Omii Sound Studio
A taste-gated AI music pipeline — nothing publishes unless it clears aesthetic scoring and a human ear.
AI music tooling optimizes for volume; this optimizes for taste. The pipeline exists to answer one question: can an automated system produce music you'd actually choose to listen to — and refuse to ship what you wouldn't?
A batch pipeline generates candidate tracks on serverless GPUs, scores every candidate with aesthetic models, masters and QA-checks the survivors, and holds everything at a human taste gate. Nothing publishes without approval. A self-serve web studio on the same core is in development.
Architecture
Node orchestrates the pipeline; ACE-Step (a 3.5B-parameter music model) runs on Modal serverless GPUs; curation is best-of-N with Audiobox Aesthetics and LAION-CLAP scoring; Python handles audio analysis and loudness mastering (librosa, pyloudnorm). The taste gate is structural — the publish step physically requires a human approval record, so the system cannot ship slop even if every model in the chain says yes.
Production posture
The interesting engineering is in the refusal path. Every candidate track is scored by aesthetic models before a human ever hears it; only the best of each batch survives to mastering, where loudness normalization and QA checks run before anything reaches the approval queue. Generation cost stays bounded because GPU time is serverless and batch-scoped, and every pipeline run is logged end to end — which candidates were generated, what scored them down, and who approved what — so any published track's provenance is reconstructable.