Frank Wiles - Founder of REVSYS

Welcome to another Founder Wisdom podcast. We have Frank Wells with us today. He's founder of RevSys. He's an expert in building, scaling, maintaining complex web applications in Python, Django, React, PostgreSQL on Kubernetes. Live in Kansas, Frank, glad to have you on. Tell us a bit more about yourself and what you've been up to.

Hi, uh Thanks for having me. Yeah, so we've been in business 19 years. I've been involved in open source since the late '90s, back in the Perl community and modperl, way back when the entire web ran on Perl. And then Django was created actually a couple blocks down the road from here, here in Lawrence, Kansas, of all places. Little hippie college town in Kansas. And the newspaper there built it for their own website, for their own web purposes. And we convinced them to open source it because at that point, Python didn't have a very compelling web framework in the Python space. And then Django kind of took over and is still the predominantly used web framework in Python.

Right. And bringing that to nowadays with vibe coding and all of this, do people still need dev agencies like yours or could they just build it themselves? It's easier nowadays.

I mean, you can build a lot on your own even if you're not much of a developer. The problem is continuing it, maintaining it, making it run well. If it's a personal tool running, you know, Replit or something for yourself, like, that's fine. I have a couple of those that I've built that are entirely— I don't even know if I've looked at any of the code— that does the job for me. But when your company needs to depend on it, you know, as not a software developer, like, do you have backups of your data? Could you restore those backups if you did have— like, even if the the Vibe tooling had built a backup for you? Do you know where it is? Can you restore it in a hurry? The other area is that, you know, as software gets more and more complicated, Vibe coding becomes less and less one-shot. People are very fascinated. I think it's an interesting test if you see Simon Willison or Ethan Mollick and people like that saying, "Hey, I told Claude Fable, build me a 3D platformer game like Super Mario Brothers," 20 minutes later, here's a game. That's great. But 18 months down the line, when you're trying to edit this one image in level 17, vibe coding starts to fall apart a little more. You know, you can't— if you don't know what to tell the LLM to do to achieve the goal, right, it's harder and harder to get that tool to not build worse and worse code over time. So we've done a lot of what we call rescue projects in the past where a team of not very experienced developers kind of built a mess and the company is now depending on it, right? And so they come to us and say, hey, can you kind of clean up this mess? I expect that in the next 6, 12 months, we're gonna have a deluge of those sorts of projects where 3 guys in a department of a large company built a useful tool, but they're having problems, they're having problems keeping it running, and they'll reach out to agencies and say, hey, can you make this kind of more real for us? Can you do some real engineering here?

Interesting. And the agency itself, is it growing? It's been 25 years now, like—

19 years that we've been doing this, that I've been doing this full time. We don't grow as fast as we could. Because mostly just my decision, I could grow this bigger. I just don't want to manage 50 people. It's not my goal. I never want to be Tim Cook or, or somebody like that, you know, like that's never been my goal. That's— I would have exited long ago and let somebody else run it. So I enjoy keeping things small and manageable to my my lifestyle as much as, as much as income. But, uh, we could grow. Um, we use subcontractors a lot, um, to kind of handle differing amounts of, of work. Um, so right now we have the most subcontractors we've ever had, um, but they're a stable of subcontractors we've been using for years. One person for like 18 years. Um, so they're, they're almost employees in everything but name and paperwork and, and that, um, which makes it easy for me to decide, you know, these 3 people together would be really good on this project, and I can just kind of set them up and they can go run with it, and I don't have to be a bottleneck in the project.

Interesting. So do you run on the roster of, what, 10 clients at a time, and the team is under 10 people as well?

Um, it's—

it—

the roster varies. Um, we probably have more more than 15 or 20 projects going kind of any given point, some of which are very much in maintenance and operations mode. So often a company— we have a bunch of projects from universities and things. They've done this research project, they've put all the data together, they've built the site, they just want somebody to keep it running. And so we're on call to make sure that, like, uh, what example is, uh, we took over the Nuremberg trial archives from the World War II Nuremberg Trial for Harvard Law Library. We've done some updates to that. It was all built in Django. It's neat. You can zoom in and see all of the discovery documents and all the evidence. It's broken out by trial and who the defendant was. You can slice and dice everything. Very interesting for researchers. We're running that for them, hosting it for them. If something breaks, we get paged and we fix it for them. So, it's not always day-to-day development.

And you said your goal in life was never to hyperscale. So, is that why you started the company 19 years ago? Was that your expectation?

No, my goal probably changed a bit. I was very much trying to follow the 37signals model of doing some consulting, and out of that, figure out a product to build, grow that product. I still have product ambitions, but I do think that after the product got to a certain size, I would step back or maybe pull like Mitchell Hashimoto did with Terraform and Vagrant and then HashiCorp. He kind of stepped back into an individual contributor role after a while, right? They brought in a different CTO and CEO because I just don't wanna do that all day. I'm still 80, 90% technical, right? I was writing code 10 minutes ago, right? I don't spend much time running the business and that's my goal.

Interesting. Yeah, what does a day in your life look like? I see you post your meals on LinkedIn and you post beach photos. I love to cook.

Many of my tech friends, I thought maybe I'm annoying people, and they're like, no, no, keep them coming, we're impressed. And we've taken inspiration, and I occasionally have some of my friends ask me like, what's the recipe for that one? That one looks really good. So a typical day is some sort of, you know, long walk, hit the gym, something early in the morning. I tend to these days be up earlier. I got up at 5 today. Not because— just because I get a couple hours before the kids get up. Then I try to block my meetings together into blocks and preferably to parts of days. So like, I really try to shove a lot of meetings to Tuesdays and Thursdays so that I have Monday, Wednesday, Friday to focus on projects. And then, you know, in 2 to 4 hour-ish blocks, I focus on one project at a time. But as you can imagine, I get pinged on Slack by my team, you know, with a question about their project or something pretty frequently throughout the day. So I get derailed fairly often. And then a couple of times a month, I end up coming in on a Saturday or a Sunday to kind of have just a focus day. Right.

Interesting. And your team, how do you make sure that they're up to the task, that they fit your standards?

So a lot of, a lot of dev agencies work on a model of having a couple of senior people and hiring lots of juniors. We kind of took the opposite approach. Pretty much everybody we've ever hired has been extremely senior. And so I'm looking more for ability to communicate than even tech skills. I can teach tech skills. So we actually have more advanced degrees in English among us than we do technical degrees. So, which really is the hard part of development is sussing out from the client what it is exactly that they want or what problem are they really seeing that we need to, you know, communicating what the solution may be is often more of a problem than the actual technical part. So, Part of growing slow is hiring slow, right? So most of the people we've hired, they've been involved in the Django community for a period of years and we've seen them grow and we've seen how they work. And so we know before we hire them kind of at what level are they at. And so we don't get surprised.

Can AI create a better Django and will AI even need languages if they can just speak to themselves one day?

So, this is something I brought up a couple of times in various conversations that I think is interesting. I don't know that we'll see another big web framework emerge except in areas of maybe a new language, right? The reason is, is, so, in the Python world, Django for the longest time and then Armin Rodenacher built Flask as a competitor and it takes a different philosophical approach. There's not really that one's better than the other, they both can do everything, it's just which one do you kind of prefer. The reason Armin went and built that was because he was running into sharp edges in Django that he didn't like, stuff that was annoying him in Django. Well, if you're not writing the code, you're not running into those. If you're just telling an LLM to go build something, you don't almost care that it's annoyed about something, or like you're not feeling it, so you don't have a motivation to go build a big thing like that. So I think we may be kind of hitting a plateau in some of the larger open source projects, and the smaller ones are all going to be just vibe-coded things that you need situationally. We'll talk more about techniques and practices and not about use this library. It's use this technique. Tell your LLM about this technique and it'll just build that piece as you need it in a project instead of using a third-party package from npm or PyPI.

Interesting. And what human skills would remain relevant if AI can code themselves and exponentially improve?

I think we're going to find that the exponentially improve is less impactful for the next few years. I'm not saying 20 years out, but we definitely can do, you know, work towards this goal until you hit it, right? But for one, somebody's got to decide what the goal is. And often, I've— experienced with various LLMs over the last couple of years. They may be heading towards a goal, but it's a pointless goal, or it's a goal that doesn't further my mission or desires in any way. It's just a goal that's picked, and it's running towards it, and it's burning lots of tokens, and it's working, but it's not really doing anything for me, right? I have neighbors that really care about their their grass at their house and they mow it and they take care of it very like golf course-like. If they were to come do that for me, I'd be like, well, thank you, but like, I don't care, right? It's just not a goal of mine to have a perfect yard. And so if an LLM picks the wrong goal, it doesn't help me that it's doing it for me. It's just wasted a few dollars worth of tokens, right?

But what if you have a digital twin and it knows exactly your goal and it works towards them? Isn't that like a simple coding equation?

Yeah. It's not a simple coding equation for sure, but if that really existed, then you would see a whole lot more meals show up on LinkedIn because I'll have a lot more time to make more elaborate lunch than I would. I don't think that we're gonna end up in a situation where we have less programmers. I think more people are going to do what, programmers do, I still think that you're going to have a senior level of people, uh, maybe we call them something different, right? Technical engineers or something. There'll be a new term, but it's the people that kind of deeply understand what's going on under the hood when the LLM can't, won't, is not quite right, is not quite doing what you want, that can step in and, and tune that. So maybe we won't end up with armies of as many notable armies of programmers, but we'll have SWAT teams or something of programmers. I remember a story years ago, there was a lady that worked for the Kansas Geological Survey. They were having a conference of some sort, and she presented a project where she had used Django and its GIS capabilities to do a bunch of mapping of all this geological stuff in Kansas. And, uh, one of my business partners attended it, and he's— and he's like, well, this is amazing. She's like, oh, but I'm not a real programmer. He's like, you built this whole thing. This was before LLMs. This was like 10 years ago. He's like, you built this whole thing, and it's interactive, and it looks like Google Maps. And like, you're a programmer. I hate to tell you, but you're a programmer, right? And, uh, you don't have a computer science degree, but you are a programmer. And I think that more people are going to have that feeling. I think we're going to end up with 10x as much software as we have today. And that means that there's going to have to be maintainers of that on some level. And we already are seeing clients come in the door with a Vibe-coded prototype of what they want, but they realize, I can't actually— this looks right, it— the screens behave correctly, but it doesn't actually talk to my Gmail, and I don't know how to do that part., right? And then we have to tell them, ooh, like, this is close, but you have all these huge security problems here. We need to fix this, we need to change this around, and now we can finish it. But that removes a lot of the real cost of software development, which is building the wrong thing and misunderstanding what the client wants. If they can vibe code their way to it's 98— it looks like software that they want, right? The forms look the way they want, it's asking the questions they want in the order that they want. How it stores its data and reaches out to other systems, they need it for it to function, but they almost don't care about how that stuff works. But they need somebody who does understand how that stuff works to put the final level, the final round of prompts on it, if you would.

And what would a 1000x engineer look like in your eyes?

I mean, we used to talk about 10x engineers a lot. And I know a few that really do get quite a bit more done than other people. I think that kind of the 1,000x, which may not be like a real scale, but like just a label, I think they already exist. I think that there's a few dozens of us out there that are doing 5 and 6 things at once with different LLMs. That are just kind of going from project to project to project to project. I don't know that that's necessarily sustainable. I think it would be— like, when you look at, like, Peter from Open Claw, focused on one thing with lots of resources, I think you can be a very 1000x engineer in that sort of way, right? I think that if you, like, in an agency sort of role, I don't know that you can. I don't, as a hobbyist, I don't think you can. But I think if you really have a maniacal dedication to one project with lots of resources and you can just focus your time, you know, all of your skills and prompts and things are all tied around that one project, you can really get some amazing efficiency. But if you're scattered across 5 side projects, I don't think any of them really— I don't think you end up being a 1000x engineer because you're doing 5 projects or 50, right?

So it's like having judgment, having a super brain as well with lots of relevant noise, so high quality, and obsession once you identify like what's the bottleneck of the problem. So ultimately, it's being a Having a great brain and being a good problem solver, that's what a 1000x engineer would be.

Right. It's always been that way. It's that now we're not as constrained by typing and time in the ways that we used to be, right? Like if you could type really fast and you didn't have to look up how that one function worked or that one aspect of that one framework worked, if you just had it all up here and you could just go, a John Carmack kind of, person, you could get a whole lot done focused on a problem. I think now it's, well, what are the problems we should be tackling right now? What can I do to streamline the LLMs to do what I want them to do? How can I put in guardrails so that I don't need to review this code, but I'm ensured that we're not creating security problems, we're not creating maintenance and engineering problems? Long-term. Those are different problems than programmers really focused on 5 years ago, right? So, like, actually next week I'm going to a conference that's all about software reliability and testing because I think that's becoming more important. It was important before, but it's way more important now. And I'm hoping that, you know, some of the smart people who are presenting there have some techniques that I can bring to my team, to bring to my communities, and say, hey, there's this way that they do some testing over here that is going to make it faster, safer, better to build with AI.

Right. Yeah, but you're pointing to an important thing about the current state of AI agents, which is that they might go in all kinds of wrong direction. That's where human taste is required. I don't think it'll take a lot of time, though, for— agents require human taste doesn't seem like that much of a technical problem. There's another problem as well, is that these enterprises are underselling the token price, right? Like it's all financed. And I think that will sort of catch up.

Yeah. So there was an article that came out recently that kind of tries— I don't know how accurate it is, but it was saying that these are— they're already profitable current models on inference within a few months of a new model coming out. Now, they're way overspending because they have so much financing in R&D and people, which is probably smart for first mover advantages. But if they were to shut off the pipeline of new models, those models are profitable at today's token prices. And pretty significantly, 60%, 90% margins. So I don't know that that— and the other thing is that long term, our incentives are aligned there. They want to make it as efficient as possible. They want you to spend more tokens. They are going to bring those prices down over time, I would imagine. But you brought up the getting to human taste. I think you're right that, you know, a couple models down the road, a model might have as good a software taste as I do today, but it won't ever have the software taste that I have a year from now because it doesn't really grow in the same way. Now, we may get to a point where models really grow on their own without human involvement.

But they're pretty good at predicting the future as well. I had the guy on my podcast that does that as a service, and the models are probably more accurate better than a human at predicting what will be trendy or what will happen a year from now.

Have they— I mean, I definitely can get an LLM to give me predictions, but like, have those panned out, right? Like, you know, have we run experiments where we say like, hey, is TikTok usage going to plateau? And it says yes. And then we wait a year and see if that really happened. Like, I don't know that we've actually—

Yes, yes, yes. With Polymarket, Kalshi, you know, like all these prediction apps. Like it's set in stone. It's, uh, we have like data of humans and we can have AIs side by side make predictions and they'll be more accurate. I mean, human judgment usually is terrible. Yeah, um, might change in the founder space of things, especially when you explore the 1 percenters. Like in, for example, if you'd ask, I don't know, Elon to predict the future, well, one, he's building it, right? So, and, or Zuckerberg, for example, their knowledge. Yeah, yeah, or Tim Cook or Steve Jobs, they were incredibly good at understanding, for one, because they built it, and two, because they constantly think about that. I personally also live in the future, but also there's a lot of hubris in human predictions, right? Like, we predict things and we're like, oh, look at me, I was right. Um, and agents don't have that, right? Agents are way more scientific and can make scientific predictions.

So also, they're, they're sort of an amalgamation of the group knowledge, right, which is always a little more accurate, right? Like, there's that experiment, if you ask an individual to guess how many jelly beans are in a jar, they're usually wildly off. But if you take the average of everyone in the room, it's usually very spot on. So that makes sense that they would be kind of good in aggregate.

And if you take the average of the best bettors, then I think that's where it gets really, really accurate. You take like the 10 best bettors, right? And right, yeah.

But I think that we're still going to have things that are Okay, in the current world, I have all these great tools and things that make my life better. And I think of the next tool that I would like. I don't know that LLMs are going to be, you know, for a while, not long term, not 50 years out, but like for maybe the next 5 years, they're not going to particularly be great at, this is something that Frank would like and need and preemptively go build it for me. Now, I might be able to construct with agents you know, like an OpenClaw that has a lot of memory about me, that sort of a personal agent that is going to get really good at Frank likes his meetings on Thursdays without me having to tell it, put my meetings on. It's going to figure out that I always schedule meetings on Thursday. It's going to notice that sort of thing and do that. But I don't know that there's going to be a general model, you know, that you're going to, you're going to use Fable-7 or something, and it's just gonna know what I want for dinner, right? Um, maybe, but I, I just don't see that as being near-term, right? I think that one of the things that these models need is that level of data to do those sorts of predictions, right? So like, for if we're going to use dinner as an example, it's going to need to know what I eat for dinner every night to figure out Frank really likes sushi on Fridays. Like, maybe I should just preorder sushi, you know? Yeah.

Or know your biology or your gut microbiome. Right. And that's very hard to track.

Yeah. Well, there's fascinating stuff that could happen with that, right? Like, there's whole classes. You know, if you would have asked me 25 years ago, would something like Oura Ring exist? I wouldn't have guessed that we would have something that small, that, you know, that ability to, to watch my biology.

Yeah.

Right. And so imagine what we have in a few years and feeding that data in. Like, it's going to get real. It's going to get real interesting, you know, over the next 5 or 10 years. I am very much looking forward to having an elderly, you know, when I'm old and retired, you know, having the robotic exoskeleton walk me around instead of being in a wheelchair and having, you know, VR experiences because I don't feel like traveling to Paris. I can just VR to Paris, you know. Um, like those— I, I don't, I don't fear getting old because I don't— I know it's not me in a recliner watching crappy TV. It's probably me, you know, watching movies I've asked an LLM to create based on characters or books or something that I like. And it's— everything is very tailored and interesting, right?

Yeah. Well, I think in the VC realm of things, what we're seeing is that, yes, VCs fund hardware and very hard problems to crack way more. We're seeing a divergence, right? We're seeing money flow to the 1%. Most money flow to 1%. And back then it was more spread out. So we're seeing more and more winner takes all. And, uh, yes, hardware and these problems that you mentioned, the exoskeleton, this is getting more attention. BS like just another SaaS app is, uh, is getting, uh, eye roll, uh, nowadays. That's what we're seeing. I also think that AIs under— will understand B2B way more because B2B is rational, it's cold. B2C is emotional and is I mean, emotions, yes, are technically— one could say it's easy to understand most humans' emotions, but they— AIs don't have these things called emotions. It's a bit harder for them to understand what will work in that, that side of the market. So it's getting interesting, this race, but it's also disturbing to see that, yeah, it's like one-percenters are taking pretty much everything home and most humans will be left out. And I also see this sort of dystopian future, right, in which the 1% keeps accelerating and the 99% gets left behind.

Yeah. I mean, I worried about that early with LLMs. What happens if— like back when we were fairly certain that VC money was subsidizing token spend or token cost. That may be a little murkier now. But I was worried, what happens if OpenAI and Anthropic 10x their prices? Right, at some point in the future. You know, what does that do? Do we all just go back to not using an LLM? Do we use it in various situations?

But they take data, right? So they wouldn't do that. The data that is fed to them is incredibly valuable. So I don't think they would ever do that.

Yeah, I don't think so either. But I was worried about that early. And now there's so many good open weight models that I'm not worried about it. I have not been using a Frontier paid model for the last couple of months. I've been using open-weight models that I could easily host here in the office if I really needed to.

And if it's the end of the world and there's a cutoff, right, that would be so bad because we use ChatGPT so much nowadays and it adds so much value. And that, that's the ultimate NPS test, right? It's like, yeah, if you cut off your software to someone, how painful would it be, right? And Yeah, LLMs have really like changed how humanity operates, right? We not— intelligence is now commodified. And me, it's also scary because it's always been my thesis, right? I try to be the smartest in the room, even if that's not true, right? I always try to do more homework, to be the smartest in every room I walk in. And that's not true anymore. AI is better now. What do I have left? I'm a human. I have agency. AIs don't, but they're about to get this, right? I'm also realistic that I could be replaced by an agent, which leaves us with one thesis, like, is merging with it the ideal path? I think that, well, so far that's—

I don't want to be first. I've experienced enough beta software in my life. I don't want to be first. But one of the things that, like, you'd mentioned taste earlier, and I do think that taste, for whatever you personally define that as, is going to become more and more important. And curation of this output of LLMs is going to become more important. I think we're going to go back to a world where I follow a few hundred people that I respect their taste, and they feed me information that they've found randomly on the internet for whatever reason. There was a— not to bring it back to cooking again, but I was watching a documentary about some chefs, and they were talking about this one particular chef that had told all of his friends all of his secrets and opened up molecular gastronomy in Spain. All of those people went on to find— found really great restaurants, and it has spread worldwide. They were all sharing their techniques and their secrets, but they all ended up with very successful restaurants independent of that, those secrets, right? And so I think we may evolve a bit into even a business world where it's not so much, here's my secret software that you pay me for. It's more, I like the way Charles does business, and he can handle this problem for me, and I'm just not interested in dealing with the edge cases of that problem. And so I'm going to— my AI is going to interface with your system to do some aspect of it because I just don't want to have to worry about, I don't know, the tax implications or the shipping logistics aspects of this or something. You know, I think we're still going to have those B2B specializations. Because like, I don't, I don't want to make sure my package gets to where it's going. I don't want trucks and drones and the headache of trucks and drones, right? Unless I'm Amazon, right? At a certain scale, that stuff matters. But for your average business, it's like, I want somebody else to do these pieces and I want to focus on my core competency piece. And so I think that'll still happen. There may be less players, there may be 5x as many players. Because it's so easy for me to switch between players. You know, hey, LLM, stop using Charles, start using Steve. Okay, boss, 5 minutes later, you're switched. Like, you know, you're going to have to have good customer service and things, you know, even if it's totally automated, it's going to have to be amazingly good service, right?

Yeah, the front end of it for sure, that won't change. But I can definitely see most of us becoming APIs eventually, just from a system perspective. Frank, thanks for being on the podcast. Where can people find out more about you?

If you search for Repsys, if you search for me, that's pretty easy to find me on Google. I'm on LinkedIn, Twitter, Mastodon, Bluesky, Threads. I'm pretty much everywhere.

There you go. So that was Frank and Charles for another