Development Progress

Lyra AI Hip-Hop Platform — Backend Pipeline

Chenye Wang · Peijun Wu · Peabody Institute, Johns Hopkins University

Versions shipped

36+

Tests passing

Deploying

Status

v0.1

Foundation

February 12, 2026

Architecture

·DAG-based pipeline: Orchestrator dispatches tasks to modular Executors
·Versioned artifact store for reproducible project state
·State machine managing pipeline transitions (IDLE → EXPORT_DONE)
·Policy config (YAML) for system-level constraints

Music engine

·Harmony module — key parsing, scale/chord generation, progressions
·MIDI patterns — boom_bap + trap drum patterns with energy variation
·Audio utilities — pad/truncate, peak/RMS, limiter, mix, WAV I/O
·FluidSynth rendering with graceful sine/noise fallback

Milestone

·21 tests passing (schema, timeline, QC, E2E smoke)
·First full E2E run (Feb 10): MIDI → render → QC → mix.wav (15.8 MB)

v0.2

Conversational Interface

February 24, 2026

ChatExecutor (new)

·Gemini-powered chatbot for brief creation and patching
·Two modes: create from scratch or patch existing brief
·Patch whitelist — only approved fields (BPM, key, mood, style) editable via chat
·Brief sufficiency check — generates only when enough intent is present

UI overhaul

·Replaced 4-page Streamlit app with ChatGPT-style project chat
·Sidebar project list, bottom chat input, floating mini audio player
·Post-generation summary: style, mood, BPM, key, instruments with AI reasoning
·3 clickable example prompt chips for onboarding

Pipeline

·Instruments selection moved after brief lock
·instrument_map.json separated from brief.json as dedicated artifact
·metadata.json as single source of truth for all project state

v0.2.1

Instrument Control

March 27, 2026

Instrument patching (new)

·detect_instrument_request() — distinguishes instrument changes from mood/BPM changes
·apply_instrument_patch() — surgically updates individual stems without touching others
·compute_instrument_diff() — human-readable diff shown to user after patch
·InstrumentPreset schema: bank, program, name, reason for full GM soundfont control

Detection examples

·"I want to try another bass sound" → instrument request ✓
·"switch drums to Standard kit" → instrument request ✓
·"make it darker" → NOT instrument — handled by brief patch
·"slow down to 120 BPM" → NOT instrument — handled by brief patch

Tests

·15 new tests covering detection, apply, and diff
·False-positive prevention: mood/BPM/generic praise does not trigger instrument detection

v0.3Current

Selective Regeneration

April 7, 2026

Stem selection system (new)

·parse_stem_selection() — Gemini decides which stems to regenerate vs. keep
·StemMixMap schema — tracks which version of each stem is in the active mix
·Version resolution: null (exclude), NEW (regenerate), int (specific version), KEEP (current)

Selective regeneration examples

·"Only change the drums" → generate: {drums}, bass + pads kept at current versions
·"Redo everything" → generate: {drums, bass, pads}, full regeneration
·Error fallback: defaults to full regeneration if parse fails

Growth

·ChatExecutor: 379 lines (v0.2) → 488 (v0.2.1) → 735 (v0.3)
·midi_patterns.py: 257 lines (v0.1) → 336 (v0.3)
·models.py: 273 lines (v0.1) → 353 (v0.3)

lyra_webMarch 27, 2026

Public Landing Page

Separate frontend codebase powering lyrapro.ai — not a backend version update.

·Next.js 16 + React 19 + TypeScript + Tailwind CSS 4
·Email waitlist collection via Formspree
·Demo audio embedded for product preview
·Dark-mode design with purple/fuchsia atmospheric gradients

Next Steps

01Server deployment — Railway/Render, shareable link for real users

02Fix Google Gemini API version issue (current pipeline blocker)

03Chatbot output testing with MiniMax / GPT APIs

04First 10 real users — collect expression feedback

05Plugin system — EQ, compressor, reverb for audio processing layer

06Company incorporation — Delaware C-Corp (target: May 2026)

lyrapro.ai · April 2026Back to home →