Niklas: Hi and huge welcome to you my lovely listeners. So glad you're here. Today you are joining me for a chat with Alejandro. Alejandro is building Bento and has a lot of experience on developing app. Alejandro, so nice to have you.

Alejandro Zielinsky: Yeah, nice to be here.

Niklas: You already told me quickly, you have quite an intercultural background, family-wise. Maybe we start there. How did you end up?

Alejandro Zielinsky: Yeah. Yeah, sure. So yeah, my dad is from ⁓ Buenos Aires, from Argentina. That's why my name is Alejandro. his ⁓ grandparents were from Poland, so that's why I have a Polish last name. And ⁓ my mother is Austrian, so I speak German as well.

Niklas: You also speak Spanish or Portuguese? Which one is it? Yeah. All right. That's quite nice. And ⁓ these days, you build apps. I think you have a lot of experience in iOS apps. How did that journey go?

Alejandro Zielinsky: Yeah, I speak Spanish, but not fluently, yeah. Yeah, so ⁓ I went to school to learn how to make video games. I'm more of a right brain person. I was gonna go into music and then my piano teacher took me out for coffee and said, you know, ⁓ I don't want you to be living on the streets, so ⁓ like okay maybe I'm not gonna go into music. ⁓ so I was looking at other things I can do and I just wanted to, you know, be creative and I thought what is more creative than building video games? get to create worlds, I get to still compose music. So I went to school ⁓ to do that and then ⁓ I was only in the game industry for a few years. ⁓ I worked at Ubisoft for a little bit. I worked in Montreal and then I realized I just wanna focus on my own games and I don't really want to do it as a career. So I switched to iOS ⁓ and I've been working for different companies, mostly ⁓ dev shops and agencies building out their iOS apps for different clients for like the past ten years.

Niklas: I've always, I've started programming quite young. I've never really programmed games. So this is a very interesting topic to me because while I like to play video games, I've never really gotten into building them and then just realized later how big of a market it is. What do you think you can learn if you start with games to do the opposite side and then build apps? What can you learn from games for building apps?

Alejandro Zielinsky: Well, I find that like like in terms of a subset of programming, building games is quite difficult. It's quite involved. you know, there's physics and math involved and you know, writing code for the GPU. And so when you transition from games to regular apps or regular software it feels really easy because now, ⁓ and I just have to render some UI components and transform some data. That's a piece of cake compared to, you know, writing a pathfinding algorithm and, you know, writing a shader to reflect the sun off the ocean.

Niklas: It's really, I think it's getting more interesting now. soul, I think also game development was the idea of developing world models. Because when I look at what's happening in AI, maybe you have an idea to that as well before we go into your apps. But when, when I started developing for machine learning, for example, a while back, and I know that you do a lot of that and for gaming as well. ⁓ I. stumbled over something called OpenAI Gym, which is kind of an environment that your agents can work in. And I always thought, okay, yeah, if you want to build like real agents, it would make sense to describe the physical environment, like in a virtual setting with constraints, and then have the agents learn during self play or something like that. I would, that I would. If I think about very good agents, would think this could go in a similar way. How do you think about that?

Alejandro Zielinsky: ⁓ are you saying like like three D world representations of of your agents? I 'cause I've seen that. I don't I don't know if that's what you're talking about.

Niklas: Yeah, I mean, it obviously doesn't have to be visualized ⁓ in any sense, right? But you need some kind of constrained environment that gets as close as possible to the real world. I mean, if we could do something like that and then have agents learn in them with different algorithms, this would probably be a path to getting kind of very well-learned agents. That's always what I thought when I looked at this, because what I did back in the day was I looked at container yards, for example, and in ports. And you can stack containers in different ways. And what they do is they have complex heuristics to optimize something like that. What I looked at back in the day, and that was 2020 or 21, was, I just described the environment. And I say I have a set of places where something can be. These kind of things, things can go in certain places. are physical limitations, like containers have to be stacked on each other, these kind of things. And then I just have a neural net ⁓ and a trained neural net to do all the stacking. And I don't care anymore. I remove the compute layer.

Alejandro Zielinsky: Mm.

Niklas: And that was always what I thought when I obviously have no idea and I've looked into transformers, but I always thought if you want to have really good agents, you would have to create some kind of environment like this and use this as a training environment for your agents.

Alejandro Zielinsky: Yeah. Yeah, for sure. yeah, it's just the beginning of this whole new agentic world, so ⁓ it's exciting to see where it's all gonna head and yeah, ⁓ for sure that's all gonna be relevant to repo robotics too, where they have to understand, you know, the real world and stuff like that. so yeah, I get a little bit of whiplash thinking about ⁓ how fast things are moving and ⁓ you know, I still I need to take time to to expand my like agentic workflows as well because I still just, you know, have a couple Claude ⁓ code instances working and talking to each other, but you know, there's people like spawning swarms and and taking all this to another level that ⁓ I haven't had too much time to look into.

Niklas: Yeah, well, feel like I've been using Cloud Code for a while, but I feel the latest big advancements really are on the more like co-work side. So when I want to create an Excel or PowerPoint or something like that, these kinds of things have gotten ridiculously good over the last year. Like if I use Cloud Co-work, for example, the output quality is just amazing. And I think that is where I felt the biggest advancements were lately.

Alejandro Zielinsky: Mm. Yeah, I haven't even played around with ⁓ cloud co-work, so I I need to check that out.

Niklas: probably from any other of the three big players you will see the same. So whether it's Gemini or Codex or ⁓ Claude. ⁓ Yeah, looking over, think that is, what I also feel is my kind of subscriptions are spiraling out of track. So I'm really losing track over time. And that was why I got really interested into. bento because I also thought about building something like that back in 2022. never moved. I got a few initial users. completely scrapped the project, but I think the idea is still as valid today as it was back when I looked at it. So maybe you can tell us a bit why you started building that and especially how modern AI can help.

Alejandro Zielinsky: Yeah, sure. Sure. Yeah, so ⁓ in the end of ⁓ twenty twenty four, beginning of twenty twenty five, I was working for a crypto company that was trying to do ⁓ Uber or Lyft ⁓ like ride sharing platform on the Solana blockchain. ⁓ it was called Teleport. And in January twenty twenty five they shut down. ⁓ and so I was like, okay, I can go and look for another development job or I can really just take a little bit of a career break and ⁓ try to build an idea. And the idea that was floating in my mind was, you know, I'm kind of financially lazy. I haven't used a budgeting app. Like I was looking at budgeting apps and stuff like that, and I I saw that they're all like built by very left brain people. They're complex. ⁓ they take, you know, some time to learn. They're very rigid systems. And I didn't love any of the solutions. And I also noticed that none of them were leveraging AI. Like none of them had ⁓ like LLMs integrated and if they did the one or two that I found would hallucinate data, they were unreliable and they were, you know, you know n not implemented very well. So I had this idea to build my own personal finance app that focused on modularity. So I I can track what I want. I wanted to create dashboards and you know, I was the idea ⁓ crystallized as like a widget system, like iOS native widgets. I just wanted to, you know, track coffee, have a coffee widget, track groceries, have a grocery widget, have like a merchant comparison widget and compare Uber versus Lyft spend. And I wanted an AI to integrate with that so I can get my agent to organize my transactions. And I wanted to, you know, I wanted to be able to say, last month I was in Mexico, tag all my transactions as Mexico, and now I have a Mexico widget tracking that spend. And so I wanted something, you know, for like a Gen Z or young millennial audience that don't really want. use these complex like YNAB or Monarch, these rigid systems and just want easy answer easy answers and you know they're used to AI giving them quick answers. So so I went to Thailand, I went to Southeast Asia and I started building this ⁓ project. And ⁓ I'm very like what holds me back is l ⁓ is my perfectionism. So I I'm very design oriented. So I just launched it two weeks ago. ⁓ so it took me a year and a half to build. ⁓ but I'm I'm really happy with what I made. Like it's I the idea was it it should be like an you know, if Johnny Ive like the guy who kind of designed the iPhone and early Apple products. Like if Apple made the personal finance app, that's what I was aiming for. Yeah, and it's called Bento because it's modular. Like it's you can transform all your widgets, and it transforms the data. ⁓ yeah, another inspiration was a mix panel, because you can like answer any product question you have and visualize it in different ways. And so yeah, Bento is the result of all that.

Niklas: How is it going now? You have launched fairly recently. Do you have first users, customers? How do you monetize?

Alejandro Zielinsky: ⁓ so yeah, I'll answer that. The the funny thing is I launched three hours before I launched, OpenAI announced personal finance integration with plaid, which I thought was like, you know, the gods kind of kicking me in the nuts a bit. but you know it's it seems like people don't really, you know, it it's not that big of a deal. But I just thought like what are the chances OpenAI kind of ⁓ Sherlocks me in my launch day. But that's why I'm happy that it's not just the AI. It's it's also this modular system. ⁓ and it's going all right. I my problem is I I I didn't I I don't do the marketing side until I launch and so now I'm ⁓ trying to hire someone to handle marketing. I'm trying to push out TikTok reels and and focus on that. So I'm not focusing on the product right now. It's it's already too polished of a V one. ⁓ So now I'm pausing product work and just trying to push it out there and get users. And it's a s it's just the beginning. Like I just launched two weeks ago. So no one knows about the app yet.

Niklas: Yeah, I I think I've launched a few apps over time and ⁓ early stage growth is just really hard. It's the same, think, as you said, nobody knows until you sometimes you can hit some kind of priority, but unless you already have an audience, it's really, really hard. Still asking, ⁓ do you see any adoption already outside of like that you can save? generate from TikTok and if you have tried different marketing channels, what's your feeling? Which marketing channels work best for something like

Alejandro Zielinsky: Yeah. So, I remember the previous apps I launched, Apple would give you a little bit of a boost in the app store. I noticed they don't do that anymore, so I think that's because of how oversaturated the app store is and all the vibe coders pumping out apps every day. so I I recall always getting a little boost like a week after I launched. and in terms of marketing I haven't I haven't ⁓ done any meaningful work on p on the marketing side on any of these platforms. I'm kind of warming up my accounts, starting to post Instagram, TikTok accounts. So I don't have a gauge of what's the best ⁓ one at the moment. But from my research, I'm expecting TikTok to to perform the best.

Niklas: Yeah, that's what I've heard from everybody. So it's also true if I look at different channels, like personally, also from the people I know, also for personal account, it seems to be the one if you put in the work where you have the best chances, like for obviously consumer audience, right? So if you look at a business to business audience, people say LinkedIn is the clear winner. ⁓ X doesn't really seem to win. Also, I use it a lot, but... Yeah, that's really interesting. When you said personal finance apps are often too rigid, what is it exactly like if you pinpoint it to a few sentences that makes a difference with your app for yourself?

Alejandro Zielinsky: Yeah. So every personal finance app out there ⁓ will prescribe you dashboards. So they'll have a dashboard and show you your net worth, they'll s show you your top categories, and let you set up budgets. and they all do that. But if you have specific ⁓ questions or you wanna set up a s a dashboard yourself, I I haven't seen any app that lets you compose your own dashboards and answer your own questions about your finances. So I thought what Bento offers that none of these other apps do is dedicated dashboards for whatever event or project in your life you want to track. ⁓ so Bento allows you to, you know, if you have a wedding coming up, ⁓ create a wedding dedicated wedding dashboard connected to whatever accounts you wanna connect it to. ⁓ so your dashboard is sp ⁓ scoped to whatever bank accounts you want it to be scoped to. And then you can Tag your transactions that are related to weddings. You can ⁓ track categories related to your wedding, ⁓ track merchants related to it, ⁓ and have that all contained in a dashboard, and then eventually we'll let you like you know, export to CSV or save it or archive it, whatever, and you can always reference these moments in your life and how much they cost and whatever. ⁓ it's also really good for for trips. You know, you you go on a Euro trip, Mexico trip, whatever. ⁓ And you have a dedicated dashboard for that. You track your Airbnb, you track Ubers, you track flights, and it breaks it down to there's so many different options to ⁓ view your data. That was the mixed panel inspiration, too. If you want to look at it in bar charts, line graphs, pie charts, ⁓ if you want some more metrics about it, like the percentage of your spend, like we let you have complete control of how you view your data, and then ⁓ Yeah, just no other app lets you have that level of flexibility.

Niklas: You also integrated an AI assistant called Benji. I think that's also a strong differentiator against other apps. How does it work?

Alejandro Zielinsky: Yeah, so Benny ⁓ is hooked up to ⁓ GBT five point four mini model. and ⁓ it has a lot of like deterministic like server scaffolding s so it doesn't ⁓ hallucinate. But it basically yeah, it has these tools. and it has ⁓ It has these intents. So every single request that you make to Benny, he'll detect whether it's like a planning intent, whether you're trying to plan something, whether you're trying to create something, like a budget or a a category or tag, ⁓ whether you want insights or whether you want to organize. And then all those paths have a different prompt and keep it like really good at those specific ⁓ like intents and all those intents have specific tools. And so ⁓ that's why when you use Benny there's like a different icon and those are like the different intents. ⁓ so yeah, so Benny is most useful to like sit like understand your spending context. It can ⁓ indirectly query the database. It has an understanding of the schema and it can only read. It can't update or touch any data. So it can ⁓ Read data, synthesize it, and then whatever you need to suggest budgets, to apply tags, whatever. And it won't ever make a change for you. It'll just ⁓ suggest changes that you have to preview, and then you'll have like a UI in the iOS app where you can like deselect certain rows if it got something wrong, and then update all. So every single ⁓ update ⁓ goes through a user review. And ⁓ it can like it's really powerful, it can like synthesize tables for you and just you can ask it like compare my top category spend over the last three months and show me how many transactions in each category and it'll create a whole table for you.

Niklas: Yeah. When you look at these things, you have built a ton of features. How did you decide what not to build? Because I think that is also very hard to be very precise on what you're building and what you're not.

Alejandro Zielinsky: Yeah. So I focused around I think it it's too feature rich for a V1, first of all. Like I have feature parity with all of my com competitors and then I have things that they don't have. And you know, there's no point building out so many features if you don't want users. Like I I don't have any that many users yet. so I definitely could have launched a long time ago with less features. ⁓ but Features I did decide to build were based around e like just ease of use. I want I want tracking ⁓ to be effortless. So right now I don't have manual account adding. Like you have to connect through a th like the trusted third-party plaid to connect your bank accounts. ⁓ because I didn't want users to have to like input every single transaction. I want it to be seamless. You connect your ⁓ bank accounts and ⁓ everything is handled and categorized for you. So all all all the decisions, future decisions were based on how to make a tracking your money experience as seamless as possible.

Niklas: this. I think it's very interesting because I remember when I worked on an app like that, it was also very rigid, but also obviously had a dashboard. I think it's a classical one, right? I was just interested in recurring subscriptions that you have because I felt like this subscription economy is just evolving a lot. And I'm just, I would say I've a lot of very small subscriptions where I pay just a few euros and sometimes I would just forget to cancel them. And it was annoying to me because if I just took a look closer, I would maybe cancel them, but that's still quite cumbersome. How did you solve the topic of connecting to a bank account or did your app connects to bank account, right?

Alejandro Zielinsky: Yeah. ⁓ so it connects like every other w ⁓ one of these personal finance apps connects, which is a third party called Plaid. It's kind of like the leader in the industry and the most trusted ⁓ third party. And so, yeah, I've partnered with them and only request read only access, so I I can only read transactions. ⁓ I can't move or you know do anything else with with your money and I also get redacted data so I c I don't ⁓ get account numbers or anything like that. So it's pretty like that was the thing I feared is just like maybe people don't have much trust in connecting their bank accounts yet, which is actually g The good part of OpenAI kind of announcing a partnership with Plaid too, because it's gonna build more trust in in these third parties. But yeah, my promise like I'm very privacy oriented, so I also ⁓ you know never sell your data. And when you delete your account in Banto, I delete every single record that you have with us.

Niklas: Yeah, it would also be a question that I have, right? Because a lot of applications, especially for like smaller companies, I would also have privacy concerns first. When the company is a little larger, it's, I think, easier to mitigate and people will directly believe that you take care of the data. That's, I think, also a big advantage of Apple. It's a disadvantage of a small solo developer.

Alejandro Zielinsky: Yeah, exactly. Yeah.

Niklas: What I'm also interested in is, I think we have seen a lot of apps that manage to build gaming features into their tools, like Duolingo, for example. Did you also incorporate something like gamification into your app? ⁓

Alejandro Zielinsky: I think my gaming background definitely influenced the design. Like it feels engaging and fun to use. ⁓ I don't have anything like explicitly gamified in it. Like I don't have streaks or anything like that, but I have thought about it. So in the future I'm sure I'll I'll think of ⁓ gamify features. ⁓ like the this is my first non game app that I've developed and released. Every other app has been a game. So it's definitely in my nature to try to gamify.

Niklas: Yeah, I think that ⁓ it's mostly helpful, right? Because an app that is, and maybe gamified is also not the perfect word for it, but an app that is fun to use, it's just, especially as a consumer app, you just like to use it more, right? And we use enough boring app or apps that are actually not fun to use. So building something that is engaging, that is nice to use, definitely wins. If you look at it, you have built, I think, ⁓ Games, have built different applications now, Fintech app, what was most fun to develop?

Alejandro Zielinsky: w I think this one was the most fun to develop. ⁓ just because it was a product that I really wanted for it to exist and it was just really gratifying to use it throughout ⁓ development. And it's also saved me money. Like Bento it makes your spending so transparent that it really shows me my habits and shows me, ⁓ I'm spending way too much on coffee and and these things. And then I, you know, I cut back and it's it actually changed my habits as well. but there's there's so many UX and design challenges with this app ⁓ that it was really engaging ⁓ to build. It was really fun.

Niklas: Yeah, I find especially apps that utilize LLMs. Usually for me, it's really hard. As soon as you start tool calling and you have to deal with different scenarios, like people will not kind of do what you expect. It's just incredibly hard to get consistency into these applications. How did you solve that?

Alejandro Zielinsky: Yeah. I mean Benny, if you use Benny he'll it'll say beta. ⁓ and that's because he's still not perfect, he'll still make some mistakes and I'm constantly ⁓ you know, fixing his brain. but it yeah, ⁓ there's been so many refactors and rewrites of of him because of just you know, it's a pretty like new ⁓ sp space and domain to just build out these agents and hook them up to tools and stuff like that. So I kept learning better ways of doing things, trying to optimize like input tokens. you know there's times where he was just way too expensive. so I kept trying to optimize him. ⁓ and then when you switch a model it it can just you know switching from four point one five point four I had to kind of rewrite his brain because of the like model specific considerations. Like it just behaves so differently and and decides to use tools very differently. So a lot of iteration. Yeah.

Niklas: Yeah, I also find that. So different models behave differently. It's also sometimes you have the feeling something works and then you realize it doesn't work. is really challenging because I think as developers, we are so used to having deterministic systems and LLMs are just not this deterministic. And also your input is not deterministic. It varies a bit and it is so incredibly challenging.

Alejandro Zielinsky: Yeah.

Niklas: to get that right. mean, that's probably also a huge opportunity and maybe some people have figured it out really well, but I notice this every time I use LLMs, like getting consistent output and performance is really, really hard.

Alejandro Zielinsky: Yeah, totally. and ⁓ that's why like running evals is really important. Like I have an eval framework with all the most common prompts and all the most common ⁓ queries that I'm trying to not have regressions on. And so every time I make meaningful changes to the prompt I run it through the evals and make sure that ⁓ they pass and that I didn't cause any major regressions.

Niklas: I think within the company you run, you also have a few games left like Letterboxx or PoliWeta and you talked about selling them online. What's up next? Do you plan to do that? Do you want to focus on Bento? Do you want to build more finance app? What are you planning to do for the future?

Alejandro Zielinsky: So I'm such a product builder. I like I love to build products. I've worked so hard on my games, Letterboxd and Polyvita. And then I launch them and I just make the classic mistake of of ⁓ just waiting for the market to find them and I I don't put enough marketing effort into them. So it saddens me a bit, ⁓ because I'm just so busy, I have my hands full with so many things that I don't have the time to give it the marketing attention it deserves. So that's why I was open to selling them. But I do want to, you know, improve my marketing skills and start I'm I'm starting to look at, you know, a AI tools and agents to help me with that as well to kinda automate some of the posting and like asset creation, marketing asset creation. So so I I still wanna give those apps ⁓ another chance and do a a marketing push. ⁓ because Polyvita is like a great game. It's it's ⁓ inspired by Monument Valley and I think there's a huge market for it. I think people would love that game and and the people that have played it do love it. ⁓ I just ⁓ my my issue is is reach. I'm very bad at distributing. I'm just I'm not a good seller. I'm a good builder. I I don't like to bother people, so inherently I'm not a good sell seller. ⁓ I just ⁓ my my issue is is reach. I'm very bad at distributing. I'm just I'm not a good seller. I'm a good builder. I I don't like to bother people, so inherently I'm not a good sell seller.

Niklas: Yeah, I'm not, so I'm like technical by nature. And I think selling, if I would say selling in a person to person sense came relatively easy to me because you just need to find the pain and articulate the value. Like if you look at real go-to-market for apps, it's a very different story. It's hard. It's a real learning. I think you can learn it. as you can also learn sales, which might be different from time to time and you can probably perform more. But to me also like marketing and sales, it's really hard work. It's not easy. Like for me, writing on X, for example, that's easy. Like that's fun to me, but doing a podcast is also easy. like creating marketing content, as you would say, this is hard work. Like I can also second that, but I... also understood more and more over time and watching people, that it's a real differentiator. Because if you fail, I mean, ideally you make a product that is so great that people tell their friends and they all come in, but you still need to kind of accelerate this, right? So if your starting group is just so small, it's very, very hard to grow it. And especially in consumer, it's really hard. So I believe, yeah, I... I can fully follow that.

Alejandro Zielinsky: Yeah. ⁓ that being said, there's a lot of ⁓ content on X and people telling you their strategies and ⁓ I'm I'm starting to follow some of those now, so I'm excited to see where that leads me.

Niklas: And, but. Yeah, it's absolutely exciting. And I think the nicest thing that you can have is actually you listen to somebody or you had that idea, you build it and people start using it and they are happy using it and it gets adoption and it gets big. That's actually really nice to see like the same part. It's just very different from the first part from the ideation and creation. If you look at Or if you had advice for people that have an app idea but keep second guessing themselves, what should they do? How do they start?

Alejandro Zielinsky: So I would say if you've never built a software product or an app before, ⁓ it's easy to kind of dream up a big idea with a large scope. And I would say as your first app, the idea matters less. If it excites you, great. But keep it small and just f ⁓ go through building it and shipping it and see how that feels. And then you can make your ideas a bit bigger. ⁓ don't do what I did and and s take a year off work and spend a year and a half building an app. yeah, w with these AI tools now, you can you can build quite quickly. but definitely still like i if you are just a vibe coder, I would say, and you're not a software engineer by trade, definitely Try to get an understanding of your systems and still like study, you know, security and databases and you should know how your app works. so yeah, I think some of these vibe coders still ⁓ are a bit naive and leak their their ⁓ secrets and tokens on GitHub. So yeah, be careful with that.

Niklas: Yeah, I have a lot of these discussions on the podcast already. And I honestly think like really good senior developers with a good architecture, understanding and design, they can just easily 10X their productivity. Even if they already were like the famous 10X developer, then they are just 100X. Like the, it's just an amplification. If you are terrible, you will just amplify your being terrible. the...

Alejandro Zielinsky: Yeah.

Niklas: You have to have this, you have to know what you want and you have to have an understanding of to judge the quality of the output. If you don't have that, ⁓ you will not get great results that I truly believe in.

Alejandro Zielinsky: Yeah. Yeah, like the way I started Bento was ⁓ I didn't actually use AI for the first few months. I just built it by hand like an artisanal ⁓ coder and then once I did integrate agents, I had a very solid foundation for them to build on top of. ⁓ you know, my app architecture was already defined. So I think that's a really good way of building too is just setting the foundations up and then having agents ⁓ leverage that ⁓ solid foundation.

Niklas: Maybe as a last point, so if people now got interested in Torbento, where can they find it?

Alejandro Zielinsky: Yeah, so you can go to bentoapp.io. it's currently region locked to just US and Canada. ⁓ and it's only on iPhone. So it's on the App Store as Bento, Colon, budget and money tracker. yep.

Niklas: Thank you, Alejandro, for being on the podcast. It was a really fun talk. And to you, my lovely listeners, see you next time.

Alejandro Zielinsky: Thanks, Nicholas. See ya.