, and you just need more horsepower, you just add a Mac Studio and now they're all sharing a gig and a half, or a terabyte and a half of memory instead of a terabyte of memory. So it's scalable, I think, in a way that the... Really, everything in the PC architecture just isn't. Craig (52:19) So Jack, you've been, well, you inspired a company to be interested in AI and the company has built a lot of amazing, amazing tools that I can't wait for people to see. But you personally went out and got yourself a little Mac mini. Was it a Mac mini you got? What'd you get? Got a Mac mini, fired up open claw, and you built something really cool. Jack BeVier (52:39) Yeah, yeah, yeah, I a Mac Mini. Craig (52:47) Dave, you did a little show and tell, which was great. We made it, we said that for each episode, we would all come with something that we've done over the past week. Jack, you built something that I thought was really, really cool. so tell us what you built. Tell us like sort of how you did it and then like how much time it took and why you thought this would be a cool use case, jump in with both feet. Jack BeVier (53:09) Yeah, so I got two updates on my homework. So one was just on Claude desktop. just started finally messing around and playing around on Claude desktop. And I wanted to just build like a web scraper that was it didn't have to be a web scraper, but I wanted to aggregate a bunch of public records resources. Specifically, I'm using Baltimore City as my my use case here. And so. I wanted to, I was curious how many illegal rental properties are in Baltimore city because Baltimore has this rental licensing requirement. So if you either, if you have a rental property, you must have a license. And my gripe with the, yeah, my gripe with the, with the legislation when it passed, yeah, what seven years ago was that I was just like, you guys are going to tax the good guys because I'm going to comply. And then the slumlords not going to comply. David Moses (53:48) I love this idea. Craig (53:52) It's a tax. Jack BeVier (54:02) And then you're not going to do anything about it. And so you're just taxing the good guys and not punishing the bad guys. And so this is stupid, right? Like this is having a counterproductive effect that you you desired. If you would enforce this shit, then you know, then I then I would be sympathetic and I'd be like, yeah, let's do it. But hey, seven years forward, fast forward. That's exactly what happened. They haven't enforced anything. Slum words are just still doing exactly what they want to do. And. and I'm getting charged, you know, 150 bucks a year per unit for nothing, you know, for literally nothing. Yeah, between like the rental, well, property registration is 30 bucks and then you have to get an inspection report. You can use section eight inspections as a proxy, but otherwise you have to send an inspector out and do a health and safety. And then you upload your, you then you have to pay somebody to. Craig (54:34) Is that what it is? It's 150 bucks. Jack BeVier (54:55) you know, compile that and the lead certificates and upload them and get your license and then track that. Right. So it's just another compliance thing that we have to deal with. So I was just like feeling, I was feeling snarky and I was just like, Hey, let's, let's go put together a list of illegal rentals in Baltimore city. And so I pulled down the assessor database from open Baltimore and then banged on the rental licensing website, and pulled in all of the. You know had to figure out the architecture but like, know, finally just in just in claw desktop just talking to it You know just talking to it the whole time I'm just I'm just like literally pasting URLs and Telling it what I want and getting closer and closer on an iterative basis to like what I want it to look like built me a little website that's a table of the of the 65,000 non owner occupied properties Populates all the rent the licensed rental properties Populates if the house has a vacant building notice, so it's a vacant property I gave and then I put another category for if you've owned it for less than 12 months, know, like, Hey, let's just give you a pass because maybe you're still in the middle of the rehab and you're going to flip it or you're going to go get a rental license. So let's, know, give everybody 12 months. But then if you haven't licensed it, if you haven't licensed it, it's not a vacant building notice. You haven't, and you've owned it for more than 12 months. Then I tag you as an illegal rental. And then, And I, and then I keyed everybody based off of mailing address, right? And then I started labeling the organization because sometimes people have different mailing addresses or different LLCs. And so I keyed everybody together so I could like see the organization and then Dave, you'll love this. I, and then I created like a bad boy score. So it's like percentage of non-compliance. So if you've, if you've got a vacant property or you've got an illegal rental, so I add up all the vacant properties you own, plus all the illegal rentals do you have? And that's the numerator. And the denominator is how many properties you own, right? So like for every 20 properties you own, how many of those are vacants and illegal rentals versus licensed rental properties? And so it's kind of like a bad boy score. And then you start sorting by that bad boy score and also sorting by like, you know, total number of properties that you own. And you can essentially be like, Hey, here's all the biggest slumlords. then here's, know, and then we scroll down, we get to the smaller slumlords and you can see like, dude. And so, you know, that Baltimore fraud that happened, the appraisal fraud that happened. David Moses (57:09) Yep. Jack BeVier (57:11) Top sorts to the top like the biggest non-compliant landlord. Yeah, biggest bad boy in Baltimore Eli gold so like it is an excellent proxy for like, you know for like good landlord versus bad landlord. That's kind of like the idea behind it and ⁓ David Moses (57:11) All not ready. Craig (57:14) Biggest bad boy in Baltimore. Jack, know it was like, love the snarkiness of the project, like, tell it, let us in. what's really, what of all of the things that Jack Baviera could have built and not seen his wife for the entire weekend, what, why that project, Jack? What is, what is, what's the use case for you there? David Moses (57:30) That's it. Jack BeVier (57:50) I don't know, man. I mean, part of it was just like, was just like, you know, I don't know. It's like real estate nerd porn, you know, like just to like, cause I, cause I also layered in, I also then was just like, well, I'm not done with that. So let's layer in the housing market typology so you can see the quality of the location. Let's bring in the permit data going back to 2015 cause that's publicly available. And so now I can see for all of the non-owner occupied properties. David Moses (57:54) curiosity. Jack BeVier (58:16) how many permits have been pulled in the past 10 years and what was the total self-reported value of all the work that was done. I then banged on the Maryland Department of Environment website for 40 hours straight and pulled in every lead certificate. And now I have the lead cert expiration date and the inspector's name and the inspection company. And so in real time, I run this script, it'll take a day, right? It'll take basically like a day and a half for this script to refresh. And so have it set up to run once a month. And so I've got a little use, you know, I got a little tool here that is, it is my picture of every non-owner occupied property, right? Like every investment property in the market that I care about, which is Baltimore city. And it's the aggregation of all of the available data sets in that one place. So if I have a question, right, about what's going on, who's doing what, who, you know, who's good, who's bad, what kind of stuff are they buying? It's all. aggregated together. next thing I'm going to do, so I had this idea last night. I'm like, all right, now that I've aggregated all this data together, why don't I point my open claw at that aggregation and be like, hey, start creating me some articles. I want to write up, start analyzing this data for me, right? Because it's 65,000 properties to scroll through. a little bit, you can do some sorts to get some interesting surface level stuff, but like, Hey, show me the new investors in town who are growing at the fastest pace, but clearly have some operational issues because they're not in compliance on, show me the most aggressive guys who are coming into town and don't know what they're doing, right? I can answer that question or like, whatever housing policy or curiosity or market, hey, I got a property that I want. I got a shell that I want to go sell at this address. Craig (59:53) Yeah. Jack BeVier (1:00:04) look up all the buyers of those address of properties of similar type over the course of past 20 months and give me the top 12 people who I should, you know, send a text to and give and pull in their cell phone numbers too, because you can pull in a lot of contact information off of the rental licenses. So like it just became like, you know, hey, I'm just going to dump all of my IP about Baltimore into this one place and we'll use it as we'll use it as a diligence tool at the courthouse steps. Craig (1:00:26) Mmm Jack BeVier (1:00:30) I don't know. It's like if someone's offering us a property or like, you know, we see, we can pull up the address and know everything about that person's situation. And, um, I could, I could add mortgage data next. I'm thinking about that, but anyway, yeah, I probably, I probably have like 20, 30 hours into it. Craig (1:00:43) You could, Jack, you could, you you could, you could add like a house canary or a data tree to like give you the actual, you know, property specs and everything. what I love, Jack, is that you start off with this grand use case and, and this amazing amount of intellectual property that you've, you know, of the past 20 years of experience in the market, but now, but now that creative curiosity, inside of you that goes, shit, man, I'm laying in bed and I just figured out a feature that I can add to this thing. And the beauty to be able to just do it, that that's to me, the real unlock, for this technology. So, yeah, I can't wait to see what you do with that. That's, that's, that's, by the way, by the way, ⁓ by the way, Dave, he showed it off, just briefly and, ⁓ it's really, really cool, man. Just a great dashboard of, Jack BeVier (1:01:26) Yeah, I put the link in the chat. David Moses (1:01:32) I want to see it. Craig (1:01:37) a lot of Jack BeVier (1:01:39) Dude, do it for Detroit, Do it for Detroit. David Moses (1:01:39) So I wonder, yeah, so I wonder, the two things I would think of to kind of layer on top of it are number one, if crime statistics are accessible and you can overlay that and see. Well, the stats are bad, but but like, if it's anything like Detroit, you know, you've got, you know, great neighborhoods and you've got. Craig (1:01:50) Well, I mean, it's Baltimore. think, I don't think we need any stats. Yeah. David Moses (1:02:05) garbage neighborhoods. So by overlaying the crime stats and possibly, the Google Street View API is like 7 tenths of a cent per image. if you can overlay, look, this is what this neighborhood looks like, and this is the crime in this neighborhood, and here are all the unregistered rentals. So city, this is not a coincidence. Craig (1:02:16) That's really cheap. Jack BeVier (1:02:28) Mm-hmm. Yeah, yeah, yeah. And by the way, if like an... David Moses (1:02:29) Look here. Craig (1:02:31) Jack, are you snarky enough to start sending reports to the city? Jack BeVier (1:02:35) I was David Moses (1:02:36) No, I wouldn't even go to the city. I would actually go, if it were me, would go to a newspaper. I would go to a newspaper and I would say, here's the statistics that I've got. Your name now is everywhere and it's like you were the one that uncovered this. no publicity's bad publicity. And then let the newspaper take care of the work for you. You don't even have to go talk to the city. Jack BeVier (1:03:05) Yeah, there's a lot of snarky shit you could do with it. Give it to the paper, give it to the Sun Papers or give it to the politicians or give it to the city and be like, hey, you're doing code enforcement really inefficiently and shittily. Here's the way to do it. Here's all the guys you need to go after. Anyway, I mean, it's also ironic because it's all their data. I'm just pulling in. I'm going to all .gov websites. David Moses (1:03:24) It's nothing they didn't already know. Jack BeVier (1:03:28) to pull this shit in and aggregate it and then like draw conclusions from it. Yeah. So they'll catch up. They'll catch up. David Moses (1:03:34) What was that line from Office Space? Thumbs up their asses. Jack BeVier (1:03:39) Yeah, that was so that was yeah, but my mind's just a bunch of Python scripts and a simple index dot HTML. So you guys are like rearchitecting how these things are thinking. And I'm just like doing Python Python scripts, scraping websites and putting it all together in an index dot HTML. David Moses (1:03:54) It's a really awesome music. Craig (1:03:56) It's a great use case though, fantastic use case. Jack BeVier (1:03:58) let me I got one more. I got one more incredible use case. This one blew my mind. I just figured it out this past weekend. So we lend money and sometimes guys don't pay. And so they're like, you know, there are bad borrowers. Right. Like and I wish these never guy. I wish these guys never got in the front door. But now that, you know, now they owe us money because they haven't paid us back and we got a judgment against them or whatever. And you want to know what they're up to, dude. So I put in. I went into open-claw, put in what we knew, know, just like names, LLCs, mailing addresses and said, hey, go, and they owe us money. I want you to be a private investigator and I want you to go scrape. I just want you to go everywhere on the internet, all public records databases, like track these guys anywhere you possibly can figure out known associates, run down those rabbit holes, just rabbit hole after rabbit hole after rabbit hole. And I just said, just go. And then 15 minutes later, it comes back with the most amazing research report I've ever seen in my life. It was like I had a team of team of expensive lawyer associates working for a month on this thing or a team of private investigators to put together this massive report. And then I read through it and dude, in I put like eight guys into it in three of the cases, it totally cracked the case. It like told us exactly what's going on. showed the conspiracy, explained the pattern of behavior. And it's the kind of thing where like if I had been banging away on the internet trying to research these guys and figure out where my money went and I had a room full of whiteboards, I'd have filled up the whole room. It would have taken me a month to fill up the whole room full of whiteboards of my research. And then like, could my brain have been smart enough to connect, you know, that, yeah, last Tuesday, I saw this pattern happen over here. And then it was over here on this wall too. Craig (1:05:43) Yes. Jack BeVier (1:05:48) I can't keep that shit all in my head. Dude, this thing just put it all together in an amazing way. Craig (1:05:54) Jack, did you give it any websites to go look at or did you literally just say go down rabbit holes as far as you can and find whatever you can? You didn't give it any context. Jack BeVier (1:06:02) I just said, go didn't give any proprietary database context or anything like that. And then when it came back, I said, Hey, are there any unanswered questions or additional rabbit holes that you should go down? Go down them. Don't stop until you're done. Everything come back 15 minutes later comes back has cracked all those like eliminated a bunch of things that weren't weren't correct and then prioritized it and then turned it into a set. And then I was like, and turn it into a set of action items. Craig (1:06:07) Wow. Jack BeVier (1:06:30) my team to go pursue basically and then send me an email summary and attach in a markdown file all the backup that you found and where you found this information. And it's like, dude, just forward it to the attorney, right? Like here's all the backup for whatever filing you want to make. Like here's all the legal research behind it. Here's what we've identified. And it seriously in three of the cases like completely cracked the case. David Moses (1:06:54) Why don't Craig (1:06:56) Wow. Jack BeVier (1:06:56) Yes, seriously. And then right there from fucking plate. I'm like, dude, I'm to go to law school just so I can sign the bottom of the damn, you know, filing like not actually. ⁓ So anyway, that was like it, the just the research, right, which is a very basic thing, but. Just open claw. No, just open claw. That was open claw. David Moses (1:07:02) you Craig (1:07:03) Right. David Moses (1:07:04) you Craig (1:07:08) And Jack, you do that in Claw Desktop? You opened up your Open Claw agent and you just... I think sometimes... Sometimes I think it's the guy who opens up Open Claw. Your agent is just inherently smarter. It's Jack Bevere. He opened up Open Claw and his agent is just inherently smarter for some reason. Jack BeVier (1:07:19) Scent it down, rabbit holes. Brute. Opus. David Moses (1:07:31) He's smarter than my agent. Jack BeVier (1:07:32) You Craig (1:07:33) I don't think I could tell Rocky to do that frankly. I'm blown away. Jack BeVier (1:07:35) I think you could. Dude, I think you could. was literally just that simple. David Moses (1:07:39) Just the way he types it, it's like, Opus knows, alright, this guy's not fucking around. I'm gonna have to work a lot harder to impress this guy. Craig (1:07:44) ⁓ Wait a minute, I got Bevere banging on me here. Wait a minute. Jack BeVier (1:07:45) What? Well, so like, and the thing is I probably, looked at the token usage because it's just pounded away in there, right? Like, and I looked at the token usage and I'm just like, I probably spent like 20, 30 bucks per one of these reports, but it's like, it's like thousands of dollars, you know, it's, like, it's a, it's a work product. I probably couldn't even pay for it. Like it's the best work product I've ever seen of this type. And I'm like, it costs like 20, 30 bucks and it's got a real definable concrete business use case. And I'm like, I mean, I get off on that stuff, right? Like, yeah, pure like this is clear ROI, you know, clear. Craig (1:08:24) Not only Jack wasn't on the AI call this morning that one of the guys on the call said he developed something to was it taking the data that you gave to him? Who was it that was talking on the call this morning? He developed like this great use case where he can go out and find the leakage I guess for the folks that aren't paying. You don't remember that on the call this morning, just two hours ago? Nevermind. Jack BeVier (1:08:47) leakage? Craig (1:08:49) Doesn't matter. I'm wondering if you gave the report to anybody here at Dominion and now they've taken it to the next step. Jack BeVier (1:08:57) no, no, I just gave it to the special assets guys. Yeah. Craig (1:09:03) Well, gents, we're into this thing for an hour and a half. Jack BeVier (1:09:05) Jose was talking about that. Yeah. Yeah. Yeah. But no, no, he's that's he's doing his own thing. He's doing his own thing. Yeah. Craig (1:09:11) Separate project, got it. All right, boys, I could talk all day. I had had a few more things I wanted to talk about that came out at the NVIDIA conference, but we can save that for another time. OpenClaw has obviously taken over the world. It is the fastest open source project, the fastest growing open source project since Linux. It's far eclipsed in terms of usage. So I think we're on the right track in terms of talk points here. with OpenClaw. Dave, any last comments from you? look like, Dave's wondering how he's going to leverage our conversations today. And he's, it looks like he's already doing it. David Moses (1:09:47) Well, just, you know, last time we each, you know, kind of made, I want to keep the theme going if we can to say, okay, give me something you're going to build before we meet again. And so we can hold each other to it. Jack BeVier (1:10:01) ⁓ Craig (1:10:01) I'm going to, go ahead Jack. Jack BeVier (1:10:04) I have mine last time was that I was going to figure out how to bring open claw into a sock to compliant sock one to compliant framework. I have been working on that. I do have progress on that so far, but I do have more work to do on it. So that'll be my that'll be my update for next time. I'm actively working on figuring out how to do what I realized here is here's what the at the risk of elongating things. It was when we go and we raise debt. Right? Like in our lending business, like we go, we just did a securitization. I just, we just finished it. And part of raising that debt was they want a lot of the bond buyers wanted to know, like, you sock one to the sock one type two sock one to compliant. And it's like an important, it's important for them, you know, like it's a, you know, some like sovereign wealth fund, give you lending you money. They want to know that you've got security in place. You're not a bunch of cowboys. And so. The debt requires SOC compliance. But from an innovation perspective, like this research tool, for example, like I put some names and addresses and stuff in there and I'm like, shit man, I should be running everyone who applies to for a loan through that for 20 bucks. I can do a deep dive on these guys. Like I should make that part of my screening process. But now I'm like, you know, but now I've, I'm intersecting customer data with open claw and that's currently not SOC compliant. So like my debt requires SOC compliant, but my equity as the owner of the company, my equity requires AI innovation. so like, I'm going to just be pounding on the conflict between SOC compliance and AI agency, you know, and agentic AI. And I'm, I'm up, I got, have like a SOC compliance. consultant and I'm talking with them about these issues. I'm talking with my IT company about it and I'm just shoving everybody. I'm just push. I'm just a pusher, right? So like I'm like, hey, I'm not like don't know. Don't tell me no. We're going to feel like we're going to do both of these things at the same time. And no is not on the menu. It's I know it's hard. I know it's new. I know we're trailblazing here, but we must figure this shit out. So that's been David Moses (1:12:08) Thanks. Jack BeVier (1:12:14) That's been a big chunk of the journey for me the past couple of weeks that I am also still on. So that'll be my update next time. Craig (1:12:21) Dave, have you let AI, your bot, loose on your inbox, Outlook, or whatever it is that you use? So yesterday, I created a technical brief that would hopefully be used by our IT guys. And all I'm asking for is for the bot to be able to read and write but no send on my Outlook mailbox. David Moses (1:12:26) Yep. Yep. Craig (1:12:46) So it can go in there, figure out what's important, what's not, move certain things to folders, create reports on me for like, you gotta get back to this guy. You gave this guy a quote, he never got back to you. My whole idea is I wanna spend as little time and outlook as possible and see only the things in front of me that are important, right? And so I ask for access for that. with our IT folks. Evidently, they're not too excited about opening up access to the mailbox for OpenClaw. But what I thought was, hey, look, I'm literally giving you a... Shocking. ⁓ David Moses (1:13:22) I can't imagine why. Jack BeVier (1:13:24) Yeah, but I'm so, dude, I'm so jealous of you. You've got it hooked up to all your stuff. You got it hooked up to your QuickBooks. I'm like, maybe, I'm gonna compartmentalize my company. Craig (1:13:32) Yeah, mean, so like, not only that, like, imagine what about what about if every loan officer had their email hooked up to to what can we learn from that? You know, how much how much leakage are we are we just are we losing in customers that like we don't get back to or don't get back to us? Like all of those things that we could learn just by having more data being being analyzed. And so I said, look, here's the here's the compromise. I don't want this thing writing and sending emails from me. It can write one, but it cannot send. So there's a server side setting in Microsoft 365 in their open graph that you can set this setting and it will absolutely not allow it. I can't change it. I mean, it's server side, right? So created a massive like 15 page technical brief that hopefully someone can read and consume and make some sense out of. Made a beautiful brief graphs. pictures, everything, and gave them all of the steps that they would need to take. Security. I even put a SOC section in there for consideration, so we'll see what they say. that would be a real game changer to have that ability in. I had it, Dave, I actually had it running with Apple scripts. It would go in and open up my email and then sort of do like a scan, and it would give me a report, but it wasn't great. And I think this would be real unlock. we could get that done. David Moses (1:14:57) So maybe, maybe it would be okay if you mirrored your inbox into something like SuperBase and then just worked with that database. Now, you know, it can make, you know, I mean, now I can see everything, but it doesn't have to access Microsoft to do it. Craig (1:15:13) That's bad. You're a bad, bad man, sir. Dave. That's brilliant. Jack BeVier (1:15:22) Pretty fucking good idea. Craig (1:15:24) That's pretty fucking- David Moses (1:15:26) I didn't think it was that good of a idea. Craig (1:15:28) No, it's really good with Dave. It's one I didn't think of. Jack BeVier (1:15:28) I think it's great idea. David Moses (1:15:31) Alright, Jack BeVier (1:15:31) I mean... David Moses (1:15:31) so want to hear what I'm working on by next time we meet? Okay. This... I'm going to be really excited if this works. So I've had this software, I think I mentioned it last time, I've got this software called Insightful on every... at everybody's machine. Alright, everybody has to... I don't care if you're commission based, it doesn't matter. You have to have Insightful. Craig (1:15:35) Yes. David Moses (1:15:51) on your computer and it runs in the background. takes screenshots every, I don't know, minute, right? And it tracks every website you visit, every program you use, how long you were in it, whether you were active or inactive, all this data. And I've collected since July 25, I've collected seven million fragments, right? And a fragment is basically just they did this thing for this amount of time. Here's what it was and there's a number of screenshots depending on how long it was. Craig (1:16:09) You David Moses (1:16:18) But what I'm going to do is actually take and finish out my where the hell is our time going dashboard. the idea is that it will, Google has a product that's like 1,300 thousandths of a penny or something like that to OCR a screenshot. Here's a screenshot, bring all the text. Now I can use a very cheap reasoning model to assign a major and a minor category to that activity. I already know how long they spent on it. I already know what website they were on. If they were on a website that's specific to a specific record in one of our software systems, I can programmatically determine what record that was. But now I can actually see what they're doing with that record. So I'll take a midpoint screenshot of every fragment. and analyze it, and then I should be able to tell in a dashboard, drill down to the person, drill down to the division, drill down to the activity. I let it invent the major and minor categories. It's got 15 major categories, and I think it's got up to 65 minor categories. And it'll categorize every single thing that it can. And I already have the categorization done. I went back through and did all of our phone calls, voicemails, and meetings. I haven't done text messages yet, but I've done the rest of those. And I'm hoping that I have a dashboard, literally, that tells me the cost of each minor category activity. And now I can rank them, and I can decide, well, we're spending way too much time on that. ⁓ Yep. Yep. Jack BeVier (1:18:01) Mm-hmm. Yeah, I'm not making any money. I'm not making I'm not charging enough for this. Yeah. David Moses (1:18:07) Yep. Actually, the idea really stems from something that Fred's Fred Lewis's sister told me. You know, she said, look, get everybody in your company an egg timer. Have them have it sit next to their desk. And every time they're going to do something right that takes longer than 15 minutes, have them, you know, write down what they did and and and, you know, put an egg timer on it and report back. you know, much time did you spend on it? And then, you you'll aggregate this data and you'll be able to look at it and say, this is where these are the time sucks. So this is, think, a little bit more of you know, AI approach to that, where all they have to do is just go about your business, do what you do, and we're going to analyze all these activities. I can, for me, cross-referencing them against Jack BeVier (1:18:49) Yeah. David Moses (1:19:00) our pain points from our client's perspective. Client is upset about this thing, customer's upset about this thing. Looks like we don't spend much time on it, or looks like we're spending way too much time over here. I can really start to get some, I think, business use data. I'm not gonna say it, I'm gonna admit that we'll have done anything with it, but by the next time we meet I want accurate. down to the fragment categorization of a high percentage, let's call it 60 % or higher of the fragments have a major, at least a major category. Actually, I would say 90 % should have a major category and then 60 % should have a minor category. Craig (1:19:40) Love it. Love it, love it, love it. Jack BeVier (1:19:41) That's a great one. I love it. really like really engineering, you know, exactly what's going on in the business, right? Like, you know, God, God mode for like, what's going on in your organization. Yeah. David Moses (1:19:56) Yeah, that's be a, but again, like, and I've explained this to people, like, I don't want to micromanage you from the standpoint. It's like, Oh, you're goofing off. Oh, you're watching this or you're wasting time on that. That's not really, you know, you get your job done. Cool. You know, Jack BeVier (1:20:05) Yeah, yeah. Yeah, want to I want to the people are good. I want to I'm evaluating the mousetrap, not you. Like it's it's the mousetrap that I'm trying to analyze. Yeah. David Moses (1:20:20) Yes, correct. Craig (1:20:23) Yep, I love that. Well guys, another second episode. Still batting a thousand on content here. Loved it. Could talk all day with you guys about it. Next time I will be wearing a blazer. I'm gonna get something probably more a tweed with arm pads. Make it look really nerdy. ⁓ yeah, maybe, maybe if you're lucky. David Moses (1:20:43) arm pads. Will you get a pocket protector? Craig (1:20:49) I might have that. So I just want to thank everybody for joining us. Love your comments. Jack, we actually had a couple people that got back to me that said they wanted that spreadsheet that you developed for, I think we talked about on the other podcast. But in any event, folks are listening. They're commenting. I love getting emails from folks, Craig at thedominiongroup.com. Feel free to send them. These guys are busy, so you can just send them to me. I'll make sure if you want to talk to them about anything, I'll get it over to them. But hope you enjoyed the podcast today. We'll see you on the next one.