Our own project · In development
Perfect Studio
A cross-platform app that generates complete songs — lyrics, vocals, and instrumentation — from an idea, using state-of-the-art generative AI models. Not a client project either — it's ours.
What Perfect Studio is
Perfect Studio generates complete songs — lyrics, vocals, and instrumentation — from an idea, a style, or your own lyrics, in whichever language each user picks. You can choose genre and mood, generate an instrumental version, regenerate just one section, split the song into stems (vocals, drums, bass, melody), or extend it. The business model is freemium: a few free songs in exchange for verified ads, and several subscription tiers for anyone who wants more.
How is it built under the hood?
The client — a single codebase for Android, iPhone/iPad, and browser — never talks directly to any AI model and never holds any access key of its own: that would be exactly the kind of credential leak this site has been avoiding from the start. Every request first passes through a lightweight orchestrator that verifies identity, subscription, monthly quota, and ads watched before accepting the job — only then is it queued.
The design we built on top of that queue is what actually makes the cost difference: every region of the world has its own job queue and its own group of machines that scales automatically — from zero to several instances when there are songs to generate, and back to zero minutes after running out of work. If nobody in a given region is requesting songs for a while, that region costs nothing. It's an architecture of our own, portable to any cloud provider that offers job queues and autoscaling instance groups — it doesn't depend on any one in particular.
Where it stands
The project runs through 8 planned phases, and four are now closed and verified end to end with real data, not merely deployed: a real user signs up, their account is created automatically, they ask for a song, the system checks their quota and credits, the job genuinely reaches their region's queue, and it comes back as a playable audio file. The app already builds and runs on Android and in the browser (iOS needs hardware we don't have yet), with all 12 launch languages translated — pending native-speaker review in the ten that aren't Spanish or English.
Music generation is now closed: the full pipeline — lyrics and music — works end to end and returns real, playable audio, not a well-formed URL; the player, with its waveform and tap-to-seek, is verified against that audio. We're currently in the monetisation phase: mandatory prior consent before any advertising code loads in Europe, mediation across several networks, an offer wall with every reward cryptographically signed and reconciled daily, and subscriptions always validated on the server through a signed, idempotent webhook, never on the client. All of it can be switched on and off from our own admin panel.
Before calling that phase done we ran an end-to-end bug audit, with particular attention to the money path: 24 findings, 23 fixed and each with its own automated test, one under watch, none left open. The worst of them: the process that generates songs wasn't idempotent, so a queue message redelivered after the job had already finished re-ran the work, corrupted the audio and charged twice. Four phases remain: security hardening, the admin panel, social features, and launch. What's holding it up today isn't code — the app is signed and packaged; what's left is publishing it and completing the account and store paperwork. None of this has an announced date.
What's already real