My AI Dev stack as a 2-person, $3.6M ARR company by Dillion Verma (Co-founder at Halo AI)

Sukhpal Saini· 20 min· 3,637 words· 17 min read· English ·Watch on YouTube

This is the full transcript of My AI Dev stack as a 2-person, $3.6M ARR company by Dillion Verma (Co-founder at Halo AI), published on YouTube by Sukhpal Saini. Every paragraph carries the moment it was spoken, so you can click any line to jump straight to that point in the video, search the whole thing for a word, or copy it out.

0:00Um, welcome. Thank you all. Thank you all for coming. My name is Dillian. Um, I appreciate that you guys all came here. It's a huge crowd. I didn't I didn't expect this. Um, yeah, so I I guess I'll begin my talk. So, my talk today is my engineering stack. And um, this is me. My name is Dillian. Uh, 4 years ago I uh, finished college and I went ahead

0:22and started building startups full-time. Uh, and since then I've been building a bunch of different things. Um, one of the things I built is called Magic UI. It's an open-source UI library for design engineers. Um, and it was really cool. Just uh, grew very quickly on GitHub and open-source. And most recently I launched an iOS application last fall in October. And uh, within 45 days we we grew it to 300k MRR. Uh, we did 600 million views in 60 days on social media. Uh, 1.2 billion

0:54views on social media after 4 months. Um, it's fully bootstrapped. Uh, no venture capital taken. I personally took on a $100,000 loan to grow this. Um, and yeah, it's just going to keep going. I'm going to continue working on this application. Um, yes. We're on the next one. Cool. All right, so the purpose of this talk is to go into my engineering stack and what I do for engineering cuz I'm an engineering I'm I'm an engineer uh, by nature, but also I I do product stuff and and develop products and grow them as well. But I think my my mindset is what's important because the engineering mindset um,

1:31is really applicable and valuable when you apply to different different uh, niches and different roles. So, uh, we'll go back to the previous slide. So, as a high-level summary, uh, what I use right now are just these three tools. I don't really go too far deep into customizing them or like trying out a bunch of skills. Um, I find the base models work well for me. I have I have gotten deep into using Pi and like setting up like these custom like skills and custom workflows, but also sub agents and all these things, but honestly like the from my experience, I've only seen like a marginal marginal gain. Like the core

2:09value is just using the AI to get work done in the real world and not like tweaking like your configuration. So, we'll we'll stay on the same thing. A conductor is a tool which I really like a lot. The reason for this is I've tried a bunch of these um IDEs which allow you to uh write code using these agents, and conductor was the one which had the least friction to getting code live. Um everything I do like it's like shipping very quickly. I I'm always shipping new code. Even this morning I was shipping some code. So, what I can do is I can just make a PR, merge it, and go live,

2:44and then and keep doing that in in repetitions and cycles. Um but other tools like Claude Code Clee or um even some other uh variations or competitors for conductor, like they didn't have the best UX in my opinion to do that. So, conductor for shipping fast, uh Claude Code I use uh for everything basically. For so, for full-stack engineering, for front-end,

3:05for for building iOS apps. Um I like to use plan mode and then going ahead and like building out the plan. Uh I always review the code frequently because like both of these models like regardless of what they do, they always um like make up stuff. So, ideally you ground what they're what they're saying

3:23in truth or docs and work from there. And then for Codex, I just use that for performance optimizations, bug fixes, and some back-end stuff. So, if I have a massive um table in the database and the performance is slow on some queries, I'll just tell Codex like go optimize all the queries, go make some indexes, go do some caching, and it'll just do it perfectly. Um but yeah, that's my overall setup right now for for coding

3:48and tools. Uh Okay, go to next slide. Cool. I want to talk about this. This is what I'm interested in right now and I've been like tinkering with for the past 2 weeks. So, Cloudflare released this thing called MCP portals and this was just released I think about a month ago or like 2 or 3 weeks ago. Essentially, what I can do is I can make one MCP for my company and I call it mcp.broda.studio and within this I have all of my tools and apps which I use on a regular basis. So, Superwall, Stripe, Linear for like project management. This

4:25is App Store, App Store Connect. Um ad platforms, etc. So, this is extremely valuable and I'll explain why. Um right now we have a workflow where someone comes to me on my team and they don't have permissions in one of the softwares to do something. So, they ask me to like, "Hey, can you please update this so I can do uh the rest of my task?" Um an example is we're doing paywall experiments in

4:53Superwall. Uh in Superwall like so in our application we have a paywall and users see the paywall and then they pay a price to use the application. We test that frequently, every single week. Uh my team asked me to test a different price and I have to go into App Store. Um it takes me like 10 minutes to like set up all these configurations in the UI, uh input the price and then I have to like go through a review process and approve it. And then once I do it in App Store, I also have to do it in Stripe because we have a Stripe application of

5:21this as well. And I also have to do this in Play Store, Google Play Store. So, it's just a lot of busy work. It takes like 30-40 minutes to do like one thing, changing the price. But now with this setup, I can just tell AI like, "Hey, I want to run a new experiment this week. I want to test if $7.99 works better than $6.99. Can you please get this live, right?" And it's going to go it's going to go to App Store, make the price, it's going to go to Stripe, make the price. It's going to go to Play

5:49Store, make the price. And then it's going to go import the prices into Superwall. It's going to go make the experiment in Superwall, and what and get it live, right? So, all that 40 minutes of busywork to one experiment could not be done in like 2 minutes or 3 minutes, right? And then if you imagine it's compounding at scale, there's a lot of like uh end-to-end mappings of these like busywork tasks you have to do. You have to like move data from here

6:12to here, review data from here to here. Um there's always this like these this busywork that is always happening. But if you have this set up and you give this to your team, they can do a lot more. Be way more productive than you can be individually and like doing these things one-off. Um I find it so powerful. I've been setting up uh groups. So, in my Cloudflare, I have a group for for example, paid ads. I have a group for um engineering and a group for monetization. Each group has access to different tools, and they can go do whatever they need to be productive and

6:48stay stay productive. Um Yeah, this is like the really cool thing. And one more thing about this, and also what's really cool about Cloudflare, um Cloudflare is very powerful because their entire infrastructure is all available as MCPs and and AI uh APIs as well. So, I actually just gave my Cloud Code um an API token to my entire Cloudflare infra account, and I told it to just go set this up for me. So, I gave it this blog post. I gave it my code, and then

7:23it made me this whole thing. Like I didn't have I didn't have to do anything cuz it just it just deployed it deployed workers um for each and individual API. It it It found the other It found the other MCP servers and put them all in. It synced them, and it just did everything. So, like the productivity gains are

7:39incredible for this. I wanted to demo this as well, but I I don't think the screen sharing is is up. Um but we we can we can go into that another time, I think. Uh And then Do you use a local website? Sorry? Wait, how do you want to demo this? Are you going to send me a link or

7:54something? It would be on my computer. Yeah, that's fine. We can We can We can do this later or something. Yeah. Well Huh? What do you want? I mean, if you can if you can We can try one more time to connect the screen share. Yeah, yeah.

8:11Hopefully it connects. [clears throat] Choose cloud that. Try try putting a app as cloud code. Oh, there's classroom five. Oh, there it is. And then should show up. Choose a code. Choose Oh, wait one sec. So search and then that would be That's 2967 Oh, nice. Okay. 2967 And also wrap up around 10:20.

8:46Yeah, that sounds good. Oh, perfect. Oh, amazing. Okay. So this is really cool. So in this chat over here, I've connected my Claude Claude uh instance to my MCP and I can go ahead and like query it for things. So as an example um I'm going to ask it um "How are my ads campaigns doing today on

9:16Meta and on TikTok?" It's going to go ahead, get the data, it's going to show me the data. And this part's like um I guess like decently valuable, but but the the next step up is when my agent can tweak the settings like when I'm not looking at it or when I'm not monitoring it. Because for example in the ads platforms, you constantly need to tweak some settings. You want to like turn off things, turn on things, increase budget, whatever. And every day you got to like review it. But if I can just give that busy work to to to an agent, then also that that removes me from the system as

9:52well. Um I'll let this do its thing. As you can see it's like searching all the tools from the MCP, finding the tool that works, and then getting the results from them, and showing them. Now see if I can show example of all the tools as well. Uh I'll let this do its thing first

10:14actually. Yeah, I'll I'll I'll let this finish first. Um but yeah, actually while this is loading, uh any any questions? Because this is this is the main thing I want to show you guys and go go over. Yes? Yeah, so it's like one MCP and it's pulling different tools? Yeah, yeah. It's it's more it's more so like an MCP, but it gives you access to other MCPs as well. Not not not tools. Tools is is a third layer or second layer. So it's MCP, which gives access to other MCPs, and then to all the tools that they have. Yeah. And you can like segment or or group which

10:49tools or which access you want to give to people. Okay. Yeah. So you you you're you have one MCP that Yes, yeah, yeah. I only have the one. The one which is like mcp.productstudio productstudio. There you go. Um I can pull the results and show that. So now I can ask it something like uh well actually I don't know if I have the edit access set up on this yet, but uh can you turn off the worst ad group in TikTok

11:19uh ads. And it can go ahead and do that. Yes, any other questions? Yes? Uh I'm wondering like if something's not acting the way that you want it to, how do you go about like troubleshooting that tool or that part in the CPU? Mhm. Uh What example of that? Like do you have an example? Yeah, like like let's say you're trying to run this discovery for I group after you first make your MCP and it's not working. How do you uh uh

11:48Oh, yeah, yeah. Yes, yeah, good question. So, this is really cool about what about Cloudflare. Cloudflare um what it did when it set this up, it made an R2 bucket, it stored all the logs for everything. So, then it by itself like checks is it live? Yes. And then it says is it is it working? Yes. But if something is something broken, it's going to go to the logs and check the logs to see what's broken and tell me what's broken. And also it'll go fix it actually, too. So, it's just like fully autonomous. The Cloudflare infra is fully autonomous and AI ready, you can you can

12:18call it. Yeah, yeah. It's really good. I don't know why this is taking so long. But yes, yes, yes. So, I don't really know the details of how MCPs are implemented, but I thought like the loading in MCPs pretty impactful on your context. Yes. So, when you're doing this, are you filling up your context with just the MCP protocols? Like if you run

12:44into issues with the context management? It's a very good question. And Cloudflare thought about this, too, and they and they fixed this as well, actually, with this setup. So, in portal uh I'm going to I'll just open up that to this as well. So, this is what it looks like. So, in portal, which is this thing what they do is they have code mode enabled. So, I think it's over over

13:08here. This thing. So, what happens is that normally when you load MCPs and all their tools, say you have a thousand tools, all of that's going to be in your context window, you're going to pay like tens of dollars, $50, whatever for that. But with code mode, what happens is that you load the tools um into a I guess it's a Cloudflare worker and then it searches for which tool is relevant and which tools to use and then it just does the the calculation on

13:40workers and gives you back the result. So, it goes from like 99% token usage to like like 1%. Like it's like it's significantly less because you don't have to load every single thing into your your contest window. So, code mode is the is the solution. This This was released, I think, a few months ago as well. Um it's a big like breakthrough in in

14:00this contest window problem with MCPs. And Portal does this automatically. So, I can load in like all my MCPs from my entire company and everything and it'll just like figure out which tools to call, get the result, and then give it back to me. Yeah. Yes. Um first, MCPs give you the ability to uh integrate your your platform to everyday applications, you're saying,

14:26right? Yeah. that mean that you can create harnesses to do everyday tasks in the background on a regular basis and I guess really MCP opens up [clears throat] the ability to utilize more harnesses at once? More harnesses at once? I mean, yeah, I guess so because you can also have the harnesses communicate with each other, too. Yeah. Yeah. There's There's all this like like I was saying this earlier, like end-to-end mappings of like this tool has to talk to like five other tools, this tool has to talk to five other tools, and then you're just always busy doing busy work nonstop running in circles. But then if you just give that

15:00to the AI and they can like communicate with each other. Also, another example of something I have set up um Yeah, it's all good. I'm almost basically done. Cool. This is an example. So, I have a a worker which runs every 2 hours. It's going to go ahead and look into my ads account and what it does is it There's a big process, but it goes into my one of my softwares which I have set up which tracks all of my creators that I

15:31work with which help me with marketing. It looks into their accounts. It finds videos which are over 100k views. It imports them into my ads manager and then it and then it puts spend behind the ads. And then it also tracks if an ad got rejected or approved and if it got rejected it submits an appeal as well to

15:51to the platform to get it re-approved. It just does a bunch of these things. So like all this busy work that that would take me time I just don't have to do anything. I just I just look at the report and it just does everything. So this is this is the the daily summary I get. So here's an example of like the denials. So I got a bunch of denials and then the the worker goes ahead and and

16:10submits it again. More updates. More a new creator onboarded, right? So all these different things are happening in the background. And the busy work is is is gone eliminated. Yes, any any other questions about this? Yeah. So I think you mentioned your platform is able to create like a worker on Cloudflare. So I guess you've given it right access

16:35to your Cloudflare account. So I mean that honestly is a scary proposition to me, but just in your own experience what's the value you have derived from that in terms of what is Cloudflare able to do and then also are there any downsides to it? Do you see any of this as a failure that we run or something like that? And do you only give it access to your dev account or do you also give it access to your

16:54prod account? Yeah, I give it full access to super admin access to the entire Cloudflare account that I have. So it has abilities for every single Cloudflare tool whatever there there is. So as an example I have my domain in Cloudflare. I have an R2 bucket in Cloudflare for storing images or assets. I have um

17:15I don't know, so many different things. I have emails set up in Cloudflare, too. So, and then so what like the more I the more I have uh I guess in Cloudflare, the tools can just go ahead and and implement things. The value which I derived is actually just this this example is a really good example, to be honest. The ability to do to do this like literally all I did to set this up was I opened this blog post

17:37that was here. This is the the main post. It was just a guide. I gave the guide to Cloudflare uh sorry, I gave the guide to my agent. I gave it my Cloudflare token. I said, "Just make this for me." And then that that was it. Like it just did everything for me. The entire setup that would have taken me, I don't know, a couple of days regularly, it just got done in basically

17:57an hour or so or even less. So, that that value the the leverage I like to I like to think of it as leverage. Like the more I I let AI do things, I'll make this full screen again. The more I let AI do things, the more leverage there is for me to I guess compound more and more and do things quicker and faster and do more tests, do

18:19more experiments, etc. Yes. I just have a clarification question. So, you basically like went into prompt mode, gave it this link and said, "I want MCPs of one particular versus I'm going to list them all." And then it went in and it just built it for you? Was that kind of what happened? Yeah.

18:33Okay, cool. Yeah, cuz it can go and search the web. It can it found me this one, this one, this one. Some of them I had to make myself. So, like the ads platforms you can see they're on my domain, product.studio. Uh that means like I looked around for them and they they don't have good MCPs or servers or whatever. So, I had to go and one by one make my own individual one for these ones. But the rest of them are

18:54pretty straightforward. Yeah. Uh I'm sorry, there's just a time uh we've got to end the Q&A session. Yes. Uh do you have anything else? Closing words? Okay. Um Closing words? I think One thing I'll say is uh uh I guess just go hard on MZPs. [laughter] So So the theme of this talk But um yeah, I mean if if you guys have any questions for me afterwards, I'll be I'll be around as well. Um but yeah, uh yeah, thank you guys for coming. I

19:29really appreciate it.

Where these words come from. This is the caption track YouTube holds for this video, written automatically by YouTube rather than by the creator. We read it, tidied the line breaks and laid it out so it can be read. The plain text version is at https://viewrankai.com/tools/youtube-transcript/FrojsW4vurs.txt.

All rights in this video belong to Sukhpal Saini. Watch it on YouTube. If this is your video and you would rather this page did not exist, tell us and we will remove it.