The Adam Sand Report - Thursday, June 4, 2026

What's going on, friends? So we're doing another live tonight, and I— if you notice my voice is different, it's because I am sick. And so I'm gonna do my very best to be half-ass entertaining. That's why I do these live streams at night, because I fully expect them to suck. And so this way, uh, I'm not— I know I'm not wasting anybody's prime time, but if you're here, I appreciate you being here. And if you join What you're gonna be learning today is we are gonna be setting up a, something new, or essentially the business primitive and base layer of any great future AI system is going to be setting up the ingestion and knowledge base, the state control of your business. And I've never done it before. I have a CTO in a company that I'm involved in who built something like that, and what he's built it, It's not for my business, not for RBP, but I'm going to— I forked the repo and I'm going to work on it tonight, and you're here to work with me and enjoy it with me. And that's about the most entertaining I can do, is kind of show you something that I'm working on. You can kind of hang out and work with me. So if you hang out and you're here, I, uh, I will— that's what I'll be working on together. Hopefully you enjoy that and you get value from that. And so without further ado, I'll hop right into it. But again, I am super sick today, and so I'm not, uh, I'm probably not going to be as entertaining as I would like to be, as you would like me to be. So I'll do my best, and, uh, if it's not good enough, well, swipe up, swipe down, go check out the next thing, and maybe I'll do better next week. All right, so I have all my files here, so I guess I might as well walk you through what it is that we are working on. So I'm just going to load all of this up and get rid of some, some stuff that you don't need in your way. Put this over here. Okay, so what I'm doing is I'm just moving some screens around so that you guys can actually watch this and I can still keep my eyes on the comments. Um, all right. So now we have this, of course, and it's going to show me this. All right, so what I've been doing, and it's— I have them over here. So my roof plot helped me design this. So I guess I'll share this screen. Share screen. Entire screen, screen dos. All right, so what I had it doing is, uh, I got it to fork a repo that our CTO put in, and they're right here. And so I essentially want to see, okay, now where did it put it? These— now I put it all in the Foundation platform, but I wanted to fork that repo, so I gotta go find They get repo over work. All right, organizations. So for those who don't know what this is— man, I feel so stupid and slow today. Which repo was it again? So if you're just hopping in, we're we are working on building a, a essentially an ingestion and operating system for an AI native company. And so we're essentially building a place where multiple agents treat as a, as a separate single brain. And the idea is that we're going to be doing this with AWS Pinecone. We're essentially creating a a knowledge layer that stands above any agent, and that this, and that this environment could be a decision gateway for, um, any agent to understand the current state. And so this is a fork of another repo for a company that I'm involved in. I'm doing this mostly as an experiment, and I have to get all the files into this repo so that because I have all the PRDs and everything, and we say I need those files moved over to this repo so that we can use it in our plan. So this is it. So I have several Hermes and OpenClaws, and so like this is an OpenClaw. You know, you'll know the difference because the OpenClaws are all like lobsters and Hermes agents are all, uh, They say it like kind of like stoic techie. I don't know what you would call this, like an Android stoic dude. That's what I would kind of put it. So right now what I'm getting him to do is to move this over. And the idea is that, um, the problem that we face, or any business really faces, is that when trying to, when trying to move to AI Why does AI not get success for a lot of people? There are business primitives and there is a current state of your business. There's tacit knowledge that exists and your agents are not aware of that. So your agents, you can train a skill, right? Like if you're, if you've been around agents for a while, you've probably got into the point of, you know, directives and executions, Python scripts, building agents, agents that have skills, and then you can use slash commands and you're basically hand rolling this thing. And it's so empowering because when you get it to work for you, that's awesome. But then you want to give it to other people, and then that's kind of where the first breaking point is. You try and give it to other people, they didn't build the skill, they don't have necessarily the environment variables. So you could upload all the skills, but then they don't have the environment variables, so then they, they don't have the API keys, so then the skill doesn't work the way they want it. So now you're trying to deal with like multiple API keys being out there, and then you've got all kinds risk. The other problem you run into is that there's a knowledge base right there, like, and we're not talking like just shoving everything into Pinecone or something like that, but you're gonna have multiple systems. You're gonna have Slack, you're gonna have, um, you know, you could have, you could have Slack, you could have HubSpot, you could have your roofing CRM like, uh, you know, RoofLink or, or, um, AccuLynx or something like that. And then you're gonna have CompanyCam where your photos are going to be, you have QuickBooks where your accounting is.. And all this stuff exists all over the place. And the easy analogy that we like to use is, okay, well, if your customer asks for a refund in 3 different places and you have agents that are, you know, checking for these kinds of emails, and you have an agent that's watching your email, and then that gets sent to the refund agent. So 3 different places the customer asked for email, they asked for it on the customer service line, they left a message, and they, you know, chat through the website. Well, then what tells the refund agent that it's not allowed to process the refund the second and third time. There's a certain kind of state control that has to exist. And so when you're building AI systems that you're going to want to actually run long term in a business, not only where humans can interact with the agents and have their own agents and agent skills, but then also agents can take action, it needs to understand the current state of business. It has to have some state control, and it also has to have access to like tacit knowledge exists in the business. So the first step of that is sensors, and we're doing a pretty good job of that right now, but I know we can do better. And so what I'm building essentially is the sensors, right? So that everything that RBP does is going into a centralized place. Now, in fairness, I have a CTO who knows how to do this. I don't necessarily know how to do this. I fork this off of another product that we have purely for my own learning and for your learning. Claudio, which is one of my open clause, he forked a repo that is for another project to say, hey, you know what, we've built this for another company, much larger company. I want to build this for RBP. Jose, that original guy, I forked his repo and I'm going to experiment with that tonight. I'm just getting all of the files put in here so that I can say, all right, It's all in here. I need to say, I want to instantiate a new coding environment that is going to help me tackle this repo and walk me through, work with me and co-work with me to get this— sorry, to build this together today. I want you to structure my first opening prompt. I'll be using Cloud Code. And so I need you to give me the prompt to instantiate that environment to get it to connect to that repo, and then give it the initial prompts that understand what's in there and what the goal is, so that it can guide me through the first steps, because I've never built something like this before. So for those of you who are new or have never been here before, I don't, I don't do these lives to be like, you you know, where I'm going to go and say, oh, Opus 4.8 destroys everything, and I'm gonna go, Opus 4.8, you know, model release, and then I'm going to come in here and I'm going to say, oh, what's, you know, introducing Claude, Opus 4.8, and I'm going to read the model card to you and talk to you about benchmarks and all that kind of stuff. I'm going to take you in a live— it's very different. I want these to be interactive, so I usually like you guys leave comments. So for those of you in the room, let me know who you are, where you're from, what you're looking for. And then I tend to try and take these as an opportunity to build something that I haven't built before. So it's still productive time because the whole idea is I want to turn energy, especially when I'm sick like I am right now, I want to turn energy into context. That context turns into tokens. Those tokens turn into economically valuable, meaningful work that generates revenue and profit. So this is a project that I've had on my plate that I'm like, man, I gotta learn how to do this. I want to know how to do this. This is like enterprise-level AI. This is something that all the my heroes talk about, and I'm like, I gotta get this skill under my belt. Like, I want to understand this primitively, and so I'm going to take a crack at building it. I have no idea how it's going to go. You're going to watch it like come together. So that's something you're gonna— you're gonna dig. Throw something in the comments, tell me what you think about it. If you're like way better than me, that's super cool because then you can help. And if you're new and this is way above your head, well, I mean, ask questions and I'll try and catch you up because other people were kind enough to catch me up. I'm lucky to have, you know, a team of people around me that know this kind of stuff even better than I do. And they teach me a lot of stuff. Alright, so what are we gonna do? Alright, so now the real shape. Alright, so you're going into this branch, okay, your role, Senior Architect, yep. README orientation says yes, yep, yep, yep, yep. Okay, back. So he knows you're spinning up Arctic's work on top of his build system, find something. Well, these commits won't be landing in his repo, this is going to be landing in this other one that I set up to test. Unless there's something I've never heard of, this is a fork that I'm doing. There's nothing going to be committing to the other repo that we cloned this from, correct? I don't think there's going to be any kind of issue there.

But yeah.

What was that sound?

That was weird.

I heard a weird sound. I don't know, some kind of, some kind of ding. All right, nothing will land in the source repo. What you're building now, git clone, okay, every attempt, okay, fork is yes, I understand. Oh, I think I know what I'm seeing. Yeah, these are, these are our docs. Yeah, go ahead and merge that branch onto main in the, let's say, this repo. This is a safe repo for me to be working in. This one doesn't matter. It's literally just me. And it's literally just a test. So you can do— we can do whatever you want in here. But I'm going to take this over to Cloud Code.

Okay.

So to make it so it's not boring, we will do We will experiment with CMUX because this is new and cool. So CMUX is something I found the other day which is pretty neat. Ah, there's an update.

Damn it.

By the way, for those of you here, I'd love to hear where you're from, what you're here with, what you're here for, what you want to see. I'm super sick, so I'm like kind of half out of it tonight. I've had a fair bit of the Buckley's. I don't know if you guys have Buckley's where you're from, but It's, uh, I'm a little tired today, but I had it on the calendar and I was like, God, I can't, I can't bail, you know, you got to be consistent. So here we go. So the reason I like Cmax is because you can open up multiple plots on here, right? So if I go Command+D I got another one. If I go Command+Shift+D, I got another one. And if I come back here and I go Command+Shift+D, I got another one. And so I could have my Codex here. Actually, I don't have Codex on this one, but I could have another Claude here. And then I could have another Claude here. And so I could have Claude working in 4 different branches of work trees. So we're not quite there yet. So I'm gonna close all these down, but that'll eventually come. But you could also, in addition to that, you can open up a— you can open up a browser, right? So you have a browser here. You can go Command+D and then change this into a browser, right? Then you could have whatever you want, Vercel on here. You want to work on something in Vercel over here, then this can control this. It's cool. It's exactly what Codex is essentially, but you can use Cloud Code in environments like Codex. Then this thing can actually take control and open up another terminal and do something else. You can spin up an agent inside the terminal. It's cool. No, it's, it's nice to have you guys here, but I just hate when you guys' fingers are broken because I got no idea what you guys— who you guys are, what you're about. If you are new here, um, and you wonder what I do, I help a lot of roofing companies grow. Uh, one of the cooler things that I have been doing lately is this thing called Roof Cloth. Roof Claw, which is essentially a custom-built Mac system that, uh, has— that I build a Hermes and a, um, Hermes and an Open Claw. Oh my gosh, yeah, Hermes and Open Claw onto the machine, safe, secure, you know, built, ready, shipped to your door, installed with the two agents on it. And we customize it to your system, build your skills, connect three— up to three APIs. You know, basically it's like your way to get a really effective, properly built, safe, secure, um, uh, uh, agent harness built on a secure machine for you that you can connect to. You know, we connect to Telegram, we do a bunch of intake stuff, then we do some onboarding with you. And it's kind of like a higher level AI training disguised as a product, because really what the real main benefit is that we teach and train the skills.. And so the whole idea is we, we build this, and now if you don't want to buy it from us, we also put the complete— we put the complete blog on how to, on how to do it yourself is here, from zero to Honu. You can get that for free. And so this is a blog for you to completely understand how we build Open Claws. And the nice thing is this is kind of a meta example of what Open Claw can do. So the whole idea was that This— oh, my OpenClaw Vito actually built this, right? And so it built it, it runs the chatbot, it answers the phone. It's a total meta example. It runs its own CRM. Like, if I go to rootclaw.com/admin, I can log into it, and it has its own CRM, and it, it, it writes the blog posts, it runs its own X profile. And it helps me run, run the business. We have it connected to AgentMail, and I'm not paid or sponsored by any of these people, but, um, you know, it runs on AgentMail. So we have— it can run its own email address, it can spin up— interesting that it's going slow, but, uh, our mayor's agentmail.to. I forget what it is. Yeah, there we go. So it runs on AgentMail, and then we always build it with SafeSecure. So we put Tailscale on the machines so that it's all set up ready to go. And then we also set up a number of the safe skills from ClawHub. So we make sure that the machine comes with a skill builder skill and a skill vetting skill and a bunch of other stuff that's good for safety purposes. And we all, like I said, we build both the open Claw. See, you can tell I'm sick. I can't even type today. Um, we build it with OpenClaw on the machine as well as Hermes agent is on the machine. And so we build all that, and then if you ever want, um, to like have help with it, we also build a lot of skills that we do for free. So if you go to our Roofing Business Partner organization, we have a bunch of like free skills and upgrades that you can use. Like, we have the— should put it as you search— we have the ProLine skill. So this is how you can make a connect to ProLine. Or we have RoofLink, I think it is. No, maybe not that. We have one for meeting recordings right here. We have, um, oh geez, what else do we have? Uh, oh yeah, we have a good local service settings. Local Service Ads. So you have an agent that'll answer the phone and work with Google Local Service Ads. All this stuff is free. You can download it, just point your Open Claw at it, you get it free. Um, the whole idea is that we just try to make it so that we make it as easy as possible for the roofing industry to start adopting AI. That's the whole point of being here, and that's why we're doing this today. All right, coming back over here. Okay, so we're building a single flight, and I guess the concept that we're building today is this single brain. Okay, and so the idea is that the first thing you want is a sensor. I wonder if this, uh, probably a bad one. Um, uh, let's go AI native company. First step is sensor. Y Combinator. So probably explain it. This is not the right video, but there's another one with a— by Y Combinator. But the whole idea is it says that when you want to build an AI-native company, you have to start with a sensor. That was the thing I was trying to show you, is that, yeah, this is the video right here, How to Build a Self-Improving Company with AI. This guy is like— that's the speaker. And the whole concept is that first you have to build that sensor that absorbs all the things that are happening in the company. Then you have to have inference, right? So you have to say, okay, well, what, what is constantly happening? So anytime an agent can go access it it has the current state. The other thing you need to do is then start training, right? Because it has to learn the business primitives. And what I mean by a business primitive is if you teach your kid how to drive a car and you say, hey, okay, I want you to— like, you're teaching your kid, you take them to a parking lot at like 9 o'clock at night, you say, okay, I want you to accelerate to 30 miles an hour and then I want you to stop at that stop sign. Well, they're going to push the gas too hard, they're going to they're going to hit the brake too hard and then you're gonna have to like feather the gas, feather the brake, you know, push a little harder, push a little more, push a little less, okay, a little bit. And then when you know how, when you go to stop, you have to lift your foot off the brake pedal and then settle it in at that like last, you know, 6 inches. Well, all of that is stuff that you inherently understand. When you go to train your kid, it's like you don't say, okay, I want you to apply 30% power to the accelerator for 3.5 seconds to accelerate to 30 miles per hour. Then you're gonna coast for 34 feet, then you're gonna apply 13 pounds of brake pressure, gradually reaching up to 18 pounds of brake pressure, and then at the last 4 feet, you're gonna ease off the brake and then apply maximum pressure within 3 inches of the stop line. Like, you're not gonna describe it that way. And that's because you don't know, you don't know how to teach someone how to drive when it's become instinct. And that's the same way when it comes to teaching AI. And so if you wanna create these like hybrid workforces, where humans in the company work with agents and they're going to have like things that work in Slack, right? So if I go here, like we have agents in our Slack, like we have Claudio and Marcus and Vito, and these agents, like people actually work in the company with them. So if you go like— actually have to be very careful here, one second, to make sure I don't expose an API key. So where do I go? Yeah, okay. Um, actually I can show this one. So I said, okay, earlier I asked, I said, hey, Tara said you're sending me something about signing up for the social media profiles. And then he said, yep, Tara confirmed this, she sent me, and he sent me over the files, right? And so I, like, she's communicating in Slack. You can see here, like, got it now. Checked our— pulled our Slack DMs, and then she— like, he was chatting, but she was chatting with him in Slack, and then she got back to me, or he got back to me in Telegram. So that's what I mean by building an AI-native company. All right, what do we got here? So it is read the PRD versions 0.2 strategy doc done, now the code layer schema Terraform all Lambdas. Okay, so it's getting an understanding. Okay, so it's writing, so it's reconciliation memo, just two design layers.

Okay. All right.

So now we start to get our project plan here. All right, so what it read, quoted back. All right, one brain, four entities, Stack already declares the conflicts, Supabase migrate to RDS, plan Kona migrating PG Vector, points to architecture.md, which doesn't yet exist. 80 AWS build, Cognito, SQS, Lambda, Gemini batch. Yep. Okay, Jose builds identity first, then ingestion, then Supabase. Okay, week by week, literal commands. Got it. The conflict is narrower than it looks. Identity kind of exists only in spec. That's the only actual coded conflicts. Nothing else is currently— is clear. All right, which bundle Okay, so we're gonna go like this. Shift+Tab, you're going to go to plan mode. All right, so what I need you to do is not only to act as the engineer to help me build it, but this is a fork of something that Jose built for a much larger enterprise, and I wanted to build something like it for myself. And so what me and Claudio did, uh, who is an Open Claw agent harness, we researched this plan and said, what is the level just below enterprise that we would do so that we're not spending, you know, um, thousands of dollars a month. We would just be spending less. We wouldn't get the same power as the enterprise, but we would be able to do this in more of a, uh, you know, small to medium-sized business. This is something that anybody else, um, like, you know, any kind of roofing company could probably, you know, build themselves. And so I want you to not only help me build it, but I want you to guide me through everything that I have to set up in AWS because I've never done a project like this before. So I need you to have a plan not only for, um, you know, what you're going to build and wire up and connect and the database structure that you're going to do, but also what you need me to enable and turn on inside of AWS so you have access and you are able to help me. Wow, sick. That is a terrible prompt. I'm like sweating, man. Same, same people have been here the whole time, no comments. I'm curious if like I'm missing comments. Is it just not showing up? I'm going to check on Facebook since we're live. I guess it would be on my name. I guess it wouldn't be— it would be on my per professional profile. All right, what do we got? Do you have an AWS account? Yes, my own account. Yes, it's my own account, and it's the same account as Jose's, but this would be kind of like a new set of services, a new environment or instance, if you will. I think that might be the problem. It might not actually be live. Then who are those people? If your fingers ain't broken, throw a comment in the chat so I know where you are. You on YouTube? Are you on LinkedIn? All right, are you on Instagram? Because it says we are live on all— oh, you know what, we're actually not live on Facebook. So you are LinkedIn, X, or YouTube, one of the three. Okay, where should the brain database live? Biggest cost and architecture driver. I believe we decided on Supabase. I'm gonna go back to Claudio. Uh, we're going to move this guy over here so we can just quickly swap. When we originally discussed this, we had talked about staying on Supabase. I believe we weren't gonna use RDS 'cause that was gonna be more of an enterprise-level system. This is a question I'm being asked. Just tell me if I should pick option 2 or if I should type something. I guess I should probably make this bigger for you guys. All right, so we PRD'd $35 a month as constraint, Telco Cloud Option 2. Yeah, what does one fault asset— what is keeping Pinecone versus We are going to do webhook queue and enrichment. Yep.

Budget.

What about those of you who are just hopping in the chat? So there's some new people here. What is it that you guys have— what are you guys working on? What do you got? What questions do you have? Or what comments do you have if you've already built something like this before? And it doesn't currently have AWS. AWS. Okay. And I have to remember this. I have to probably stop sharing this screen for a second. So I'm just going to go get logged in on my actual Chrome browser. Okay, so just copying this stuff over, but this is something I'm not going to share on the screen, not on a live, that is for sure. And for those of you who are just— oh, what are we talking about? It's the wrong password. There we go. All right, bring the screen share back up. All right, for those of you just jumping in, we are building something a little bit different today, but I am doing something— set up a new I've set up a new instance for this, so I will say here You guys are quiet. You guys are super quiet for a sick guy. I don't know if it's a good thing or not. I'm gonna try and listen to some music to, uh, try to ease the fact I can't stand the sound of my own voice because I'm so sick. Comments, definitely hop in because I'm here to, uh, I'm here to kind of bring you guys along the journey. I don't like doing live streams where you just, uh where people, I just say it, um, people just like don't do anything of value. They just kind of say, oh, I'm going to read an AWS model card or Claude code model release or something like that. I actually want to build stuff. I want to build stuff that, um, that, you know, I have not done before. And so then you guys can kind of come along the journey, and I tend to go a little slower that way because it's not something we're doing every single day, right? So, and then for those of you who are just hopping on as well, you're wondering what I do. I build a lot of custom agent stuff for roofers, roofers only. I don't do windows, I don't do concrete, I don't do decks and fences and garage doors, but I do build, uh, Mac Minis and MacBook Airs with OpenClaw and Hermes agents on it. And then build that safe, secure, and ship it to your door as a roofer. And I think I gotta switch to fast mode just to burn some, burn some tokens tonight. What's this? It's a new feature. Show banner, use banner to hide. That's new. Man, this feels slow tonight. Like, Codex is so quick. If I'm over here, I get a lot more done out of this. Maybe I'll work on another project while I'm doing this. Uh, what was that debug mode? Is that turned on? This URL— can you set the URL parameter to get us the actual lines for the camera travel path and the scene box as well, lines showing up again. So this is kind of a cool project to work on. It's a website, and it's almost like a little bit of a video game. It's kind of like making a camera fly through space. And so we had to kind of design a 3D box. If I kind of show you— show you kind of the theory behind it. And again, would love to see you guys' comments in the chat to see like who's here and what you guys are— what you guys are wanting to see. Um, I'm not signed in. Why is my So the idea of this page is essentially this is like a map of how the globe is going to go. And so we had to map like this 3D box, right? And then map the actual JSON. Like, essentially think of this like XYZ. So it's like 75 by 75 by whatever, 700 tall or something. And then each of these camera— this like path of it spiraling through space, each of these sets XYZ coordinates as the camera comes down. And then you have XYZ coordinates for the relative location of the objects that all kind of morph into each other through a part of like a particle shader. And then the— there's the positioning of the relative objects, like the brain rotates 90 degrees, and then rotates, right? And then there is the XYZ kind of yaw, pitch angle of the camera as it moves through space. And so with this, you can see here, now that we turn this debug mode back on, I come in here, you'll see like, okay, so it's rotating around the box. You see that? It's like it's rotating through space, but the brain doesn't move the 90 degrees we wanted it to. So that's like one thing that we need to fix. You see the brain goes up and there's some weird light refraction happening here. But then we go down and this camera's moving and you can see it moving through here. And here we are. And then it's gonna go, oh, now we're in the next camera location. And then bang, there's our light bulb. We're gonna follow our line down some more. And there's our globe. It's not, I have to like adjust it 'cause the, angle is in the middle of the ocean right now, and this is Antarctica, so it's tilted a little bit away from the camera. So we gotta adjust that. And then as that morphs over to here, I don't want all these objects to be bigger. And then we morph over here into the robot and then morph down here into the quadruped. And so yeah, it's kind of cool. So I could probably kick this off on a project. Did this kick off and check on a— get a checkpoint? Yes, it did. I would like to isolate the brain object for a second. And what I mean by isolate is just isolate our work to the brain, because what I find is that as we scroll, the relative motion of moving through space makes it appear as if the brain isn't actually rotating, because I believe we are rotating approximately 90 degrees while it is also rotating, so we don't get the actual rotating effect of the brain. So if I was to have modifiers, I would modify the brain size and make it about 40% bigger, and then I want to make the rotation about 135 degrees to see it actually be able to rotate as we're moving down. I would also like the starting position of the brain to be about -30 degrees. All right, there we go. Do that so you guys got more cool stuff to watch. All right, come over here. So what are we doing? So how should we set up, you guys, poses and price accuracy. So you say, but the difference is, so Um, is there a way to create another instance in the same Amazon? So this is completely new ground for me. It's like working in, uh, working inside of AWS. So we're just figuring out something new today by ourselves. Yeah, it's like we are literally rotating. As you can see, like that there, that's essentially the stopping point. And then you can— because you see a little bit better here, right? You can see that we're going to eventually end up right here. That's another camera stop, right? So it kind of comes here, and then we are in that box that's showing the rotation. But you can kind of see that that's the line that lines up with right here, right? Because then we come down here to the light bulb. So you come down here, and then we come down, bang, there's a light bulb, right? And we're kind of in this camera— cam spot right there. Postgres 7 column staging. Yeah. There we go. The brain's doing a little bit more rotation now.

We would probably start—

so the brain is bigger. It does.

It's just a bit of a—

I have to think, what do I do when the— I don't love the rotation there. Feels like the brain went backwards. I think the brain needs bigger particles closer to the screen. It's tough to say what— are we looking at the back or look at the front? I feel like we're looking at the front and then it goes to the side. But does our perspective change that much? Like we're on top of it I don't think it goes quite— I guess you do feel underneath it. I think it needs more of a pitch. Is it possible to rotate it again another negative 20 degrees so that we see more of the side of the brain? The idea is that it rotates from the front quarter. Left of the brain and then rotates to the front quarter right of the brain as we're going down. So I believe that we go another negative 20 degrees and that should get the brain to be at the right starting point. So you can see here what we're building, right? So we're saying, okay, Roofing Business Partner operates on Avoma for call recordings, JustCall for phone calls, like that's a phone system. HubSpot is our CRM. Gmail, right, is how I opt, right? And then we have ClickUp, right? And so it changed it to say, okay, these things become these, these have these Lambdas, these workers, right? And then we're using Gemini 2.5, uh, Flash Bash. I actually think we're going to plan to switch to 3. 5, that, um, because Gemini 3.5 is really fast and it's pretty smart and it's still pretty cheap. And then we're using Supabase with Postgres and using this pgvector to create the embeddings to essentially create this like vector types, like a just essentially a knowledge base, right? And now So then this becomes the brain. So then every time— so the whole idea is that this stuff is going to— every time ClickUp task is opened or closed, every time a, a meeting is done, the Avoma meeting transcript is going to get ingested automatically and it becomes this like giant brain. And then instead of it living on a machine, right, it's this giant brain that that like any other agent can go access, but it's connected to 16 people's Just Calls, you know, 16 people's HubSpots, 10 people's ClickUps, you know, everybody's Afomas. And that way everything is happening, it's like a sensor. So if you think about like what happened with, uh, how do we get autonomous driving cars, right? There were sensors in cars and people drove them, and even when they weren't they didn't have Tesla Full Self-Driving turned on, the, um, the Tesla was collecting all the driving data and sensors. And now if you look at like other companies that are coming out, like you look at Bedrock Robotics, right, what are they doing first? They're putting sensors on big yellow iron, right? And so all these sensors are saying, well, like what happens when a driver is looking, like when he sees this and he applies X amount of pressure, like how does that actually dig and how much dirt actually comes out? This all comes from sensors first, right? And then so if you're doing this in your business, you're essentially saying, okay, well, we need, we need to first say like, well, what happens when it hits a rock, right? Like when it's digging sand, that's one thing. But what happens when there's a rock there and they sort of like, the bucket's going to try and get underneath the rock and like, you know, there's going to be dirt that it actually pulled out, but then there's going to be, as you can see here, there's dirt that goes back in the hole, right? And so the problem of solving for these kind of big yellow iron diggers to figure out for themselves, they have to get all this training data, right? So first starts with sensors, the sensors collect data and then that data can learn from it. And so the whole idea is they get smarter over time from all the— all how we operate it, right? And so this is just making it so that eventually this understands, like, okay, well, when I run into— when I ran into this concrete, like, how do I break that concrete, right? Like, how much pressure does it take to actually cause that concrete to break? So when I see this in my camera, right? I eventually, as a robot, like, I learned this, right? And so it always starts with sensors, and that— I mean, that sensor needs to be across your entire company. And so that's essentially what we're doing here is saying, look, we need everything that we do at RFP to go into a sensor. And so what this thing is doing right now is speccing all this out because we came to it with a product requirements document, a PRD, and the PRD dictates like what we want to be what they're going to look at. I might be able to actually get that hosted. Are you able to package up the HTML files that you had on your other machine and then send them to me in a— even in a— like, I just need to be able to download them and open them up on my machine. I think he actually already did this. I guess actually working on this right now locally, um, but if I go files, these are all the decisions. Actually, maybe didn't. I actually show it to you now. There we go. All right, so I can kind of show you that. Is this okay? So now you guys can actually see the plan. Okay, so we have the engine, the RBP US Corporation. We have RoofCloud products, like, so we have the product we sell. We have this foundation, and we have foundation platform, which is like our agent clear. We have an agent roster, right? Claudio, Marcus, Vito, Mayor Klosky. So this is OpenClaw, Hermes, OpenClaw, OpenClaw, right? And then saying, okay, so we have all of these different places that actions exist, right? And then we say, okay, I can scroll past much of this. Yeah, so see, first we have the sensor. Then we have policy, right? So this creates like role permissions, approval gates, who can see what. Then we have agents that can take action based on the fact the sensors and policies created this data layer, right? And then once we have actions, then we have a quality gate that says, okay, like before anything, just like the refund question from earlier, like what, who is that LLM judge that says, okay, we got 3 requests for a refund, right? And for step 3, for it to take action, we have to have policy. And this thing basically says, okay, well, I'm going to have an LLM as judge that says, what is the current state? Has a refund been processed? So the first time it would say no, approve the, the refund. But the second and third time it would say no, we can't actually process the refund for that request because you've already done it. And then there's a learning, a learning process where it goes back in. And so this is right here where we are. We are basically on this first step of like defining the schema, the structure of the database. So if I go into Supabase, actually I shouldn't probably do that, 'cause there's gonna be something there. But if I go, if I was to go to Supabase, you could see the database structure already exists, but now we're just saying, okay, well we don't wanna have to manually ingest with cron and stuff 'cause it was unreliable. We're going to actually build an engine that, that runs all that in. And that is what we are building today with this. There we go. Y'all are very quiet today, and I'm super sick. So we're at the hour mark, I haven't had a single comment, so I might just chill for the night and let this run, but I might hop off the live. So I appreciate y'all being here. If you guys enjoyed it, leave something below. And, uh, I will keep you up to date as this project keeps going.