Zics Vault

Personal music knowledge base

Zics Vault

A private library of enriched tracks — searchable, playable, explorable. Built with a full ingestion pipeline and AI-assisted metadata.

0
Tracks
0
Artists
0
Collections
0
Hours

Catalog snapshot — live totals load when the database is connected

Volumetry

Library snapshot

Representative catalog figures — connect the database for live totals.

0
Tracks indexed
0.0
Storage (GB)
0
Albums
0 kbps
Avg bitrate

Top genres

Enrichment coverage

Genre tagged88%
Decade tagged76%
Mood tagged82%

Top moods

  • melancholic1106
  • nostalgic757
  • energetic605
  • uplifting446

Top decades

  • 2010s1122
  • 2000s478
  • 1980s357
  • 1970s285

Sources

ComputerTracksGB
macbook-20162,10028
macbook-20231,90024
other1,55720

Collection types

ML clusters by lens

fused (128)sonic (127)metadata (128)semantic (127)

Playlists

legacy (45)user (12)

How it works

Ingestion pipeline

From raw MP3s on S3 to a searchable knowledge base. Runs on your Mac; the web app reads precomputed data.

Tools

S3 scan, CSV, iTunes XML

Output

Recordings, artists, playlists, enrichment jobs

LLM

No — deterministic

S3 audio → ingest → enrich (librosa) → identify (transcribe + web search) → facts (MB + web search) → standardize → dedup → cluster (UMAP) → curate (LLM) → postcluster → Postgres → Next.js app

Engineering

Architecture

Four app modules over a Postgres knowledge base. Pipeline runs locally; Heroku serves the UI.

Tech stack

Next.js 15PostgresAWS S3OpenAIPythonrechartsHerokuDrizzle ORM
┌─────────────┐     ┌──────────────┐     ┌─────────────┐
│  S3 (audio) │────▶│ Mac pipeline │────▶│  Postgres   │
└─────────────┘     │ ingest→postcluster│     │ knowledge   │
                    └──────────────┘     └──────┬──────┘
                                                │
┌─────────────┐     ┌──────────────┐            │
│   OpenAI    │◀───│  Next.js 15  │◀───────────┘
│ facts/curate/identify│     │ Heroku (web) │
└─────────────┘     └──────────────┘

AI layer

Intelligence

LLM-assisted enrichment with citations and cached responses. Operational cost details stay in admin Settings.

Hybrid facts

Per-track genre, mood, decade, and description from MusicBrainz, Wikipedia, and LLM synthesis — with source URLs.

Transcription identity

Unlabeled or filename-only tracks get Whisper transcription plus web-search identity with confidence scores.

Acoustic dedup

SHA256 exact-match and acoustic fingerprint candidates — reviewed manually before any merge.

Cluster curation

Collections named and described by a reasoning model; post-cluster LLM proposes merges and completions.

Discovery Lab

Natural-language queries translated to structured filters for semantic search.

Cost-aware batching

LLM responses cached by prompt hash. Pipeline runs track usage for operator review in Settings.

LLM cache

—

Cache stats appear when the database is connected.

Usage analytics

  • · Sessions & unique visitors (admin)
  • · Tracks played & queue depth
  • · Search queries & filter patterns

Models in use

  • gpt-5-mini — facts synthesis, discovery, postcluster
  • gpt-5.5 — cluster curation, identify (web search)
  • whisper-1 — transcription for identity fallback
  • text-embedding-3-small — semantic clusters

Status

Development progress

A living project. Core pipeline and library are built; usage analytics and sharing are next.

70%

7 of 10 milestones shipped

Authentication & accountsBuilt

Cognito sign-in; per-account playlists and listening

Library consolidationBuilt

Tracks, artists, collections unified

Hybrid facts enrichmentBuilt

MusicBrainz + LLM with citations

Music map & clusteringBuilt

Four lenses, fused coordinates

Acoustic dedup & identityBuilt

Fingerprint + transcription fallback

Landing pageBuilt

Public overview with live stats when database is connected

Usage analyticsBuilt

Plays, heard/completed, listening dashboard

Placement builderIn progress

Interactive playlist candidate review

Public sharingPlanned

Curated read-only views

Mobile PWAPlanned

Offline queue, installable