Case study

Ensayadero.studio

A rehearsal room that listens back.

A rehearsal tool for bands: upload a recording and get it back split into stems, with MIDI and an editable chord chart, in a multi-track mixer. In invite-only beta.

See it live

In private beta. If you'd like to try it, I'll set you up with an account. beltran@rengifo.es

The Ensayadero home screen: what goes in, what the pipeline does, and what comes out. — The product explained in three columns, above the recorder that takes a rehearsal straight off the room mic.

A session playing: transport, stereo meters, a channel strip with its waveform and a loop marker. — Playback with per-channel mute, solo and EQ, and loop markers for the bars that need repeating.

The chord chart in editing mode, with sections, chords, lap counts and timings. — The chart the models guessed, open for correction: sections, chords, laps and timings. The band's edits survive reprocessing.

Self-hosted audio ML, no paid APIs

A Python pipeline — Demucs for stem separation, Basic Pitch and Omnizart for transcription, Omnizart and madmom for chords — runs on TensorFlow and PyTorch, which is why it ships in the same Docker image as the Next.js server, self-hosted with Coolify, behind a job queue that recovers interrupted work on restart.

Next.js on Supabase

Next.js 16 and React 19 with RLS-aware API routes; Supabase for Postgres, Google OAuth with an allow-list and role claims, and Storage for stems, MIDI and charts. Playback in the browser with Tone.js and wavesurfer.

← Back to the résumé

Self-hosted audio ML, no paid APIs

Next.js on Supabase

Need an engineer who ships end to end — and still sweats the pixels? Let's talk.