# Could One Person and AI Build the Whole Thing? > Rainfall cleared the App Store, but the launch was never the thing I was testing. The real question was whether one person and AI could build the whole stack. Published: 2026-06-30 · Updated: 2026-06-30 · Topics: AI & Machine Learning, Product Strategy & Development, Engineering & Development, Startup, Beauty & Wellness · Author: Alan Wizemann Rainfall cleared the App Store today, which is a nice sentence to be able to write, but the approval was never really the thing I was testing. The launch is the visible part – the part you can screenshot. The question sitting underneath it, the one I actually built the whole product to answer, was simpler and more uncomfortable: could one person, working with the AI tools we have right now, build an entire product across every layer, the code, the backend, the writing, the voices, the music, the brand, and have the result be good rather than merely possible. I did not want a demo that proved a model could write some Swift. I wanted to find out whether the whole stack could come from one person and a set of models, and whether anyone on the other end would be able to tell. And I gave myself almost no room to hide, because Rainfall went from an idea to a finished, launched product, on the App Store, in twenty-one days. The idea itself came from a small, ordinary failure. I was in the backyard during a rainstorm trying to meditate, and the app on my phone wanted me to choose a category and a duration before it would do anything at all, which is precisely backwards, because the moments you most need help are the moments you are least able to navigate a menu. So Rainfall works the other way around. It reads the shape of your day on the device (a short night, a packed calendar, a restless heart), infers a kind of weather for your mind, and offers a single practice it thinks fits, already loaded, instead of a library to dread. That is the product. This piece is about how it got built, because the how turned out to be the more interesting story. Start with the code, since that is the one place most people already assume AI belongs, the layer it is allowed to touch, even if they will tell you in the same breath that it still gets plenty wrong. Rainfall is a native Swift app on the device paired with a backend on Cloudflare (Workers, a relational store, edge-cached catalog snapshots, object storage for the audio, and durable workflows that survive a server being evicted mid-render). The model wrote a great deal of that, but the part that made it work was not the code generation, it was giving the model a memory that lived inside the repository itself, through a tool I built for exactly this problem and named Memophant. Every session could read the decisions that came before it, so I was not starting from a blank chalkboard each morning, re-explaining the architecture for the hundredth time. The decisions stayed, the model read them before it wrote a line, and the thing that is usually the worst part of working with AI, the amnesia, mostly went away. That is the difference between an assistant that is fast and one that is actually building alongside you. The hardest call was privacy, and it is the one I am proudest of, because it is structural rather than promised. The parts of the app that read your sleep and your calendar and your heart, and infer how your day is landing, physically cannot reach the network; the module boundaries forbid it, and the backend has no endpoint that will even accept a health signal or a mood. Nothing is uploaded, there is no account to make, and the server only ever sees an anonymous request for an audio file that carries no hint of why it was chosen. I did not want the kindness of the product to depend on trusting me; I wanted it to be true in the wiring. Then the content, which is where most people would expect a wellness app built this way to fall apart. The meditations are written by a model, but every script passes through a blocking safety gate before a single second of audio is ever rendered, because "move fast" is not a phrase you get to use casually when someone vulnerable is listening with their eyes closed. The voices were the real surprise. Stock AI voices are trained for voice-over, all forward-leaning confidence and even spacing and a faint smile in the delivery, and meditation needs the exact opposite, the trailing cadence and the generous silences of someone who has held space for people before. So I did not use stock voices at all. I designed bespoke ones from written character descriptions (a near-whisper "just above a whisper at 6am," a low male voice "at the end of a long session"), auditioned them, rejected the ones that sounded like they were selling something, and locked the handful that sounded like a calm friend in your ear. The soundscapes are generated the same way, a small owned library of beds rendered once and held, never conjured fresh on every play. None of this was a token gesture, either: what actually shipped is eight bespoke voices, 180 guided sessions across the catalog, and five original soundscapes, every second of it generated, synthetic, and owned outright rather than licensed from a stock library. I want to be honest about what that experience actually was, because the temptation is to tell it as a story about a miracle prompt, and it was nothing like that. The model wrote the code, and I made the architecture calls. It rendered the voices, and I threw away the ones with salesperson energy. It drafted the meditations, and the safety gate and the taste behind it were mine. The leverage was genuinely staggering, a single person producing across infrastructure, clinical-adjacent content, audio production, and brand – work that not long ago would have needed a studio and a team and a year. But the judgment never left the loop. AI collapsed the cost of producing each layer; it did not touch the cost of knowing which layer was good, and those turn out to be very different problems. So, the answer to the question I started with, as of today, with the thing live on the App Store and real people about to use it: yes. One person and these tools really can build the whole thing, every layer of it, and have it hold together. The market still gets the final vote on whether it is good, the way the market always does, and I built in the honest off-ramps to listen if it tells me no. The app even ships under a license I wrote called Heirloom, so that if I ever truly walk away from it, the source opens automatically and the thing outlives me, because a promise that says "help is at hand" should not quietly depend on me staying interested forever. What I keep coming back to is that the constraint has moved. For most of my career the hard, scarce thing was the building, and now the building is the part that compresses almost to nothing. What is left, the part that is suddenly the whole job, is taste: knowing what is worth making, and being able to tell the difference between something that looks finished and something that is actually done. The models will make anything you ask for. Deciding what to ask for, and refusing the ninety percent that is merely impressive, is the work that is still entirely yours. Rainfall is live on the App Store now, if you want to see what twenty-one days and a great deal of taste actually produced: --- Canonical: https://alanwizemann.com/articles/could-one-person-and-ai-build-the-whole-thing