Kevin Williams: Hello out there in AI Land. this week Matt is gallivanting around Europe in yet another one of his hacker houses. So I brought a friend of mine, Sabir Summerkant, is the founder and CEO of Growth by Sabir. And like Matt and I, he is deep, deep, deep in the weeds. but Samir's, I would say, claim to fame is that he was in e-commerce like director. consumer online merchandising and sales development. even way before I was. And I like to joke that I was like one of the early people in that field. But Sabir was one of the guys who actually created that field. So there are a few experts who know more about the ins and outs of direct consumer commerce than Sabir. so I don't want you to tune out because D to C is super, super interesting as far as a leading indicator of where technology goes. behind the websites that you see are often incredibly lean teams that are incredibly tech forward because they have to do a lot more with a lot less, given just how challenging it is. Also, they have to deal with huge data pools as far as their advertising and their product sets and whatever. And there's a lot that we can learn about these businesses on the bleeding edge of AI and marketing. So I'm delighted to have Sabir join us today. Sabir: Kevin, thank you for having me. Really appreciate you. Kevin Williams: So I guess let's start out with a little bit of an orientation of what you're seeing. you know, the audience hears a lot from Matt and I about what we're seeing in our sort of B2B trenches. I'm very curious from a generalized business perspective of what you're seeing in in sort of your core market. So Sabir runs Growth by Sabir at the end of the day is individualized or group-based coaching so that he can help people level up their skills around AI with with a general focus around direct consumer. But so can you help us sort of frame who these clients are and sort of let's orient around what you're seeing out there in the wild. Sabir: Sure. Well, first of all, you know, thank you to your audience for listening. so let's go through it. It's not gonna be the top the our conversation should not be just limited to e commerce. because I do help across the board AI automation, right? what I'm seeing is that, you know, one, we have to kill the hype, right? Kevin Williams: Yeah. Sabir: That's that's the number one advice, you know. every founder out there, whether they're B2B or B2C. do fall into following a lot of hype, And even when you when you see YouTube videos from pretty pretty popular influencers, AI influencers, right? They show you one shot, this model is insane. You know, it's in the title, in the YouTube title, right? Or podcast or whatever. you need to get away from that, right? We have business to run, right? Whether it's B2B or B2C, we have to focus on What is the right mix of AI? And it's not a single shot, right? Single shot doesn't exist, right? Maybe I I would believe because I've been in this field for quite a long time, since I was six years old, actually, right? I've been in it. I'm 53 right now, you know. So I've been in it for quite a long time. I've seen iterations of it. There's a hype cycle. You have to get over that hype cycle and focus on what's real. So today I I will be sharing with your audience. What the right mix for you to actually achieve the best results from the AI powerhouses that we have of models that are out there, right? That that's one of the biggest ones. Like that's a foundationally like step zero, right? You have to get rid of the hype and focus on what need what works and and working with folks like you and me, right? Which, you know, we we actually guide our clients towards not wasting time doing. also and taking all the wrong steps and focus on doing the right things and and achieving results. Like even if you're automating 10 different things in your organization, that should be that should be the goal, right? Those 10 Kevin Williams: Mm-hmm. Sabir: things need to work. Those work AI workflows, the quality of those workflows, how the results that you're getting on the other side, how you're compressing months and years of work into weeks and hours of work, right? That's what you need to focus on. Kevin Williams: So help me yeah, let's calibrate a little bit though. So again, you work with companies, you work with much larger companies, of course, but your sweet spot is sort of one to ten million in revenue and getting people going. Like if if if you were to look at at you know, air quotes typical one to ten million dollar company, where should they be at this point in time? Like, okay, there's a lot of hype. There, there are people out there who really are doing like absolutely amazing ninja AI stuff. A lot of people who who aren't and they're they're they're deeply feeling the FOMO. So like help us calibrate on the reality of where like Brandis Generalis should be here at the end of two thousand twenty six. Sabir: One, you have to be in it, right? I actually exited from my Rapid Two X program, which we will talk about you know, later on. I exited about three brands that did not like that w the program was based on AI. You know. Kevin Williams: Yeah. Sabir: You know, if I can only teach you so much, right? you Kevin Williams: Okay. Sabir: know, af you know, after you know, if but you if you have decided that AI is not the way. you know, if you were a retailer when e-com was coming out and you said that, e-com is just a trend, right? In fact, in my career, I was told that internet was a trend, and I quit that job basically, and it was with a big six accounting firm. You know, that was not that's not what we what we want to do. So what I see is that because there is so much garbage out there, unfortunately there is, right? people, you know, founders try to follow that garbage and then they don't see results, right? And then they get turned off. So even for those three brands that had a that that complaint, right? because they were getting a lot of hype information that was coming to them that it should be going faster or whatever. No, Kevin Williams: Mm-hmm. Sabir: but the thing is your business, you first you need to fix the business side of things before you automate. Otherwise, you're automating garbage, right? And Kevin Williams: Yeah. Sabir: and And if you automate garbage, garbage in, garbage out. You know, it's a one of Kevin Williams: Mm-hmm. Sabir: the computer science sayings, right? Garbage in, garbage out. If you automate garbage, you it AI is gonna scale that garbage. And that includes data, which is very dangerous for B2B and B2C businesses, right? Because your life depends on your data. But if your attribution data is wrong, if you're you know if you're not giving the full context to AI, it doesn't know any better. So it guesses based on what it knows with the context you have given it. So first Kevin Williams: Mm. Sabir: things first, you have to give context. It's like meeting a person at a party, right? And and that person doesn't know you. So you say that, I'm in I'm in the fashion business. There are certain assumptions they'll make about the conversation to have a conversation with you. That's what AI does, right? But if Kevin Williams: Mm-hmm. Sabir: if that person, imagine if they were working with you for let's say two or three weeks and they're Taking a pretty deep dive into your business, they have a better perspective in giving answers properly. Right. That that's a like a real life analogy I would I would like to use, right? So it's not just adoption of AI and AI workflows and what's what looks sexy and shiny right now. It's also your context about your brand. What you're putting into the AI is what you're gonna get out. And the more you do, like for our members, I'll give you an example. We have put in so much context into these businesses, into in in AI, right? That it actually says the human required tasks that need to be done because the API doesn't support it or MCP or CLI doesn't support that. It says, do you want me to draft an email or create an Asana card, to-do card for Susan on your team to handle this for you, right? Kevin Williams: Sure. Sabir: That's that's the level of integration you want on on AI side of things. Kevin Williams: But let me let me push because that's not quite tactical enough. Like we get it. Like context is important, data is important, data organization is important. But like what does it look like? you know, one of the pieces that that that that I highly encourage at this point is believe it or not, a lot of organizations don't even have a daily driver. They don't have like an agentic harness in place, they don't have the basic data connectors, people on their team don't know how to use it. That's sort of what I'm talking about as far as as far as calibrating to what does your toolbox look like in that one to ten to know that you're sort of on track and whether or not your FOMO is justified. Do you walk away from the from listening to this feeling like, okay, all right, we're we're sort of heading in the right direction, or my gosh, what do we have to catch up? So like calibrate on that. Like what what does that look like to you? Sabir: I actually have a chart I could show and for your podcast listeners, I'm gonna talk to them through it, but I would encourage them. Kevin, you said that you have a YouTube channel. I would encourage them to actually look at the screen share so that they can see what I'm presenting. so Kevin Williams: Okay. Sabir: if you allow me, I'll I'll share the the the the screen with you. Okay. Kevin Williams: Sounds great. So again, so AI at work, I think it's well, I think it's the real AI at work at at on YouTube, but we'll put it in the show notes if you wanna check this out specifically. But Sabir will be very descriptive about what he's he is describing as well. Sabir: Yeah, so this is on my X account. you can follow me on Sabirs S, you know, so that's my account. and I've pinned this at the top for for your listeners so that they can see it. It's called the my latest AI audit workflow, right? And I'll I'll walk you through what this chart means right on on the screen. you have to, there is no such thing as a a magic model, right? So there is no such thing, right? You have Opus 5.5, which is highly capable from Anthropic. And then you have GPT Soul 6.1, which is incredible at what it does. And then even the other 5.5 models that came out of Anthropic from for Sonnet 5.5 and Haiku just came out about about a week ago, right? How do you you know you know, most people what they do is they just set it to one model and they just say, okay, you know what, I'm gonna just use that model. But the but the problem is from my experience, if you don't use a mix of models, what you do is this is where you gonna run into a lot of issues, right? Edge cases, errors made by and assumptions made by one model because it just missed it, right? Those kinds of things happen. So that's one aspect of it. The other aspect is you know, most people are on if they're running a business, most people are have a hundred dollar plan. you know, per month that they're paying into Anthropic or and and Kodaks, right? Kevin Williams: Let me let me inject right there. So so a lot of as far as those table stakes type things, guys. Most of you need your you and your people to be at least on the premium or pro plans in either Cloud or GPT to getting the most out of the system. if you're still eking by on like the $25 plan. Sabir: No. Kevin Williams: Frankly, you're not using it right. You're not doing it right. So that to me would be part of the calibration. And yeah, it feels a little painful when you click on that GPT and upgrade to premium, but all of a sudden, premium is going to allow a whole bunch of new functionality that you can lean into that just simply isn't available at the lower tier. So anyway, so that's a point of calibration. We all kind of need to expect that we're at that hundred dollars. Maybe it's gonna go up. Don't tell Sam Altman this, but I would. probably pay as much as a thousand dollars a month at this point just for like my out of the box like daily driver in Claude and or or in GPT and agents and things like that. Anyway, continue. Sabir: So I I will I will I'll add on to that a little bit more, right? If you think about because the thing is, I've lived in this field for a long time, right? $100 for an anthropic plan and $100, both of them, and also $100 for the Kodaks plan from OpenAI, right? Is table stakes. If you're running a business, that's table stakes you need to. If you're if you're going with $25, $30, it doesn't make sense because You know, you're gonna run out of you know, tokens or whatever because you're not even optimized, you know, and when you want to do serious work, it's gonna become an issue. So table stakes, because if you hire even a junior developer on your team to do a fraction of the work that model these models do, right, at a minimum, you're paying them 60 to 80 K a year, minimum with somebody who came out of computer science field with like no experience, basically. You're paying them almost that much, right? So compared to that, it's it's $200 per month, $100 each, and it comes out to $2,400. And I want to give you t this tidbit also. I never automatically take the annual discount. It doesn't make sense. Because if if if one of the labs falls behind and I've now prepaid for the whole year, it doesn't make sense. I would rather take that $100 away from that from that lab. And give it to somebody else who's who's doing a better job, right? So stick with the monthly plan, even though net net you're paying a little bit more. it gives you that freedom of switching if if you need be. So let's go through this, right? So when when I have any kind of a task, right? Any task. This could be a micro task, it could be a major task, right? my orchestrator, my main model that I have set my clot code to is Sonnet 5.5 on high. Kevin Williams: So so add flavor. This audience is hands-on keyboards in a lot of ways. So when you say a task is surfacing, where is that task surfacing? And in the case of Claude Code, where is that interface happening? Sabir: Yeah. So the task could be anything. Like you you could have done some sort of gap analysis. Like if you're a B2B business, most likely you're on on a platform like Salesforce or you're on you're on BitCommerce, for example, right? maybe that task is you are we we will go through one example. This is relevant to regardless of whether you're B2B or B2C, right? has to do with site performance, for example. The task might be you're auditing. And I'll show you how to do that as an example today, right? You're auditing and you're seeing you know, where the gaps are, what's causing your site performance issues and why site performance issues are gonna be important to handle as step zero, literally in any business out there that has a web presence, right? So that's where the task is coming from. Now, depending on your skill set. You could be doing that entire audit, including this entire audit that I'm showing on the screen here. You might be running through this, even to do the planning phase of it to identify what those gaps are, what's causing your slowness on your site, as an example, right? You're doing that work, you're running it through this audit. It's finding even all the edge cases and everything. And then it's you know, going through and you're fixing it and so on and so forth. So even What's funny about this workflow is I've actually run the workflow on itself to Kevin Williams: Mm-hmm. Sabir: make sure that there are no edge cases in the workflow that I may have missed. Kevin Williams: Okay. But again, is this an app that you're you're you're hosting the app or is you running this as an edge function, a cloud worker? Like w where is this sitting? Sabir: I'm in cloud code. I'm I'm in cloud code on on Kevin Williams: Okay. Sabir: your Mac, on your PC, on Windows or or Linux. I'm running Kevin Williams: Mm-hmm. Sabir: the all of this in cloud code. You could run it on the cloud app also with on the on on your desktop. You could run that. I personally prefer the terminal because it's a lot faster for me. And and I've also even experienced that running cloud code on a Linux machine. is eight to tiny ten times faster than running it, eight to ten times faster than running it on a Mac OS or Windows, you know, Kevin Williams: Hm. Sabir: because because of the operating systems itself, you know, cause and and Linux does not have that barrier. I've actually test even tested that. So more more serious workflows that I want to run 24-7, I do run them on Linux. Anything that I'm working on during the daily driver, I might be working on it on my Mac. But I have the repository on GitHub or or my own repository where I'm checking things in so I can go bit between the different operating systems to do the work. Kevin Williams: All right, proceed. Keep keep on trucking through the flow. Sabir: Yeah. Yep. So Sonnet five five on high effort. By the way, when I'm even describ telling you to put it on high or max or whatever, there is a reason behind it because this has been tested extensively, right? So sonnet five five on high does all the planning, all the dispatching, all the scoring, and all the fixes. So this is like your your universal model. This is the one that does all the work, right? haiku five five that came out. needs to be on the max. And haiku is very good at it's very cheap for first of all, right? And it does basic operations really well. Finding a piece of text inside of a file by using a tool in your operating system, right? To go to graph it or something like that. It does a great job. For it to do complex things, it doesn't do a good a good job at it. But for basic stuff, I want it to you know sonnet to actually spawn eight sub agents With Haiku Worker, right? And Haiku five five max to take care of the basic stuff so that Sonnet is not burning up my plan, basically burning my tokens and my plan. Right. So this is very effective at token optimization also, so that your hundred dollar plan can stretch for a long while. I I barely run out of five hour usage or weekly usage. Kevin Williams: So just a reminder for the audience that Anthropic runs on literary references. So their lower end model is haiku, then Sonnet is sort of the workhorse, Opus is is is the the higher end inference model, and then Fable, which is getting a lot of press in a lot of ways. I'll say for my own workflows, I find Fable or Opus five five to be remarkably efficient in a lot of ways. Slicing and dicing this way. it really, really makes a ton of sense at any sort of scale because it can get really expensive. But in smaller one off type workflows, you you mentioned settings on high. you can get away with a lot through a high setting on Opus these days and preserve tokens. It's pretty amazing. Sabir: Yep. Yep. I mean, by the way, you you brought up Fable, Fable five point five. Skip it for now, right? let's see what Fable five five comes out. We'll see if it's worth it, because I actually test this constantly, right? Opus five five on high. I like I totally agree with you. More than enough, right? There is nothing you don't need that, right? This is optimizing the entire plan so that you are maximizing your hundred dollar usage per per month that you're getting, right? So when when this runs, it creates the artifact. Now the artifact could be a piece of code, maybe it's updating Shopify code or your or your theme, it could be HTML, it could be a document or analytics report, anything like that. All the connectors you talked about that you have connected to your clot code, maybe it's pulling that data and it's actually doing a first draft of of that thing that you asked it from the task to do, right? So it's creating that after it creates it. So these two guys are working together, they create this. This is created. Now, after it's created, it says that, I'm done. AI is very confident about completeness, right? That that, I completed 100% to your satisfaction. This bunch of garbage, right? So Kevin Williams: Hmm. Sabir: you know, okay, you know, you say you completed it, let's make sure that you completed it correctly. So, what happens is now this workflow. starts up two subagents, a SOL editor. Sol is GPT 61 Sol on High, right? Read only. And then it creates it's the other sub agent, Opus Auditor, which is Opus 5.5 on high. It feeds this artifact that was created from the for the task or a bunch of artifacts, right? It feeds it to both of them with 10 scoring criteria out of 100, right? Based on this is very dynamic, right? So it determines so the orchestrator dep determines what the scoring criteria is, and then it gives it to Soul Editor, auditor, and opus auditor to actually audit the work that it has completed. And it these two auditors' job is read-only. They only find what the problem is and they report back. And why do we do read-only? Because w with these higher-end models. With with GPT Sol and Opus, they are expensive on rights. You know, that's why you have the input cost and output cost of tokens, right? The inputs are super cheap, but the outputs are expensive, right? So these auditors, we don't want them writing anything. We just want them to, we don't want them fixing anything in here. We just want them to find the problems and feed it back to the orchestrator. And then the orchestrator needs to fix it. And this cycle right here runs three checks basically. So three times it runs to make sure that you know, it finds, you know, it find does round one, comes back with a bunch of critical errors, you know, you know, high errors, medium errors, and low errors, right? With with the details and with the success criteria of what how to fix it. This guy, the cheaper model, actually fixes it, right? It fixes it and then round two starts. And it does the same thing and round three goes, right? Every round. So by round three, you might start out, and this this one gives a score of soul gives a score, a soul auditor gives a score of 32 out of 100. Opus says, Kevin Williams: Mm. Sabir: it's 42 out of 100 in round one. But it says this is how you fix I found these errors, and these are the things that need to be fixed. It goes back, round two, it that's becomes the input basically, right? It runs through it, it fixes it, it feeds it back. It runs it to round two now. It reads, Sol says that, now that number jumped from 42 to now 75, right, out of 100, right? But I still found some medium errors and low errors because of the fixes you've you've done. Opus you know, increases the score to 82 out of 100 or something. And typically what I do is the the success criteria I've said it because these two are very aggressive at scoring, right? So eighty-five out of one hundred is considered success, right? And the higher that if you want to do more rounds, you could actually get it to ninety-six out of one hundred, right? But I I typically stick with three and three tends to get it well above eighty five, like ninety-two, ninety three or something like that. And then even after the final round, it tells you I still have these things open. Do you want me to still fix them? Kevin Williams: So as it I I I I I suspect that we might have some audio glaze over going on because there's a lot of detail in there. And I I orienting around a couple of like primitives in terms of architecture, I think could be really important for people to understand. So essentially what you're building in here are a series of loops. And being able to build loops, you can actually do qualitatively. you can do it. This is what I would call a deterministically driven loop as far as you building an edge function that is going to loop back on itself. It's going to call a specific model that's doing a specific task. And when you're trying to create a repeatable action across an organization, there are, you know, 10,000 insurance agents or whatever it is, and it needs to be looped back in exactly the same way every time. This is the eight sort of underlying structure that that that people like. Matt or Sabir and myself would build under the hood as far as to make those loops happen. And however, there are different ways of skinning this cat. the loop function is actually it's a skill. And we've talked about skills on this show before, as far as it being essentially a set prompt. So if you wanted to do this in a less like coding type way. You could actually do loops of skills and the skills which you can build through the skill editor. And I really feel like not enough people are understanding how to deploy skills. So hopefully this is going to resonate. That you could be building a skill in Claude that calls haiku specifically to do a loop and establishes the threshold for passing. So Internally, we have a skill that we call Naysayer in coding. And it creates loops that it has to pass before it can go on to the next stage. And the next stage is a skill called Simplifier. And what Simplifier does is it looks at all the mess that Naysayer created and trying to get it to pass and it cleans up that code and then it runs Naysayer one more time. So it's handing. Sabir: Hm, very similar. Kevin Williams: It's it's it's a similar idea, but it's it's actually driven by a qualitative input. It is for for those who aren't thinking in in pure code, you can actually describe that. You can say, All right, what you want to do is run it through the code. Naysayer's a little tough that way, but simplifier would be a little bit easier. And you're going to run that over this until it can proceed to the next bit. And you know. We talk about like hallucinations and to be honest, in 2026, people get what a hallucination is, but these are tactics to deal with hallucination. They're tactics to deal with drift in your code and your output to make sure that you're not building garbage into the process. One of the first things you said here today was if your process is wrong, you get garbage in, garbage out, right? Sabir: Yeah. Kevin Williams: Well, if you have something that's wrong in the process, you're handing, you're creating garbage somewhere in there, and then you're automating that garbage as it passes down. So being really regimented about the way you think that's you're scaling the Sabir: You you're scaling the garbage. Kevin Williams: garbage, exactly. And so so these loops, which sound really complicated, but I mean you could go back to 1950 and like think about organizational structures and like human-based feedback loops that organizations had to build. In order to like check things. a lot of what we're building here with relatively complicated language is a reflection of how organizations have always worked and being able to have checks and balances and feedback and whatever it is. So don't get dissuaded by this sounding complicated because this is simply actioning process that you're probably already doing. Maybe you're already doing. Sabir: And how to how by the way, how to how to implement this is simple. you go to my X account, it's this is pinned, you cop, you save this image that's featured, right? The the workflow, save it, paste it into Claude, and then the text that's next to my that that's next to my to to the post text, copy literally up to here, right? Kevin Williams: Mm-hmm. Sabir: you just put this in here. with that image and Claude, because it's it's a multimodal, it understands image, it will actually take this and implement it in your workflow. That's it. That's how simple this is. Kevin Williams: You could probably also take this and build a skill with a skill builder. You could build a Sabir: yeah. Yep. Kevin Williams: very complicated skill that is Sabir's Awesomeness skill. And then whether you're using dot and GPT or you're using just your normal vanilla Claude like interface, you could say literally use Sabir's Awesome Flow skill to assess X, Y, and Z. Now this one, this one look it's looking a little bit generic. So you could do that with this. You might then have a sub skill that's like SEO optimization skill that uses Sabir's awesomeness skill to do this. And you can start calling that from a whole bunch of places. What gets really powerful then is once you have, I'm just gonna keep calling it the Sabir Awesome skill, like you could that you could then you Sabir: I like it. I like that title. Kevin Williams: could then plug it into other operations because this is really working. And you could then invoke this skill in other places in the rest of your world. I might be tempted to to pull like the the double check part of this out as like a subskill. And then all of your apps, all of your workflows, all of your processes are approaching it on a consistent basis. And what's super killer is if Sabir version two comes out because now Fable 18 is out and like, you know, you need to change this. Because the severe skill is being invoked in the rest of your system, you can change that skill under the hood to adapt to changing circumstances without rebuilding the rest of the workflow. Sabir: Yeah, I totally agree with you. I mean, this has gone through so many iterations. Like let's say Sonnet ten comes out, right? And haiku Kevin Williams: Yeah. Sabir: is fifteen and Opus is you know, fifteen, and then there is a new model, Bell model, D E L Kevin Williams: Yeah. Sabir: from from GPT that's rumored, right? Comes out. And that that actually proves that it's better. You could literally say that okay, swap out my sonnet five point five with sonnet fifteen, right? And Kevin Williams: Mm-hmm. Sabir: could you could actually just literally replace it and it will it it will replace that so 85 or more the the you know sonnet actually fixes every finding so nobody else touches the fixes except for the orchestrator if it passes it just moves on to the next if it stop it will stop after three and ask Kevin Williams: Mm-hmm. Sabir: you and say hey you asked me to run three times I got it up to this point but I still have these things open and and you do you want me to do something about that? Kevin Williams: So, all right, I'm gonna push one more time on this because there are people listening who are like, okay, this is cool, but like how the heck do I get started? Again, this has to be running on like an edge worker somewhere. So so I guess define an edge worker for us. And what would be your recommendation as far as that like basic the the basic tool set? Is it superbase? Is it Cloudflare? Like, what are you building this on top of? And we know once you've built this cool thing that looks like a a a you know a flow chart, like yeah, okay, you can do what Kevin said and you can build a scale and you can just do that in GPT, but that's not actually the way to scale this. So where does it go? What's hosting it? Like what's the nuts and bolts that are sitting under this? Sabir: So for the for the daily driver, like there are certain things that I do on a daily basis, right? checking my all my inboxes, right, through all my social, Gmail, all of that stuff. So this runs right in front of me on my desktop. so that's that's one type of workflow that's just kind of manual that I run it because I I I don't want AI to just respond to emails on my behalf, right? Because I wanna I wanna be I wanna have some input in it because I don't want it to guess what my response would be, right? So on on that side of it, you're running it, you're running it on under your clot code and and you're keeping it to that, right? Now, when it comes to running it on the edge, you can run it. Cloudflare does a great job with workers, right? You could actually Kevin Williams: Mm-hmm. Sabir: run that in there so that you know, when if you are if this is part of your app, meaning that where you know it it's a SaaS product that you have and you want When the results come back, you want it to be fully audited because this is a reporting dashboard, let's say, and it's a critical dashboard for insights and reports and stuff like that. You don't want it to be just one model single shot that it it gave you the answer and you're putting it with pretty charts. Pretty charts could be garbage charts, you know, right? It doesn't have it doesn't have the right things in it. So Cloudflare Workers actually is is is a great use case of this that you could actually put it in there. And then you could get when you get the results out of it, then then that's the thing that's gonna feed the result, the reports back to the dashboard as an example use case. Kevin Williams: Sure. So CloudFiller workers are also absurdly cheap. it we just for just sort of because of inertia on this show we talk a lot about super base, mainly because that's my tech stack and that's Matt's tech stack. but cloud filler is great and again, really cheap, really scalable, but it does require a a few other approaches. The other piece here that I'd like to pick apart. Is is how you handle this changing over time. So just a second ago, I was talking about you need to adapt to Sonnet 10 or whatever it is, but you have three, four different Sabir: Four. Kevin Williams: models that are baked in here. how do you control that from a central process? And once somebody has done what we just said, as far as taking a screenshot and build this, what is the right approach in your opinion? And I have opinions, so I'll I'll I'll voice this out as far as keeping your models up to date with the latest and greatest and what does that look like when the new model comes out? Because PSA don't just upgrade because models behave in different ways and you can break things. So how do you think about that as models continue eva to evolve every 15 internet seconds? Sabir: So I have this is a borrowed concept from the old days, right? You have a dev environment, you have a staging environment, and you have a production environment. Production environment, you need to fully vet before anything makes it to production. You mentioned Superbase, right? Superbase, if you have it, if you have that something running in production, like a SaaS product, right? You don't want to just update and replace the models because it might be garbage, right? We we've had that sort of experience recently also with with different GPT models, for example, right? And and some people did not upgrade. Or Opus five came out, people didn't upgrade. They waited until Opus five five came out like literally around the corner. Right. And then Opus five five was so much better. Right. So you when with when it comes to production, you don't want to touch that unless you go through this cycle. So you do need to have a dev environment When new things come out, you should be testing it. And by the way, in this post, if you could come down here, I I said currently using Muse Wonder and Muse Wonderful from Meta, right? Kevin Williams: Mm-hmm. Sabir: To see if this orchestrator worker auditor workflow needs to be updated with Muse, right? And Kevin Williams: Mm-hmm. Sabir: I said I tried using Muse Auditor, right? I tried to add that in here, but the problem is that they have some kind of a problem with their plan usage. Kevin Williams: Yeah. Sabir: that it it burned through like that in before even round one was completed. Round Kevin Williams: Mm-hmm. Sabir: one, right? And that was the fifty dollar plant, right? It that shouldn't happen, right? And I tried this many times. And I also and I also I'm sorry, I I don't know why I went there, but I Kevin Williams: Yeah. Sabir: also what I did was I did Argon. I Kevin Williams: Mm-hmm. Sabir: also did Argon. The problem with Argon, Gemini four Argonne, that's when it came out, this was on October sixteenth, right? A couple of weeks ago. So when it came out, I also tried that. So when these kind of things come out, this is what I do. I actually try them to see does this work, right? Does it work in my process? Or it would it replace Sol Auditor, for example? Or do I need to add a a a third one that's going to give me even more better results? Right. So it's a mix of models here, right? So I even did Gemini 4 Argon. Unfortunately, this is my experience based on my workflows and the work I do. The hallucination is a serious problem with Gemini 4 Argon, in my experience, right? So it it actually was hallucinating problems that didn't exist. So when Sonnet got it back, it said that no, I double checked it. And also the audit Sol Auditor and Opus Auditor checked its work and said that no, that was garbage. That that's not a real problem, right? So I actually am testing that, right? Then deep seek Deep Seek, actually, I have another post here. Deep Seek V4.1 Flash 1017 update came out, right? On October 17th. Right. I actually tried that. Now I added that. It did not replace my Sol Auditor and Opus Auditor. I added it as my third auditor for certain workflows because I do have a $10 plan with command code that gives me about a hundred dollar credit with with Deep Seek. So and it's super cheap. as an auditor, it does a great job. It's very fast also compared to a soul, it's very slow, but it's very good at finding edge cases. Opus is medium, I would say, from a speed standpoint. Deep Seek is faster. However, let me put a huge disclaimer on it. the in this workflow that I'm showing you, these are all American models, right? Kevin Williams: Mm-hmm. Sabir: And Deep Seek is a Chinese model, right? Highly capable. But there are issues with privacy, with with there were there have been some issues with whole repositories getting uploaded by GLM, for example. That's another one from Z dotai lab, right, Z Po lab. So there there are things like that. But in my dev environment, when I'm testing these things, you know, it makes sense to test it. Would Would Deep Seek make it to my production server? The answer at this moment is no, right? Unless I totally understand okay and completely understand my privacy options that I have and w and how my data is gonna stay private with me and and everything that I'm doing it stays private with me. Kevin Williams: Okay. all right. So yeah, different model model categories and approaches. you know, we've we talk about this a lot here, that the that the switching cost is relatively low as long as you understand what models fit into similar categories. Like if you're using s you're using anthropic Opus Four five point five, you probably are going to have the best luck in GPT land with soul just to keep your life a little bit easier. now you can play around with different approaches to that to figure out what it is. But one is you do have to check because they have like just because Seoul and Opus are are serving the same class doesn't mean that they're going to have the same outputs. So trust but verify is one. The other in in terms of a build hygiene perspective, I would say is it's it's it's really important not to accidentally build around hard-coded models. It's the the the whatever you're using to build, ranging from you know cursor to cloud code to codecs, whatever, if you don't have specific build instructions that are built into your agents MD or your Cloud MD, it will hard code models. And I stumbled across the across that. I have like 600 references across 50 repos to sonnet 4.6, which is going to go bye-bye. And that's a real problem. So you can it'll Sabir: It will start breaking. Kevin Williams: start breaking. So what you need to do as as you mature in your build, like it's it's cool to do kind of whatever you want to do on your first build, whatever, but You know, I have we 120 different repos in our organization that are running all of these different things, is having a common language in the organization around things like model choice. So for instance, creating categories in the code like that are essentially calling a variable instead of calling the model specifically. So call. Sonnet, like most recent sonnet, and then you can have a library that's built somewhere about what most recent sonnet is. And thus within the app, or you can be much more clever about this from a systems perspective, all you have to do is update the library, and then all of that percolates through the rest of this. So I it you could build that to be automatic, but that makes me super nervous just because I I want to know. Like there were huge differences between, you know, Opus five and opus five five, huge differences in the ways the way Sabir: Yeah. Yeah. Kevin Williams: that they worked. And if you just blindly built something that was use latest Opus, like your stuff started to break. So one, don't let don't avoid the temptation to let the coding agent build hard coded references and think of things in terms of a library, which people with a coding background, they they do this naturally, but sort of Normal people have no idea. Like, cool, I'm gonna use the best model. And then 18 months later, the best model is a super old model, and now you have to update. Sabir: Yeah, I mean i as an example, this entire flow that I'm showing here, right? For example, you could have something called LLM config, right? And Kevin Williams: Mm-hmm. Sabir: LLM config.env, you know, that's a that's a file that says, my sonnet orchestrate my orchest orchestrator is sonnet orchestrator and sonnet orchestrator is sonnet five five on high, right? my haiku worker is haiku five five. Let maybe later on you know GPT six one Luna comes out, right? And then you want to replace that the haiku worker with 6.1 Luna, for example, you could actually change that in that one file. And then all this entire workflow could be pointing to that LLM config dot env that that would reference that. So if if anything changes, you just go to that one file, change it, and all of your workflows are going to change from there. basically. But test, test, test. Really important to test it. Kevin Williams: So somebody comes into your your tender embrace and they're you're doing tutelage through this and you're introducing them, you're trying to match where they are and and like amplify their basic skills, I imagine. and then you start building workflows. like what do they look like after six months? Like what are the types of things that they're they're they're taking on in their business? yeah, I'm just gonna leave that with you. Sabir: So a lot of compression happens. So that's that's the number one, the number one thing, right? Things that were they were carrying a lot a lot of founders carry either they carry it on their Google Calendar or their Trello board or Asana, right? A bunch of stuff that's just sitting there. All of that could be automated or or at least coded through AI so that it's a it's a done deal. Basically you check it off that the to do list is done, right? So there's a lot of compression. That happens from there. In our in our case, specifically with Rapid2X program, which is a built-on AI, you know, from the ground up, typically in the first 12 weeks, these e-com businesses see revenue increase because we want to have a monetary number associated with the work that we do, right? 108% Kevin Williams: Mm-hmm. Sabir: increase. This is average across 200 plus brands that have gone through the program. in the first 12 weeks, they see a hundred and eight percent increase revenue growth. profitably because Kevin Williams: Cool. Sabir: we we also optimized things like meta ads and broad ads and MER and stuff like that. These topics are also relevant for B2B audiences Kevin Williams: Sure. Sabir: too. Yeah. Kevin Williams: That's I I mean it's great. I think the overall guidance for anybody is get started, find workflows that are available. I I this is usually where I make a pitch for process engineering approaches, like don't get scared, but like you have to understand why you're building something and what the components are of it such that you can amplify it. And then you want to have building blocks in the system like we're seeing here that catch errors so they don't compound. And those building blocks can be plugged in. this is sort of where the the the the the reality hits the ground around this concept of compounding advantage in AI adoption. That the reason why brands who dove into this in 2023, 2024 are so much farther ahead is yeah, there have been a lot of scars, expensive mistakes, and whatever. But fundamentally they're building all of these blocks, and those blocks build up to new and better things. So similar to the adages about planting a tree, that you know, the best time to get started was ten was, well, in this case, four years ago. The next best time to get started is now because you've got to build these blocks on top of each other. Sabir, this is awesome. I love to see the this is a I feel like we've violated some of the core tenets of the show. We shouldn't talk about models. So we're not the show that talks about all these models, but the model stuff is important to like remind people of now and again. And I think he did a great job of laying that out. So how do people find you and how do they get in touch with you? Sabir: So you can go to my site, growth by severe dot com, growth B Y S A B I R dot com. And on here, you would if you are running B2C e-commerce, D to C e commerce business, we go we have the Rapid2X protocol is a one year program. We go through nine sprints. Each sprint is about six weeks each, where everything is built built on AI. And just to kind of give you a testament to to the program from a human standpoint. We have had fashion founders who have had zero technical skills be become extremely efficient on clot code. And now they see everything they do translates to code, basically from an AI standpoint. And they're very efficient at it. that's from the Kevin Williams: Yay. Sabir: human side of things. And and as b and like you said, to the time to start is now. the thing is, even if you're at zero and you you have no clue about tech, you know, starting with it, believe me, twelve weeks from now. you will you won't recognize yourself. You won't believe what you're achieving. and and you thought that you needed a staff of IT workers, right? versus what you are able to do. And it's not just that, it's also numerical analysis and stuff like that related to your business, you know, to understand your business metrics, like what is cost cost of goods, for example. Believe it or not, most most founders don't even understand all components of cost of goods sold, right? But with AI, Kevin Williams: Mm-hmm. Sabir: it becomes a lot easier. Kevin Williams: Sure. All right, sir. Thanks for being on and hopefully we'll see you again in the future. Sabir: All right. Thank you for having me, Kevin.