Etyde (Music Learning Platform)

Overview

Etyde is a music learning platform connecting private music teachers with students through practice-centric workflows. It replaces the 3-4 separate tools teachers currently juggle (scheduling, lesson notes, practice tracking, sheet music) with a single integrated platform.

Target users: Private music teachers (guitar, piano) with 2-10 students, their students, and parents.

What It Does

• Lesson management with plans, outcomes, notes, and RRULE-based scheduling with timezone handling and iCal export
• Practice sessions with a live timer, auto-complete after inactivity, and integrated metronome and tuner
• Sheet music rendering for PDF, Guitar Pro, and MIDI files via alphaTab, with notation/tablature toggle and track controls
• AI-generated practice sessions that combine a student's goals and repertoire into tailored exercises with progressive difficulty (Practice Sets)
• AI notation generation producing AlphaTex rendered as interactive sheet music, with quality gates and graceful fallback
• Student goals and repertoire tracking with growth dashboards
• Parent portal providing read-only visibility into practice activity, lessons, and goals
• Real-time messaging between teacher and student via Supabase Realtime

How It Fits Together

A Next.js frontend talks to a Fastify API backed by PostgreSQL (via Supabase with row-level security) and Drizzle ORM across 34 schema files. BullMQ + Redis handle async work, with an AI worker running Ollama (Qwen 32B) on an M4 Mac Mini connected over WireGuard. File uploads go through a three-step S3 flow (presign, upload, confirm) to bypass the API for large files. The whole stack deploys via Docker Compose + Ansible + Traefik on a Hetzner VPS, with Watchtower for auto-updates.

Architecture Decisions

• RLS from day one — Multi-tenant isolation lives at the database layer, not in application code. Claude Code managed this across 34 schema files without leakage.
• Generate-on-read for recurring entities — Lesson slots and practice set sessions are computed on query, never materialized. Keeps the data model honest and avoids stale-data bugs.
• Three-step S3 upload — Presign on server, upload direct to storage, confirm back to API. Large files never touch the application server.
• Graceful AI degradation — Notation generation fails 3x? Strip notation. Still fails? Deliver text-only session. The student always gets something.
• Live auto-save everywhere — No "Save" button. Debounced mutations reduce cognitive load for teachers writing lesson plans.

Iteration Story

The project shipped across three milestones. The v1.0 MVP proved the core teaching/learning loop worked. v1.1 added the operational features teachers actually need daily — scheduling, calendar views, the parent portal. The interesting shift came in v1.2 (AI Studio): rather than bolting AI onto existing features, we built a separate practice generation pipeline where AI composes multi-session sequences with progressive difficulty. The key learning was that AI notation generation needed aggressive quality gates — the 89-94% success rate sounds high, but a 10% failure rate in a music lesson is unacceptable, so the fallback chain became the real product design work.

Weaknesses & Open Questions

• VPN brittleness — The WireGuard tunnel to the Mac Mini is a single point of failure for AI features. Graceful degradation is in place, but the failure mode still surprises users (features just quietly disappear).
• Inference latency — 30-90 seconds per AI generation is fine for batch/async but rules out any interactive AI features.
• Redis SPOF — No replication means jobs are lost on restart. This needs proper persistence tuning or a managed alternative.
• alphaTab vendor lock — Requires a webpack plugin, blocking Turbopack. No ESM support on their roadmap.
• PWA deferred — Offline practice tools never shipped. This matters most for students practicing without reliable internet.
• No real user validation yet — The platform is architecturally solid but hasn't been tested with actual teachers and students beyond the developer.

Ecosystem Role

Etyde routes AI inference through Roughneck's plugin-based job queue and reports health to CNC for monitoring and error classification. It shares the music domain with AbletonBuddy (learning vs. production) and Evolver (instrument mastery vs. music theory/practice).