My Startup App Was Dead. Then I Built an AI for It. | Devlog #1

Added: 2026-05-12

[00:00:00]So, Finance Guy is actually becoming real. And before I get into that, hi, hello. It's been a month since I last uploaded. I kind of got busy with finals and then I went to my uncle's wedding, the whole shebang. But it's fine.

[00:00:10]Anyway, we're here now. Anyway, Finance Guy. So, some of you might remember that I promised publicly on this channel with my whole chest, "Next week, this either turns into something real or I find out why most startups die right here." Pro- probably both. It is May. And okay, fine. The last time I actually touched the codebase before this week was January. So, there was a bit of a small gap, a brief intermission, dare I say. And so, yeah, for the past 4 months, the app kind of just sat there. I was not actively working on it. But here's the thing. I opened it 3 days ago and went absolutely feral. I mean, I looked at the codebase all the way back on April 28th after I came back from the marriage, and it was a basic finance tracker. And I just looked at it now, and it has a full machine learning pipeline and on-device inference with ONNX runtime, ML governance infrastructure, anonymized training views, and a full guided onboarding system. All in 3 days. I don't fully know how that happened.

[00:01:04]Well, I do know how it happened, but I'm going to try to explain it as best I could. So, April 28th, the comeback begins. I decided the app needed a full overhaul, not a little cleanup, everything. The thing is, right now, the whole project looked like it was vibe coded, right? I mean, it looked stupid, it looked generic, it looked ugly in my opinion. And so, I was like, "Hey, let's just rebuild everything first." But this time, it was different because I finally had a clear identity for what Finance Guy actually is. Finance Guy is going to be the little coach in your pocket. It's not just a tracker. The whole thing is built around the idea that managing money should feel like leveling up, not doing homework. So, when I rebuilt the UI, I wasn't just making things look nicer, I was building towards that identity. A lot of gamified elements, as you can see, which is kind of standard.

[00:01:50]I might remove them afterwards cuz I'm not really feeling the gimmick as much.

[00:01:53]But the idea at the end of the day is you enter in something onto the app, and the app essentially tells you whether you're being dumb or whether you're not being dumb. And of course, along with that, on April 28th, I decided, "Hey, I'm going to ship the actual UI, and then I'm also going to ship all of the Superbase migration with it." The bones of the app were finally getting real.

[00:02:13]January me may have planted the seed, but I wanted that seed watered in 72 hours. So, it's finally April 29th, the second day. This is the day I built Blue Eyes. If you don't know what the Blue Eyes Predictor is, it's essentially my machine learning model where the idea is before you make a purchase, the app checks in with you. Not in an annoying way, just like, "Hey, you're about to spend money. Let's think about this for more than 2 seconds." And while you're thinking, it would essentially run a prediction in the background. It scores how impulsive this purchase is likely to be, it explains why it flagged it, it recommends what you should do, and then it records what you actually decided so it can get better over time. And the full big picture of this This is the MVP being shipped. So, this is a very heuristic model right now, where essentially you put in the data and then it spits out stuff. Eventually, I want it to get to a point where the machine learning model has trained so much on your data where before you make an impulse spend, it's going to notify you.

[00:03:07]The app is going to be like, "Oh, look, you tend to get coffees on a Monday morning. Let's not do that." So, I built the pre-spend check-in flow and the core prediction service. The app can now create a spending intent, which basically means I don't add it to the actual spending, it's just my intention to spend. And then after that, it scores its impulse risk, yada yada. I had to create a bunch of new Superbase tables.

[00:03:27]Again, this is a very good learning moment. When you're coding these big apps, it is very important to get your systems down first. If you have no system in place beforehand, no vision, you're going to be scrambling like me, making new Superbase tables and connecting everything over and over again. And then, I turned the Blue Eyes into a real ML system. Because here's the thing, it's very easy to call something AI-powered nowadays and have it be a bunch of if statements wearing a trench coat. I didn't want that. I wanted a real model that gets better over time, that's transparent about how it works. Again, we're Again, we're following the FAIT principles, if you don't know what those are in machine learning. And that doesn't just collapse when the model isn't available. So, I built the governance layer, right? So, we have anonymized training views. The data the model learns from can't be traced back to individual users. We have the opt-in research consent, which basically means you choose whether your anonymized data contributes to the model improvement. We have a model info screen in the app that actually explains what Blue Eyes is and how it works. All of these are present within the settings.

[00:04:29]You can click the link so you can manage your consent and see what the model knows. This is something that I feel like you can't learn from AI. This is proper machine learning principles, and this is kind of what separates a gimmick from a real product feature. The governance, the transparency, the fact that it's designed like something you'd actually trust. I don't want it to be a black box you send data into. I want you to understand, "Hey, this is what the data is contributing to." Then, I added the budget recommendation engine. Blue Eyes looks at your recent spending, your projected income, your active quests, your impulse-prone categories, and then it generates budget caps with a one-tap apply. So, it's not just telling you things, it's actually doing something about it, too. All right. So, final day, April 30th, or yesterday from when I'm recording this. I was kind of relaxed this day. I had an interview. I wanted to play football with friends. So, we didn't really do that much. You know, I cleaned up the add transaction UX, which had been quietly bothering me. See, the whole point of the app was that, "Oh, we're not going to give you automatic sync because automatic sync would kind of ruin the idea of you being accountable with your own spending, right?" I wanted it so that you open the app and within 30 seconds, kind of like MyFitnessPal, you've already logged in your spending. Uh and then I built the onboarding. So, we had the proper guided onboarding, you know, there's a tour that walks you through the main loop, you know, we've got the treasury, income, expenses, limits, quests, the Blue Eyes. Uh it stores the completion state, it's integrated into the app layout, and then it's replayable from settings if you ever need a refresher.

[00:05:58]Because here's the thing about features.

[00:05:59]You can build the most technically impressive system in the world, and if someone opens that app and doesn't know what to do in the first 30 second, none of it really matters. So, onboarding is really what makes everything else land. It's the thing that turns the codebase into an actual product. There's a couple of things that I'm in the middle of right now. Firstly, I might rework the machine learning pipeline simply because some of the stuff here is kind of vibe coded. The ONNX I personally don't really understand how it works yet. So, I might rework it a bit, get it to work with something I'm more familiar with, and then we'll see. But I'm not going to talk about those today because they're not committed yet, and I've learned my lesson about announcing things before they exist. So, that's where we are. 4 months of nothing, and then 72 hours of losing my mind, and now Finance Guy has a machine learning pipeline, and on-device inference, ML governance, AI budget recommendations, gamification infrastructure, which is pretty gimmicky, and guided onboarding. Is this shipped? Not yet. Is it more real than it has ever been in its entire existence? Genuinely, yes. I'll upload again hopefully by next week, talking more about this app because I'm really focusing on it right now. Uh but yeah, let's see how that goes.