Subscribe
Share
Share
Embed
The development world is cluttered with buzzwords and distractions. Speed, focus, and freedom? Gone.
I’m Nicky Pike. And it’s time for a reset.
[Dev]olution is here to help you get back to what matters: creating, solving, and making an impact. No trend chasing, just asking better questions.
What do devs really want?
How can platform teams drive flow, not friction?
How does AI actually help?
Join me every two weeks for straight talk with the people shaping the future of dev.
This is the [Dev]olution.
Marco Martinez (00:00:00):
I recommend anyone when they're starting off, create something small. Create something that's easy, not low stakes, nothing that's going to be in your workflow today. You want to first take baby steps because it really helps when you want to create the big product.
Nicky Pike (00:00:17):
This is [Dev]olution, bringing development back to speed, back to focus, back to freedom. I'm Nicky Pike. You know what's missing from every tech conversation, actual live demos. Welcome to [Dev]olution two code where the BS stops and the code starts. Here's the deal. My guest has one hour to put their code on the table. No slides, no promises, no. We're building towards that, just fingers on keyboard, code on screen, and real results. The rules are simple. You build something, show us. If it works, prove it, and if it's production ready, then ship it. Otherwise we're going to skip it. And today I'm putting one of our own in the hot seat. Marco Martinez is the co-creator of this show. He helped build [Dev]olution from the ground up, but today he's not behind the scenes. He's on camera and he has to play by the same roles that he helped create.
(00:01:06):
Marco is the community marketing manager at Coder and he decided to solve his own business problem instead of waiting for engineering. Here's the thing, Marco is not a developer. Six months ago, the only Python that Marco knew and I quote was a really big snake. Now he's built a multi-agent system that monitors Discord, run messages through a Lama, AI routes them to Slack for approval, and then posts back to Discord and it's heading to production. This is what Vibe coding actually looks like in an enterprise context. It's not a toy app. It's a real business problem that needed solving while engineers focused on core products and a non-technical person who said, hold my beer, I'll figure it out. Before Marco shows us what he's built, here's the challenge on the table. The market is screaming that value coating is real Cursor, just hit a $29 billion valuation, nearly tripling in six months. Lovable went from zero to 200 billion in a RR in under a year and is now valued at 6.6 billion. We have tools like Blink that are letting people build agents that build other agents, but what actually happens when someone with zero engineering background tries to build production software? It's time to find out. Marco, welcome to the [Dev]olution.
Marco Martinez (00:02:15):
Thanks for having me, Nicky. I'm glad that I finally get to be a guest.
Nicky Pike (00:02:19):
Yeah, get to be a guest on a show that you helped create, man. Sure. Anything you want to state before we get started?
Marco Martinez (00:02:25):
It's really great to be here and I'm really excited to share what I created. I think it's going to inspire other people to create it. I mean, that's kind of tooting my own horn on something that I created. It's not like the second coming of Software Revolution or anything like that, but it is something that when I've talked to people about, they're like, whoa. It gets them starting to think about, Hey, I can actually do this. So I'm really excited.
Nicky Pike (00:02:48):
Well, that's awesome, man. Well, let's get started. So first of all, Marco, you're a marketing guy. You have access to engineering teams. Why the hell did you decide to build this yourself instead of just filing a ticket and waiting for the experts to get to you?
Marco Martinez (00:03:01):
Yeah, that's normally how the process would go, right? The other companies, you would've to file a ticket even before filing the ticket, talking with engineering to see if they're even going to attempt to do this right?
(00:03:13):
And then you file a ticket, you hopefully it's going to get done in the next six months or year. But that's time that I couldn't really waste and not saying that the process is a waste, it's we're moving so fast these days that we have to be on top of our jobs, our projects faster than ever. So luckily at Coder we have a lot of resources, we have a lot of tools, and also from the ground up, our leadership has said, Hey, go tinker. Go create, go do something that will help your job. We'll make it more efficient and you have the runway to go do it. Go learn, go mess up. You're not going to really break anything, so just go try. And so that inspired me to go like, alright, I have this problem and I now have the tools to actually fix it. So that's kind of the way I approached it in this sense compared to previous companies that I used to work for where there was a very rigid process because the tools weren't there three years ago. Right.
Nicky Pike (00:04:23):
Well, and so you're the community manager. One of your roles and responsibilities is you help manage the Discord channel that goes into coder. So walk me through your actual problem, Discords, intro channel. You're constantly traveling, you got notifications everywhere. What was it that was breaking that made you think I need to write a bot here?
Marco Martinez (00:04:41):
Yeah, I wouldn't necessarily say break, but what was happening was my process was starting to slow down. We started off the whole community process in January of last year. We had had some efforts at community beforehand, but in a startup everyone's wearing multiple hats. And so community was being touched upon throughout the years, even before I was at Coder. But really last year we really wanted to make it a key part of our organization to really take community seriously and make sure that we had somebody dedicated to run community with that. Now that I'm not wearing multiple hats, I have to now focus on community, but just all the aspects, it's not just managing Discord, it's attending meetups, creating strategy, going to events, doing other marketing responsibilities. Because even when you reduce the number of hats that you're wearing, you're still wearing a couple other hats still whenever you're with a startup. At the beginning when I took over, I went into the Discord and wanted to make sure that everybody got a personal message, a message that welcomed them, made sure that they knew that there wasn't a bot behind a response of them saying hi in our intros channel because our intros channel on Discord, the place where people would introduce themselves when they would discover our community.
(00:06:02):
When I go into communities and someone responds to me saying, Hey, welcome, this is where you go, just guide you in the direction of where you need to go, instead of just letting them say hi and not responding to them, I feel that that interaction, it has to be personable because if it's not, I feel like there's a disconnect there and you're going to just maybe sometimes lose community members who were waiting for some type of welcome when they come say hi. So I decided to personalize every message that came through, and at the beginning when we were kicking this off, it was slow enough that I could respond to everybody in a timely manner and a timely manner for me was within the hour. I would respond a lot of the times even faster than that just because if somebody messages and I see it, I like to respond really fast just to let them know, Hey, thanks for coming.
(00:06:55):
What are you working on? It's not super detailed, but it's enough to get them engaged and we have a back and forth after that. So as the year progressed, I started to travel more. I started to change time zones more, losing two hours, gaining two hours, losing 10 hours, going to Europe and events out there where there's gaps in internet coverage. And so there's really larger gaps than what I had established before at the beginning of the year, which was like an hour, and they were starting to get four hours and half a day to 10 hours. And that's when I was like, alright, this is not spinning out of control, but it is slowing down my response time. And that's what I didn't like for my own personal reasons and also what I was trying to establish in the community. So I was like, alright, I got to create something that handles at least the foundational part of this, but doesn't lose out on the human aspect of the interaction. I didn't want to create something that was just ai.
Nicky Pike (00:07:55):
Well, and it's Discord. And as the community grew and as it started to explode, I imagine you were starting to miss messages. They were getting lost in the fold. So I'm guessing that you needed something just to help remind you, Hey, this person came on while I was doing something else.
Marco Martinez (00:08:08):
Yeah, exactly. Deliver those messages, those notifications in a place that I was always in because I am in Discord all the time, but the notifications that are priority on my phone and computer are in Slack. I live in Slack all day. And so I was like, all right, I need to see these messages faster than I'm getting them just through Discord. So I need to get them in a place where I'm always going to see them. And Slack was the place that seemed like the best place to have this whole thing kind of start from.
Nicky Pike (00:08:39):
Yep. And we'll get into all the technical details when we get to the two code portion of this,
(00:08:44):
But one of the things, I mean, you're not a technical person. You're not a developer. And when we were talking about this, you went to film, you studied film in school and you kind of drew this parallel between mobile and how smartphones democratized video production, and you see that the same way as AI democratizing software. So there's this great big cavern between, I have an idea and I built a thing. For most non-technical people that gap will kill the project. They don't know where to start. Where did you kind bridge that and how did you get started in writing software that you weren't familiar with?
Marco Martinez (00:09:18):
Yeah, it's interesting because when looking at the last three years and the big AI boom and seeing these tools come out, it reminded me when the first iPhone came out in 2007, that changed everything for the internet, for how we interact with information, the accessibility to it. It reminded me a lot of when AI and ChatGPT and Antrhopic all came out and they came out with these tools that all of a sudden gave us immense access to information again. But then with the rise of agents, it gave us to this other step where it happened even faster than the smartphone revolution because the smartphones took a little bit to get to the point where the modern cell phone camera started to be a big player, but even before then, just the modern cell phone camera started to, like you said, this big canyon between your idea and the tools to create it.
(00:10:13):
It was just in the past, so big, right? I went to school for filmmaking at San Francisco State, shout out to San Francisco State, and in there I was still using a lot of the old tools, film cutting, film, learning how to edit with film, very analog. But when the digital age came for filmmaking, there was that same uproar that we're seeing now in ai, which was, oh, these tools, that's not really real filmmaking because it's not on this massive cinema camera that the Hollywood filmmakers and directors would use. It's not true art because it's not in this specific box that the Hollywood industry had been living in for a very long time. And I see the same perils within AI and software development because you're starting to see that the term vibe coding is either seen as a cool term or kind of a little distasteful to say, because it might rub some software developers the wrong way because it's like you're giving tools to everyday people that don't know how to use these really big tools because they don't have software degrees.
(00:11:19):
And it was the same thing with filmmaking. It was, oh, you're giving these cameras now everyone has a camera. They've never gone to film school. How will they know to create art and art the way we know? So you're seeing these similar parallels happening, and I just saw it immediately with AI because I lived through it early in the early two thousands, and I'm seeing it now, but at a faster rate because the tools are coming out faster and the tools are becoming even more accessible. And like I said before, agents are now giving us the ability that canyon that to exist has now become a small little land bridge that you can cross over create the idea. You have an agent that can actually create that app for you or whatever idea that you have. It's faster and more accessible to the point that it's not even, it's scary because people are just starting to get to the term of vibe coding and using these agents and tools, but I feel like it's becoming less and less scary and it's going to just start opening more and more to people having this idea and being able to execute it with an agent or just
Nicky Pike (00:12:30):
With ai. Well, I mean we're seeing that we saw the same thing. We've seen movies filmed entirely on iPhone. It's one of the things coming out and when we saw the iPhone camera come out, it is still look at the influencers, right? Influencers in social media, most of them are using iPhones and things of that nature before they progress up, before they get better at their craft. We're seeing that same thing in AI where I've got an idea, I'm not quite sure how to do it. This is going to go back to what you created, but I know what I need, but now I've got kind of a partner that I can talk with that can help me through that technical stuff and I can help iterate with them on there. And I'm guessing that when you did this, that's what you did. You went in and you treated AI like a partner and said, Hey, this is what I need, and started establishing ground rules and requirements for it, but you let somebody else do the technical work for you.
Marco Martinez (00:13:16):
Yeah, it clicked for me with the rise of AI agents, and I'll get to that in a second, but my workflow at the beginning was like the last couple of years has been just have a conversation with ai. I learn from it. It gives me the information that I need to go execute on my side. It wasn't until the rise of AI agents and them being accessible to someone like me, non-technical, that I realized, oh, I actually have a partnering crime that's next to me building with me. And it's not just a back and forth conversation, but somebody that a junior software developer that has an MIT degree can create the piece of software with me. And it clicked for me in that moment that I was like, wow, this is just big moment that we're in right now because you have the accessibility to these agents that can create this idea, bridge that gap instead of it being a canyon, being this small little footbridge that you need to cross from idea to reality, and it's just all there for you to do within a couple days depending on how you build your stuff.
(00:14:23):
It's way faster than it used to be before agents came out. So that's the big thing that clicked for me, which was like, oh, I have this now partner, that big junior software developer that has all the knowledge in the world that can create this for you that we never used to have before.
Nicky Pike (00:14:41):
Well, and it's not just coding that you were learning, you had to learn other technologies as well. You told this story about how you were creating everything and you had to create your own repo and you were everything on the main branch, and then when you were troubleshooting something, you lost everything. So kind of walk us through what else did you have to learn and tell us more about the story about what you learned from it.
Marco Martinez (00:15:00):
Yeah, it's an interesting story because when you go through this process of tinkering around, and especially not having a software degree, you go into these things thinking, oh, I can do this. I'll just follow the steps, and you're following the steps that they're giving you. But you got to remember that this AI doesn't also know how much experience you have unless you explicitly tell them, Hey, I don't know anything about software development, right? I didn't do that. I don't
Nicky Pike (00:15:25):
Know, Jack.
Marco Martinez (00:15:25):
I didn't do that from the get-go, and I have since told my ai, Hey, I am a beginner at this. Please hold my hand through this whole process. I don't know anything. Even if I even know stuff, I still want it to hold my hand. I'm noticing that when I do that, it actually gives me all the context that I need that I might miss out on. For example, early on when I was tinkering, I was creating my first to-do app with an agent that we have in house called Blink, and I'll get into more of what Blink is later, but I was using it and it's like, okay, you need to go create a GitHub account, a Supabase account and a ELL account to get this to-do app all up and running. I did all those things and it was great. I got to learn, set up my first repo, how to set it up, how to set up the licensing, how to set up Supabase for the database and Vercel to host the to-Do app. So I'm going through this whole process and my repo, it just has the main branch, and at that time had no clue about branches. I just thought, oh, GitHub repo, I'm done. It's all going to live there. So as I'm iterating and tinkering, because the whole process, especially when the to-Do app, I recommend anyone when they're starting off, create something small, create something that's easy, not low stakes, nothing that's going to be in your workflow today. You want to first take baby steps is what I'm trying to say.
Nicky Pike (00:16:51):
Crawl before you walk.
Marco Martinez (00:16:51):
Exactly. Because it really helps when you want to create the big product that I did after it to learn about all these things I'm about to tell you. So I'm creating the to-Do app, I'm iterating on it. I'm having a conversation back and forth with our agent and we're doing all these changes, and then all of a sudden it was either I asked it to do something or we committed a change and the entire app just broke and we tried troubleshooting it, and I can't remember if we nailed it down, but what I do remember was that it educated me on branching, and I remember having a conversation with you at an event we were at and you kind of educating me on branches and like, Hey, did you know this? You know, could branch your GitHub repo so that you're not messing with the main branch.
(00:17:45):
You got to treat them as your main branch is what's production ready and goes live, and then you can create a development branch or however many branches you want so that you're not messing up the main branch. That was a big light bulb moment for me because I was like, wow, okay, this is amazing. And these are the kinds of things that I'm like, I can see why people might get intimidated to learn that, oh, I have to do now branches, I have to learn about all this stuff. I guess I have a tinkering mind and want to learn more. But I think for other people that are getting into it, it's still a low barrier of entry because you have that partner in crime sitting with you, creating this stuff, right, to kind of help you and be like, Hey, mainline branch. Yep, don't do that. Create a development branch to tinker.
Nicky Pike (00:18:30):
And I think there's a lot of mid-level and senior engineers out there that say, well, this is why people shouldn't be vibe coding. They don't know the basics of get, but Pain is a great teacher. You learned a lot through that process. And I would encourage most people out there to remember, Hey, Mr. Senior Developer, there was a time there that you didn't fully understand branching as well, Marco, you're not creating a financial app here. You were creating a to-do app. This was something that helped teach you, and I think that's a good statement for other people out there. You had an idea and you went and learned and you're going to learn stuff through this. And that's kind of the process is going through it. And on that same lines you talked about with your to-do app that you had a conversation with your ai, that's how you got things done, Hey, change this, do this. But then you discovered, well, the AI is fairly smart. If I give it a product requirement document, if I really write out a specification for what I'm doing, which is what I think you did with your Discord app, that it made things a lot better. You actually said it took you from like 60% up to 90% accuracy on your app once you started doing the PRDs. Walk us through how you figured that out and what changes you made there.
Marco Martinez (00:19:33):
So another bit of credit to you, again, we were at another event and I was telling you about my idea and you're like, Hey, have you thought about PRDs? And I was like, what's that? And it's a document that software developers use to create their own software, but you were reshifting the context around, Hey, use it to build with ai. It has a better time understanding you when you can feed it this PRD because AI knows that because it can also, it's a really great software developer and it understands how to software develop with PRDs in mind. So when I started to tinker with it, as soon as you told me, I was like, it was a light bulb moment for me. I was like, wow, this makes so much sense. Because before it was just a conversation of I was going back and forth with my agent and the context window was just getting longer and longer and longer.
(00:20:27):
And right now anyways, the way AI agents or just AI in general is built, all the interface is the ui. You're scrolling and scrolling and scrolling just to see exactly the conversation you had to fix there and you have to scroll up to find it. And then hopefully you can reiterate it back to the AI that was to me before was fine. It was a way of like, oh, this is how you work with ai. This is the most efficient way. I guess when you introduce PRDs, it opened up the ability to not have to keep scrolling up and down and that I can, basically, what I did was after you told me about PRDs, I went back to my ai. I'm like, Hey, pump out a PRD doc for me, based off of our entire conversation, it was able to pump out the whole, it organized everything, the goals, what not to do, everything, all the context, all my tokens, all of that was just amazing because I was doing that already, but just not in a formal way.
(00:21:31):
And it was just a lot clunkier. It's funny before I see it now as a clunky way of developing software where now with working with AI and PRDs, I was able to create that PRD doc be more efficient in the way I was iterating on my Discord app because if something didn't break, I'm like, oh, let me just go into my doc, tweak it, and then I didn't have to go, Hey, do you remember back two days ago when we were talking about this? And the context token window is massive. Hopefully. Hopefully you'll remember where with the PRD doc, I just edit it, what I needed, what wasn't working, updated, the PRD doc uploaded it back to the ai and it was just a faster workflow. I wasn't stuck hoping it was going to be able to remember what I had asked it before going to now, oh, I remember exactly.
(00:22:26):
I have all the information. It was just a workflow was at least like you said, from 60 to 90% faster. And now that's the way it worked. Now, whenever I create new projects in ai, I create a PRD doc for everything that I'm working on. I'm working on a new app for a screenwriting app, and the PRD doc has been tremendous in my workflow there, and it's sped up the development process. And even I told a couple friends of mine who I've been encouraging to get into AI and AI agents about the PRD doc, and you can just see their eyes just light up light bulb moments in non-technical people, and it feels like more education around PRD docs to non-technical people. I hope you're listening because it'll change the way you develop software in AI and even on just the basic level to even complex apps. So I highly encourage anyone to use PRDs.
Nicky Pike (00:23:22):
Well, and the great thing is, is that PRDs are going to give you a little bit more repeatability. If something does happen to your hardware environment where you're developing this, you've got that PRD to back you up, and if you need to hand this off to maybe somebody you've got issues you can't fix, you've got a full product spec there that a more senior developer or more educated developer is going to be able to look at and understand what you were doing and what you were attempting to do. So I do think that's a better way because then you're not really being the software developer, Marco, you're being the architect. You're saying, here's the problem I need to fix and here's how I kind of want to fix it rather than actually writing code, which is, that's kind of the basis around vibe coding, and I think it's way we're going to see enterprise going in the future.
Marco Martinez (00:24:07):
That PRD doc really bridges the gap between giving that PRD doc, for example, the screenwriting app that I was working on. I could give the PRD doc to my wife who is vibe coding as well, and she can take that and just feed it into her AI the same way I can give that PRD doc to a senior software developer. It's that same level of doc. It feels like the PRD doc is the bridge and the bridge to non-technical, super technical that it can understand where you're at in your development process without you having to be like, Hey, I have this app and here lemme just share my AI chat with you and see if you can get it from there. It is something that that workflow would never work with a senior developer, even somebody would not want to cipher through all of your 10 day chat history, right? No one's going to cipher through that, but if you have a PRD, it's a simple doc to go through. It's laid out in an easy digestible way for the technical and non-technical to see what you're working on and where you're at in your development process.
Nicky Pike (00:25:13):
Well, and I'm sure there's people out there that are listening going, but Marco said he used AI to create his first PRD, and that's not really creating A PRD. Well, no, it is, right? You learn from that. Now when you go and try to do other projects, you've got an idea of what A PRD looks like, and now you can modify and iterate on that PRD. There's nothing wrong with having AI help you with something. It's no different than going to your senior engineer and saying, Hey, give me a template by which I can go in and do this, or go into your project manager who gives you a PRD template, or here's the requirements that we need. It's the same thing. Again, we're just replacing the human pair with the virtual pair and this all worked for you.
Marco Martinez (00:25:50):
Yeah, exactly. And I see when you bring up like, Hey, Marco, that's not the way to create PRDs, right? It's the same kind of rhetoric that you get with, Hey, that's not how you film a movie. Have a cell phone like this. That's not how you film stuff. You have to have it in landscape. And what do we have now, people being okay with filming it in vertical video. That's a big norm now. So I feel like the same is going to happen in AI even faster. So yeah, I'm sure the PRD doc is not, it was created in ai. I get that. But like you said, I understand what the PRD is now and can start from that from the get go instead of having to have a back and forth AI conversation and then create a PRD from it. I think there's just now multiple ways to create PRDs for the people. So I find that very refreshing and it's just opened up the way I now develop software.
Nicky Pike (00:26:51):
Well, and here's the part that got me really excited. I mean, after you went through this process, you said that AI actually increased your interest in trying to understand the underlying tech. So now you're planning on taking some Linux courses, you're looking at programming formal program study in 2026. This is exciting to me, but why go deeper? I mean, why do you want to go deeper instead of just staying at the vibe coding later?
Marco Martinez (00:27:13):
I think it's the same, and I think I'm going to keep making these parallels between the cell phone camera and this. When you get a tool that is so robust and you can at least master level one, which is hey, turning it on recording, being able to film that, which was for me, oh, working with an AI agent, it suggests these tools for me to use, and I say yes because more than I do, and they worked for me. I go, okay, I've achieved level one. But my inquisitive mind goes, Hey, is there a better language than Python? Is there a better app than superb base? Or is there a better workflow? Is there a better LLM model than the one I chose for my Discord bot, right? I went in with doing some research on some stuff, but I know that there's better languages, better ways to set up my coder workspace.
(00:28:14):
I just went with the default because it was kind of the intro level. But after going through the whole process, I was like, whoa, I can't wait to start diving into what's the better language? Is a Docker container the right way to set up my workspace? Should I use a Kubernetes cluster? Should I use a vm? Whatever it is, I'm just like, I need to get into tinkering more. And the way to tinker more is to learn more about those programs. And I'm not going to sit here and tell you by the end of 2026, I'm going to be a full fledged software developer, but what I am going to be doing in 2026 is at least getting that first level knowledge of like, Hey, this is what's behind there, right? Linux, what's behind this is go, this is a different language that you can use.
(00:29:00):
Maybe use a different IDE, maybe use all of these tools because there's always more efficient ways to create one piece of software. Everybody has a certain way they're doing it. And I just found that now that I have the first level, the foundation set, it's made me go like, wow, I need to start learning some more stuff so that when I go talk to my AI agent, I have a little bit more knowledge, and I think the AI will reward me with that knowledge because it'll be like, oh, you already know this. Let me suggest this. So
Nicky Pike (00:29:34):
I think that's the part that most people are missing, right? AI is there to help, but it's also encouraging people like yourself to actually go out and learn more. The people that say, well, if you don't know what you're doing, how can you be doing this? Well, this has given you the start. This has given you an idea and it's interested you, and now you want to go out and learn more. And I was sitting there laughing the whole time when you're like, there's a better language. Well, now you're going to get into the whole Python versus Java versus all these different frameworks are going to come yelling at you and saying, well, we're the best. You should use us. But yeah, it really is kind of one by one. Alright, you told me about the journey. Now I want to see where you took this, right? You've got this agent system that goes into Discord, it runs things through Llama and Slack and back. It's time to show us how this actually works. Are you ready to put up or shut up?
Marco Martinez (00:30:20):
Yep, yep, yep, yep. I'm
Nicky Pike (00:30:22):
Alright. Well, let's talk architecture. So first walk us through the flow, Discord message comes in and then what llama slack Discord, what happens here? Do you have a diagram? Anything that you can show the people that shows how these pieces connect?
Marco Martinez (00:30:36):
Yeah, let me share the diagram really quick. So I created this diagram so that it would help me with the whole workflow. I built it out where a message comes into our intros channel. This is the current workflow it goes, then the message would go into the Discord agent that reads it and crafts a response. From there, it crafts the response using llama ai. And I chose llama AI because the 3.1 Instruct model was really good at its open source and it allowed me to, and it was a good model to read and write content for me in this small way. So it reads and crafts the personal response, it shoots it back to the Discord agent, which then sends it to Slack. I was mentioning earlier, slack is where I'm at like 90% of my day. So it shoots a notification in there, and then from there I can review the message and the way it's created. It's like, Hey, I have some options. Accept AI generated message, edit the message or skip because I didn't want to lose the human aspect of it. What I'm using this spot for is so that it can at least get me part of the way there.
Nicky Pike (00:31:54):
So you're kind of keeping the human in the loop there. You went back and said that being personal was very important to you. So this is something to help automate the workflow, to give you an optionality of what you want to say, but you're still keeping the human in the loop. You're still making that personal choice on what the response is going to be back.
Marco Martinez (00:32:10):
Yeah, exactly. So you make the decision to edit within the Slack bot. If you say no, you can skip the whole message of say somebody else messaged or it was like a spam message or whatever it is, but it gives you the options to accept the AI message, edit or skip. So then once that's all figured out, it shoots it back to the Slack bot, and then if it's approved, it sends it into Discord. The Discord agents creates a thread in Discord and replies as the coder welcome team. That's awesome. And then from there I reply with a follow-up thread. So I wanted to make sure that even though the message is coming from a human through an ai, I still wanted to come in as the community manager would a follow-up comment, which has been successful because it creates this conversation more than just the coder welcome team, but it's also followed up by the community manager and people have really liked the way we've communicated there. Very nice.
Nicky Pike (00:33:08):
Was this how you started off your project was by making this flow?
Marco Martinez (00:33:11):
Yeah, I created this workflow because it was like, I've seen other software developers do it, and I was like, oh, I think this will help. I'm a visual person when it comes to ideas like this. And I wanted to create a diagram for this that hey, I could follow and I could also pitch it to other people and be like, Hey, this is how I was working and thinking about the agent. And then this was just a little bit more of how it works with the agent, the workspace, and why I ran it this way.
Nicky Pike (00:33:43):
Okay. Well that you consulted multiple ai. So there was blink, you talked about a llama, you talked about Claude, and then you came down and you settled on an Llama model for reviewing the messages. So walk us through how you did that evaluation, because I'm going to take it, you're not an AI expert, so how did you pick the model when you don't have an engineering background?
Marco Martinez (00:34:05):
I had to pick a model with the help of ai. I wanted to, the stipulations in and anything that I create is how can I do this a hundred percent free or near a hundred percent free? And at the time, I didn't realize internally that we had API keys from Anthropic and OpenAI built into our Blink agent.
(00:34:28):
And so I was like, alright, that's not working. I need an open source model that I can use and won't have to worry about token usage or anything like that. So I did some research, found out that the 3.1 Instruct model from Llama was really good at reading and writing for messages like this and longer social messages, which was kind of where I landed with this model. I found a couple papers on it that it's like, yeah, that backed that this model was created for those reasons. After that, I didn't really tinker with anything else because it was really easy to set up on my Mac and easy to integrate into this whole thing. So I went with 3.5 and didn't really tinker around with anything else after that.
Nicky Pike (00:35:10):
So are you hosting this model alongside your application? Where are you running the model at?
Marco Martinez (00:35:15):
Yeah, so the model is all running within a coder workspace, and I don't think we talked about the journey into creating that, and I'll just be pretty quick with it. When I was developing this whole bot, I wanted it to run continuously. I needed it to pay attention to the messages that were coming through and didn't miss anything. Now, I created that first locally, but what I realized quickly was, hey, it created this wonderful app. It can look at everything, look at all the messages that's coming through while my laptop was up and running, and that's all great and all, but for 12 hours of the day, I'm off my computer asleep and I need that thing continuously running and sending me notifications. I didn't want my laptop up and running all the time, I'll admit. I was like, yeah, that makes a lot of sense. I also work for a self-hosted cloud development environment company that I could put this up into and have it run continuously and just edit in the cloud, edit in a terminal through my Coda workspace. So from there told Blink our agent, which let me just give you a quick little insight of what Blink is. Blink is an open source framework that you can build and deploy AI agents on-prem simply by using natural language
(00:36:41):
For what you want. And it generates the AI agent for you. It handles the infrastructure behind the scenes and lets you run the agent locally or in the cloud. That's blink in a nutshell.
Nicky Pike (00:36:53):
Well hold on there. So this is a two code episode. So you're telling me that you use Blink to help you write this. Can you show us what that looked like and how that conversation
Marco Martinez (00:37:01):
Went? Yeah, so let me share. So here in Blink you have your chats on the left side. It's a very similar kind of workflow as any AI kind of open AI anthropic interface. And from here, as you can see this long chat thread that I have in here, but basically I asked it to create the app for me. It then goes in and creates the Discord agent, it creates the llama agent to come in and bake it in. And throughout this chat conversation, it went in and basically gave me all the tools that I needed to set this up and set up the coder workspace.
Nicky Pike (00:37:45):
So Blink in this instance helped you, so now there's been a lot of talk about Blink and how it's going to be the agent to write agents. Is that kind of how you use this here? But you also use Blinked out actually write the code itself?
Marco Martinez (00:37:56):
Yeah. Yeah. So hold on, let me get down a little bit more. But as you can see here, it wrote some of the code, it chose the language, what to write in. I can bring up my PRD doc here.
Nicky Pike (00:38:08):
That was going to be one of the questions I was going to ask later is just show me what the PRD looks like. But
Marco Martinez (00:38:12):
Yeah, exactly. This is actually easier. As we were talking before, scrolling through that whole context is a lot. And so with this PRD doc, I was able to ask Blink to create this for me. It broke down the executive summary, my pain points in an easy way, the primary goals to respond a hundred percent of Discord intros within five minutes. It creates the success metrics based off of all my conversation that I had,
(00:38:39):
The nongo replacing human community management entirely. That's not something we want. It created a persona, which was me, and then we went down with like, Hey, this is the architecture overview and this is what I loved, that it broke it down in all these ways that like, Hey, this is the model that we're using. I want it as a Discord agent. And then here are the external integrations. It's so easy to understand and write. And then after that, we get our technology stack. We're running in no js, the language is TypeScript. And then as you can see here, version Troll and GitHub. And then it basically goes through the whole workflow here, which
Nicky Pike (00:39:18):
Now have you learned anything from this PRD? Are you using this? I know, and we're going to talk about it here in a little bit. You've got some other projects coming up, but have you learned anything about the PRD about if I modify this has an impact on my AI, or I get different outcomes if I change certain things? Are you really deep diving into using requirements to drive AI versus conversation?
Marco Martinez (00:39:40):
Yeah, entirely. When I'm having my conversation with my ai, I'm feeding it the PRD, and then I'm seeing that 90% success rate with it. And if it's something that's not working or I'm not seeing something in the Discord agent or whatever app that I'm creating in, I'll go here and tweak the PRD in the requirements and it'll be like, alright, must log all intros detected.
(00:40:03):
If it's not logging that in, I'll be like, Hey, please, please double check all intros detected. And then I'll go into more of why I am doing that, and then I'll feed it back into the ai. And I'm seeing the fixes be implemented way faster than me just saying, Hey, I noticed this when I was using the PRD to make the changes. I was seeing my success rate go from just the conversation and saying, Hey, can you please double check all the intros detected? And just putting this sentence into Blink and putting that in here without the context, I was seeing a 60 70% success rate in implementing it. And not just with Blink itself, but with any AI or AI agent that I was working with, because I've used other agents before and they all kind of have similar ways of working. And then when I was feeding the PRD doc into this and saying, and adding this line here that I just added, it would have all the context and be able to digest it easily. It knew how to read a PRD doc so well, and it could just implement it and be like, oh yeah, I made that fix for you. And I was seeing a 90% success rate in the way I was developing this software. And so, oh, go
Nicky Pike (00:41:25):
Ahead. So give me an example. When you're writing this and you've got your PRD out there and you're making changes, there's got to be, you had to run into some times where AI didn't completely understand what you were telling it. So how do you debug something that you didn't write? I get that you got your specification, but the code that's coming out, I know that by your own admission you don't fully understand the code. How do you troubleshoot something that's not
Marco Martinez (00:41:48):
Yours? Yeah, a lot of that workflow is honestly through the AI itself. So I'm like, Hey, look, could you debug on what's going on? The app is not reading the messages or I'm not getting a notification anymore. Or there'll be all of a sudden a new update on how, as I was showing you in the diagram, it was giving me three options. When the message came through, I was starting seeing four options and then two options, and we were debugging that way. So I was using the AI to debug and tell me, Hey, have you noticed anything? And then as soon as it's like, oh yeah, I'm noticing this, I would go in and then edit the PRD doc with updates here and just be like, Hey, here's another error scenario that I just noticed, and I would add it here. And then from there it would be like, okay, once we had fixed that, the fixes came through and I was seeing better a response rate there.
Nicky Pike (00:42:46):
And I know there's going to be people saying, well, if you don't understand the code, how do you know it's secure? Well, let's be honest, right, Marco, I mean, you're acting as a supervisor here. You're relying on your AI to go in and do this stuff for you. You're not claiming to be a software developer. But on the flip side, you're not writing a payments app, you're not writing something that's holding secrets. So have you been in a company where you got a lot of really great engineering talent? Have you had any of your engineers kind of push back on you and make fun of you or say, well, we should probably check this over to make sure that this is working right?
Marco Martinez (00:43:20):
No, the people that I have shared it with, the engineers that have been like, oh, that's really cool that that's working. I think to your point,
(00:43:28):
The information that is going on in this isn't like a payments app where you have a lot of secrets going on. It isn't a big piece of software that is controversial in a way that I'm handling the information a certain way. I am looking to partner with our engineering team once we get this into production, because right now I just have it working on my coder workspace, but we haven't put it into an EC2 instance that could run 24 7. So I think when we get to that point, I'll probably bring in an engineer and ask them to take a look at the PRD, see if there's any ways I can iterate on it or work on it better and then get it production ready, not there. I can admit that for sure, but I do think I need to bring in an engineer once it is production ready and that we're ready to bring it into an EC2 instance.
Nicky Pike (00:44:17):
Well, you mentioned something there, right? You're looking, we've got this running right now. You're basically in beta mode and you're running this out of a coder workspace. But you talked about that progression. You started locally and then you were like, Hey, I work for coder. I'm going to move this into a coder workspace so that I don't have to rely on my laptop. And now you're talking about moving it into EC2. So is there any of those steps that you wished you would have maybe skipped over or what led you to making each one of those steps as it comes to the progression of this application?
Marco Martinez (00:44:45):
I wish I would've just skipped over the whole running it locally on my laptop. I mean, that's the biggest one, just shooting. And I have now I just straight go to a coder workspace and work within here just because it's easier to go into my terminal window here and edit my workspace and go in and edit through the terminal window here. I just find it more efficiently than working on my terminal on my Mac and editing through there. So that's the biggest one is like, Hey, just jumping into the coder workspace, because what I was finding is that I was going to eventually end up here anyways. And so if I had just skipped that step and just go into a coder workspace and build from there, it just bypasses all of that local setup that eventually I was going to toss because I want to eventually have it in an EC2 instance or whatever cloud provider that you decide to put it on. But I wanted it run in the cloud 24 7, and that's going to be through coder workspaces.
Nicky Pike (00:45:41):
If you're running this in beta mode right now, you're not turning it off. But one of the other advantages it would seem to me is having that kind of ephemeral nature, turn it off and on, maybe being able to access it from my laptop when we're at an event, you want to talk about an issue or maybe have the engineer go in and be able to jump in and take a look at it. Did you find any of those values around using the coder
Marco Martinez (00:46:01):
Versus Yeah. What was great is that, for example, I shared this co of workspace with you at an event and you were able to easily get into my co of workspace diagnose what was happening, and you recommended, you're like, oh yeah, this is how you need to tweak some of the settings in your co of workspace so that it doesn't shut down every day. Because I think the template that I had was shutting down every day just basically because I didn't have it set up the way I wanted to when I first was kicking this off. So you helped me get into my workspace, and then for example, during the holidays, I had it shut down and it was easy to bring down and bring back up because Coda Workspace are ephemeral. So that was another great thing is that when I do want it down and not working, it can shut down and not worry about it and then bring it back up when I need to. So in that instance, it worked really well.
Nicky Pike (00:46:48):
Alright, well fire it up, man. Show us. Let's see what this thing does and how it works.
Marco Martinez (00:46:52):
So what's really great is that today we actually got somebody in the interest channel say Hi. So we'll go to Discord. As you can see here, Renee at 10:31 AM said, Hey everybody. I'm a solution architect at IBM and I'm trying Mux for Agent Multiplexing and see how it goes. Now this is exciting because I think this might be our first person talking about Mux, which is really interesting. But what I'll show you here, so the message comes in here and then based off of my diagram, it then will shoot a notification into the, I have this spot here, Discord intros approval, right?
(00:47:28):
So I'll click here, and then you can see here it says who it's from, what channel, and then it says it brings in their message and then it says, AI suggested response, right? And it creates this response here. And what's nice is that it's actually pretty good. It's a fantastic to have solution architect with IBM here. So then exploring M for agent multiplexing. What challenges are you aiming to address by integrating with this technology? I think that's pretty good, but I want to edit it. And what I'm noticing, and I'm probably going to have to iterate on this, is that this follow-up question isn't something that I have in the PD doc, it's actually doing by itself. And I actually don't want to do this because I will be following up on my own. And this is something that I specifically had in the PD doc not to do. So I need to go edit and debug this. So this is actually a nice little example because you're seeing, hey, it did part one, great, but it created a new bug because this shouldn't be existing. So
Nicky Pike (00:48:31):
It went rogue on you,
Marco Martinez (00:48:32):
But not too bad it didn't send the message that's completely rogue. But down here you can see three responses, send AI responsive. I just wanted to send this response or skip for this example, I'm going to hit write response. I'll go in and then what I like to do is tweak this and be like, Hey, welcome to the coder community. And then I'll tweak the messaging here. So because that bug here, and then I'll say, it's great to have you here and learning
Speaker 4 (00:49:07):
And using Mux. What project are
Marco Martinez (00:49:13):
You working on right now and how did you find out about M? Okay, simple as that. Click save, and then you can see it updated here. And then I'll just hit send. Okay, so it says response sent, and then what this does, and our Discord, as you can see now, it actually created a thread, which is really great, really.
(00:49:44):
And if we click in here, I have it created a thread because what I was told early on, it's better to have threads than just straight replies. A lot cleaner for going through the timeline. But as you can see here, it created a thread are coder welcome team here with our little astronaut mascot comes in and has the message. And then what I'll do next is be like, hi Renee,
Speaker 4 (00:50:10):
Thanks for joining our wonderful community. How did you tell me more about your
Marco Martinez (00:50:22):
Project? So then you have, like I said, you have the coder welcome team saying Hi. Then you have the community manager saying Hi. And it creates this chain of communication going, Hey, there's multiple people saying hi to you and encouraging you to divulge more information about what you're working on and sharing that in a nutshell was how it works. It looks very simple, but in the backend, I have the Discord bot, I have the Slack bot working. I have an LLM from Llama all running in a coder workspace. And it's a little bit more complicated, but it's executed very simply, which I really love. So that's the Discord bot in a nutshell, and saved me a ton of time because I'll get that notification here on my phone, and I am able to just respond on my phone super easily and even on Discord on my phone as well. So it's just made my workflow in responding cut down on that 10 hour gap, four hour gap to a more reasonable time. Now, I don't know, I haven't timed out my response time yet, but it's a lot better than it used to be.
Nicky Pike (00:51:34):
Well, and just being able to keep track of the new people coming in and like you said, creating those threads so you got some follow up. I think that's awesome, man. I love everything about it. I do think one of the things that I would ask is, this is something that you've done. What part of this still kind of feels like dark magic to you? I imagine there's still parts that still kind of feel like they're way above your head. Are you still running into things daily? Are you running things weekly? What feels different?
Marco Martinez (00:52:00):
I think it's the debugging. I'm like, oh, I just noticed that part. Why is it doing that for me? I want to know why it's doing it, so then I can improve on the way I update my PRD. And then learning, okay, hey, do I need to change up maybe a different language or how it's set up so that it doesn't create these bugs? It's having me want to learn more about the debugging process and implementing it properly. So that's the kind of first thing that comes into mind on a daily basis and weekly basis that seeing
Nicky Pike (00:52:32):
Well, and you've got visions. I think I heard you mention that you're working on a screenplay feedback app that you're wanting to build next. What's the vision for that one, and are you going to go through the same process for that?
Marco Martinez (00:52:42):
Yeah, so I'm already going through the same process If you want to feed your screenplay into ai, as we know, sending 120 pages with dialogue and descriptions takes a lot for AI to be able to give you feedback one-to-one feedback because
(00:52:59):
Screenwriters are now using AI to get coverage for their screenplays. And coverage is like feedback to tell you, Hey, this is going to sell and it's not going to sell. And so when my wife and I are, my wife's trying to be a full-time screenwriter and eventually I would love to, and what I'm noticing is that when I would upload the full script, the AI was having a hard time knowing the page number because the software that we write screenplays with doesn't attach the page number to the exported file that you have only in one format, which is PDF. And so I was wrapping my head around, I was like, God, do I need to create an agent to match the page number to the actual text file so that when I ask it, Hey, what's wrong with scene one? It will go, oh yeah, this is what's happening in scene one, page two.
(00:53:52):
Because what was happening before is that it would almost basically misplace the scene because the context that it had for it was guessing where the scene was based off of where it could be on the structure. So I was like, okay. I was like, no, an agent's going to be too much. What do I need? And I was like, oh, I need an app that can basically read A PDF because that has the page number and then read a text file of the screenplay, all both created with the screen rating software, final draft. You get those two files and I created an app where you upload the PDF, you upload the text file, it merges both so that it has the page number baked into the text file for some reason. You can't export a text file through the app to have the page number.
(00:54:42):
And then what it does is that the AI then reads both files, merges it into a text file that AI can properly read. And then that's what I feed into the AI that I'm working with. And I've had a hundred percent success rate since then. And I was like, how is no one else doing this? What's going on? So I'm thinking about open sourcing that app here soon once I have it production ready. But what's nice is that I learned from everything from my Discord app with the PR building. I did that same thing for this app and I was able to run, I created the app in two days where the Discord app was done over a couple weeks, I was learning everything. That's the gap that PRD kind of cut down. I went from two weeks of development to two days of development, and I was able to create a beta app with two days because I had a PRD doc to guide me there. And I think that's a tremendous leap in such a short amount of time.
Nicky Pike (00:55:40):
Yeah. Well, and that's the democratization of software, right? There is people that have great ideas. To your point, why is nobody doing this? I'll make it go happen myself. This is what that's going to actually allow to happen. And I think for those that are watching here, thinking that I can do this, I do want to take this practical. For those that have zero technical background, just like you, they've got problems that they need to solve. What's step one for them? Where do they actually start? What is your recommendation?
Marco Martinez (00:56:06):
First step, go tinker. Go into your AI and have a conversation. Ask it what's possible, do a little bit of research, just ask it some questions. Always double check your work. That's the biggest thing that I recommend. Don't just go off of what it says, double check, see if it's actually going to work. But yeah, go in and just test out your idea with the ai, see if it's possible. I think that's the biggest barrier is that people have ideas, but they're not willing to take the jump by just simply asking a question. Sometimes a lot of people are afraid of asking that simple question of, Hey, has anyone done this or can I make this app happen? Tinker with a small idea first so that you get the foundation under your belt and then the app that you really want to create, save that for the next iteration of your idea. Because I think if you tinker with something small first, it'll give you the foundation that you need to really go in for the big one.
Nicky Pike (00:57:08):
Nice. And you're kind of blogging about this for yourself and the journey that you're taking between these apps and how you got started with AI programming. And this is on coder.com. Tell us a little bit about the blog.
Marco Martinez (00:57:19):
Yeah, so I've been documenting my entire journey, it's vibe, coding journey with me. The first one was on my to-do app like we talked about in this. And then the second one that just came out just before the end of the year was around the Discord bot and what I went through. It's very high level, which is great, and it breaks down my entire thought process of getting in and building the app. I'm going to have a follow part two of that app where it gets into the really nitty gritty technical details of it, why I chose certain models, the code behind it and some of the decisions behind it. But yeah, you can follow that journey right now on coder.com/resources and go in there, find my blog. Also, what I also recommend is going to coder.com/chat and joining our community there because there's a lot of vibe coders in there as well.
(00:58:12):
Sharing ideas, tinkering, doing what you're doing, learning just as much as you are. There's an array of people in our community from super technical all the way down to my level of the simple Vibe coder. Yeah, you can follow the journey on coder.com and I'll be blogging throughout the year on this vibe coding journey and blogging about other ideas that I have around our community and just overall software development because I'm finding that I'm becoming an accidental software developer that's learning, and I'm hoping with my blog that it'll enable people that are a little hesitant to jump in to at least dip their toes in. And then once they do jump into the deep end like I did and go learn
Nicky Pike (00:58:58):
Well, and I think you make a great use case. I mean, like we said, lovable 6.6 billion valuation for people that are wanting to do vibe coding and no code. So having the ability for coder things like coder out there that has tie-ins to AI where we have an open source component to it. These are all, I think it's a great use case. I think it's going to be inspiring. We'll make sure we put some links on there in the description when we about this later so that people can follow along and definitely keep in touch with Marco as he goes through because he's got a lot of good stuff coming. And with that, Marco, we end the episode with the same questions every time. So you know this, you started from scratch, you learned how to navigate software development and you used AI to bring all this to life, and you did this without an engineering background. So keeping that in mind, what does it mean to you to be a coder?
Marco Martinez (00:59:45):
It means that I get to create software from ideas that I've always had and create software in a way that I never could before. It's a great time to be alive right now because the tools are there to make even the smallest idea into a reality and in a reality within a couple days. And I think the progression of being able to have an idea to creating something is going to get faster and faster and more efficient, that it's going to be a thing of the past where we're going to be thinking about the term vibe coding in a negative way or seen as slightly negative, and it's just going to be a secondhand way of just creating the same way the cell phone was creating easy videos. I think we're going to get to that point where you're going to have an idea and you'll be able to spin a very well made app. And I think later down the line, it'll be so good that you'll be like, oh, I can actually make some money off of this if I really wanted to. So that's what it means to be a coder in and today it's that I can create and I can have an idea create something within a couple days, and I think that's amazing.
Nicky Pike (01:00:59):
Yep. Well, and you keep comparing this back to mobile filmmaking, democratizing video. Give us the spicy version of this. What percentage of professional developer work do you think could be done by domain experts with AI tools in the next five years? Feel free to give me a number that you might think pit will piss people off. How do you think that the average citizen is going to take ai and do you think that you're going to see people that are not professional developers out there making software?
Marco Martinez (01:01:25):
Yeah, it's interesting. I keep going back to that analogy that I was around cell phones and making videos with it, and you see that huge spike from when that was first introduced to all of the people making stuff on YouTube. I think we're going to see something similar but in a quicker pace where within five years I could see a healthy 40 or 50% of people are going to start having ideas and creating within the next two years. And then I think the next five I could see that jump within 70 to 85%. And that's an aggressive amount. And I think it's because of how quickly we're getting these tools and how younger generations are already creating stuff.
Nicky Pike (01:02:18):
That's a good
Marco Martinez (01:02:19):
Point in a faster way than my old brain is able to create, right? Younger generations are getting access to these tools faster than we ever did, and their minds are just pumping out stuff left and and I think that's why the high percentages of 70 to 85% in five years could be realistic because of the trajectory we're on with the tools coming out so fast and the high quality of those tools coming out.
Nicky Pike (01:02:47):
Yeah. Well, I think that you've got a very high number when it comes to creating stuff, but you bring up a good point. We've got kids that are coming out. As technology comes out, kids are picking it up much faster than we are as the older generation. So I think the utilization of AI is going to hit that 70 to 80% number. I don't know that we're going to see 70 to 80% people creating software on their own, but I do expect that we're going to see this explosion of ideas just like you did. People that want to fix problems or they want to make something work for them. So I know I don't have to ask you this, you helped create this channel, but I'm going to, anyway, we're going to consider you a full-fledged member of the [Dev]olution.
Marco Martinez (01:03:26):
Yes. Oh, definitely always been through and through. So can't wait to come back and chat more about the future of ai.
Nicky Pike (01:03:34):
Yeah, one of the original [Dev]olution members. All right. Well thank you, Marco. Love it. Look forward to looking for more from you, and we'll make sure that we get your blog posts down in the comments so that people can follow along. Until next time, thanks everybody.
Marco Martinez (01:03:47):
Thanks everyone.
Nicky Pike (01:03:48):
Alright. Thank you for listening to [Dev]olution. If you've got something for us to decode, let me know. You can message me, Nicky Pike on LinkedIn or join our Discord community and drop it there. And seriously, don't forget to subscribe. You do not want to miss what's next.