Initial haunt-fm implementation

Full music recommendation pipeline: listening history capture via webhook, Last.fm candidate discovery, iTunes preview download, CLAP audio embeddings (512-dim), pgvector cosine similarity recommendations, playlist generation with known/new track interleaving, and Music Assistant playback via HA. Includes: FastAPI app, SQLAlchemy models, Alembic migrations, Docker Compose with pgvector/pg17, status dashboard, and all API endpoints. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-22 08:36:36 -06:00
parent 897d0fe1fb
commit 7ff69449d6
39 changed files with 2049 additions and 0 deletions
--- a/.env.example
+++ b/.env.example
@@ -0,0 +1,23 @@
+# PostgreSQL
+POSTGRES_DB=hauntfm
+POSTGRES_USER=hauntfm
+POSTGRES_PASSWORD=changeme
+
+# Database URL (constructed in docker-compose, override for local dev)
+DATABASE_URL=postgresql+asyncpg://hauntfm:changeme@localhost:5432/hauntfm
+
+# Last.fm API
+HAUNTFM_LASTFM_API_KEY=
+
+# Home Assistant
+HAUNTFM_HA_URL=http://192.168.86.51:8123
+HAUNTFM_HA_TOKEN=
+
+# CLAP model
+HAUNTFM_MODEL_CACHE_DIR=/app/model-cache
+HAUNTFM_AUDIO_CACHE_DIR=/app/audio-cache
+
+# Embedding worker
+HAUNTFM_EMBEDDING_WORKER_ENABLED=true
+HAUNTFM_EMBEDDING_BATCH_SIZE=10
+HAUNTFM_EMBEDDING_INTERVAL_SECONDS=30
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -0,0 +1,54 @@
+# haunt-fm
+
+Personal music recommendation service. Captures listening history from Music Assistant, discovers similar tracks via Last.fm, embeds audio with CLAP, and generates playlists.
+
+## Quick Start
+
+```bash
+# On NAS
+cd /volume1/homes/antialias/projects/haunt-fm
+docker compose up -d
+docker compose exec haunt-fm alembic upgrade head
+```
+
+## Architecture
+
+- **FastAPI** app with async SQLAlchemy + asyncpg
+- **PostgreSQL + pgvector** for tracks, embeddings, and vector similarity search
+- **CLAP model** (laion/larger_clap_music) for 512-dim audio embeddings
+- **Last.fm API** for track similarity discovery
+- **iTunes Search API** for 30-second audio previews
+- **Music Assistant** (via Home Assistant REST API) for playback
+
+## Key Commands
+
+```bash
+# Health check
+curl http://192.168.86.51:8321/health
+
+# Log a listen event
+curl -X POST http://192.168.86.51:8321/api/history/webhook \
+  -H "Content-Type: application/json" \
+  -d '{"title":"Song","artist":"Artist"}'
+
+# Run discovery
+curl -X POST http://192.168.86.51:8321/api/admin/discover -H "Content-Type: application/json" -d '{}'
+
+# Get recommendations
+curl http://192.168.86.51:8321/api/recommendations?limit=20
+
+# Generate and play a playlist
+curl -X POST http://192.168.86.51:8321/api/playlists/generate \
+  -H "Content-Type: application/json" \
+  -d '{"total_tracks":20,"known_pct":30,"speaker_entity":"media_player.living_room_2","auto_play":true}'
+```
+
+## Environment Variables
+
+All prefixed with `HAUNTFM_`. See `.env.example` for full list.
+
+## Database
+
+- Alembic migrations in `alembic/versions/`
+- Run migrations: `alembic upgrade head`
+- Schema: tracks, listen_events, track_embeddings, similarity_links, taste_profiles, playlists, playlist_tracks
--- a/28
+++ b/28
@@ -0,0 +1,28 @@
+FROM python:3.12-slim
+
+# System deps for librosa/soundfile (ffmpeg for AAC decoding)
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    ffmpeg \
+    libsndfile1 \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+# Install Python deps
+COPY pyproject.toml .
+RUN pip install --no-cache-dir .
+
+# Copy source
+COPY alembic.ini .
+COPY alembic/ alembic/
+COPY src/ src/
+
+# Install the project itself
+RUN pip install --no-cache-dir -e .
+
+# Create cache directories
+RUN mkdir -p /app/model-cache /app/audio-cache
+
+EXPOSE 8000
+
+CMD ["uvicorn", "haunt_fm.main:app", "--host", "0.0.0.0", "--port", "8000"]
--- a/alembic.ini
+++ b/alembic.ini
@@ -0,0 +1,36 @@
+[alembic]
+script_location = alembic
+sqlalchemy.url = driver://user:pass@localhost/dbname
+
+[loggers]
+keys = root,sqlalchemy,alembic
+
+[handlers]
+keys = console
+
+[formatters]
+keys = generic
+
+[logger_root]
+level = WARN
+handlers = console
+
+[logger_sqlalchemy]
+level = WARN
+handlers =
+qualname = sqlalchemy.engine
+
+[logger_alembic]
+level = INFO
+handlers =
+qualname = alembic
+
+[handler_console]
+class = StreamHandler
+args = (sys.stderr,)
+level = NOTSET
+formatter = generic
+
+[formatter_generic]
+format = %(levelname)-5.5s [%(name)s] %(message)s
+datefmt = %H:%M:%S
--- a/alembic/env.py
+++ b/alembic/env.py
@@ -0,0 +1,69 @@
+import asyncio
+import os
+from logging.config import fileConfig
+
+from alembic import context
+from sqlalchemy import pool
+from sqlalchemy.ext.asyncio import async_engine_from_config
+
+from haunt_fm.models.base import Base
+
+# Import all models so they register with Base.metadata
+from haunt_fm.models.track import (  # noqa: F401
+    ListenEvent,
+    Playlist,
+    PlaylistTrack,
+    SimilarityLink,
+    TasteProfile,
+    Track,
+    TrackEmbedding,
+)
+
+config = context.config
+
+if config.config_file_name is not None:
+    fileConfig(config.config_file_name)
+
+target_metadata = Base.metadata
+
+# Override sqlalchemy.url from environment
+db_url = os.environ.get("HAUNTFM_DATABASE_URL", "")
+if db_url:
+    config.set_main_option("sqlalchemy.url", db_url)
+
+
+def run_migrations_offline() -> None:
+    url = config.get_main_option("sqlalchemy.url")
+    context.configure(url=url, target_metadata=target_metadata, literal_binds=True)
+
+    with context.begin_transaction():
+        context.run_migrations()
+
+
+def do_run_migrations(connection):
+    context.configure(connection=connection, target_metadata=target_metadata)
+    with context.begin_transaction():
+        context.run_migrations()
+
+
+async def run_async_migrations() -> None:
+    connectable = async_engine_from_config(
+        config.get_section(config.config_ini_section, {}),
+        prefix="sqlalchemy.",
+        poolclass=pool.NullPool,
+    )
+
+    async with connectable.connect() as connection:
+        await connection.run_sync(do_run_migrations)
+
+    await connectable.dispose()
+
+
+def run_migrations_online() -> None:
+    asyncio.run(run_async_migrations())
+
+
+if context.is_offline_mode():
+    run_migrations_offline()
+else:
+    run_migrations_online()
--- a/alembic/script.py.mako
+++ b/alembic/script.py.mako
@@ -0,0 +1,26 @@
+"""${message}
+
+Revision ID: ${up_revision}
+Revises: ${down_revision | comma,n}
+Create Date: ${create_date}
+
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+${imports if imports else ""}
+
+# revision identifiers, used by Alembic.
+revision: str = ${repr(up_revision)}
+down_revision: Union[str, None] = ${repr(down_revision)}
+branch_labels: Union[str, Sequence[str], None] = ${repr(branch_labels)}
+depends_on: Union[str, Sequence[str], None] = ${repr(depends_on)}
+
+
+def upgrade() -> None:
+    ${upgrades if upgrades else "pass"}
+
+
+def downgrade() -> None:
+    ${downgrades if downgrades else "pass"}
--- a/alembic/versions/001_initial_schema.py
+++ b/alembic/versions/001_initial_schema.py
@@ -0,0 +1,121 @@
+"""Initial schema
+
+Revision ID: 001
+Revises:
+Create Date: 2026-02-22
+
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from pgvector.sqlalchemy import Vector
+
+revision: str = "001"
+down_revision: Union[str, None] = None
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Enable pgvector extension
+    op.execute("CREATE EXTENSION IF NOT EXISTS vector")
+
+    # Tracks
+    op.create_table(
+        "tracks",
+        sa.Column("id", sa.BigInteger, primary_key=True),
+        sa.Column("title", sa.Text, nullable=False),
+        sa.Column("artist", sa.Text, nullable=False),
+        sa.Column("album", sa.Text),
+        sa.Column("fingerprint", sa.Text, unique=True, nullable=False),
+        sa.Column("lastfm_url", sa.Text),
+        sa.Column("itunes_track_id", sa.BigInteger),
+        sa.Column("itunes_preview_url", sa.Text),
+        sa.Column("apple_music_id", sa.Text),
+        sa.Column("duration_ms", sa.Integer),
+        sa.Column("genre", sa.Text),
+        sa.Column("embedding_status", sa.Text, nullable=False, server_default="pending"),
+        sa.Column("embedding_error", sa.Text),
+        sa.Column("created_at", sa.DateTime(timezone=True), server_default=sa.func.now()),
+        sa.Column("updated_at", sa.DateTime(timezone=True), server_default=sa.func.now()),
+    )
+
+    # Listen events
+    op.create_table(
+        "listen_events",
+        sa.Column("id", sa.BigInteger, primary_key=True),
+        sa.Column("track_id", sa.BigInteger, sa.ForeignKey("tracks.id"), nullable=False),
+        sa.Column("source", sa.Text, nullable=False, server_default="music_assistant"),
+        sa.Column("speaker_name", sa.Text),
+        sa.Column("listened_at", sa.DateTime(timezone=True), server_default=sa.func.now()),
+        sa.Column("duration_played", sa.Integer),
+        sa.Column("raw_payload", sa.dialects.postgresql.JSONB),
+    )
+
+    # Track embeddings (512-dim CLAP)
+    op.create_table(
+        "track_embeddings",
+        sa.Column("id", sa.BigInteger, primary_key=True),
+        sa.Column("track_id", sa.BigInteger, sa.ForeignKey("tracks.id"), unique=True, nullable=False),
+        sa.Column("embedding", Vector(512), nullable=False),
+        sa.Column("model_version", sa.Text, nullable=False, server_default="laion/larger_clap_music"),
+        sa.Column("created_at", sa.DateTime(timezone=True), server_default=sa.func.now()),
+    )
+    op.execute(
+        "CREATE INDEX ix_track_embeddings_hnsw ON track_embeddings "
+        "USING hnsw (embedding vector_cosine_ops)"
+    )
+
+    # Similarity links
+    op.create_table(
+        "similarity_links",
+        sa.Column("id", sa.BigInteger, primary_key=True),
+        sa.Column("source_track_id", sa.BigInteger, sa.ForeignKey("tracks.id"), nullable=False),
+        sa.Column("target_track_id", sa.BigInteger, sa.ForeignKey("tracks.id"), nullable=False),
+        sa.Column("lastfm_match", sa.Real),
+        sa.Column("created_at", sa.DateTime(timezone=True), server_default=sa.func.now()),
+        sa.UniqueConstraint("source_track_id", "target_track_id", name="uq_similarity_link"),
+    )
+
+    # Taste profiles
+    op.create_table(
+        "taste_profiles",
+        sa.Column("id", sa.BigInteger, primary_key=True),
+        sa.Column("name", sa.Text, unique=True, nullable=False, server_default="default"),
+        sa.Column("embedding", Vector(512), nullable=False),
+        sa.Column("track_count", sa.Integer, nullable=False),
+        sa.Column("updated_at", sa.DateTime(timezone=True), server_default=sa.func.now()),
+    )
+
+    # Playlists
+    op.create_table(
+        "playlists",
+        sa.Column("id", sa.BigInteger, primary_key=True),
+        sa.Column("name", sa.Text),
+        sa.Column("known_pct", sa.Integer, nullable=False),
+        sa.Column("total_tracks", sa.Integer, nullable=False),
+        sa.Column("created_at", sa.DateTime(timezone=True), server_default=sa.func.now()),
+    )
+
+    op.create_table(
+        "playlist_tracks",
+        sa.Column("id", sa.BigInteger, primary_key=True),
+        sa.Column("playlist_id", sa.BigInteger, sa.ForeignKey("playlists.id", ondelete="CASCADE"), nullable=False),
+        sa.Column("track_id", sa.BigInteger, sa.ForeignKey("tracks.id"), nullable=False),
+        sa.Column("position", sa.Integer, nullable=False),
+        sa.Column("is_known", sa.Boolean, nullable=False),
+        sa.Column("similarity_score", sa.Real),
+    )
+
+
+def downgrade() -> None:
+    op.drop_table("playlist_tracks")
+    op.drop_table("playlists")
+    op.drop_table("taste_profiles")
+    op.drop_table("similarity_links")
+    op.execute("DROP INDEX IF EXISTS ix_track_embeddings_hnsw")
+    op.drop_table("track_embeddings")
+    op.drop_table("listen_events")
+    op.drop_table("tracks")
+    op.execute("DROP EXTENSION IF EXISTS vector")
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -0,0 +1,53 @@
+services:
+  haunt-fm-db:
+    image: pgvector/pgvector:pg17
+    container_name: haunt-fm-db
+    restart: unless-stopped
+    environment:
+      POSTGRES_DB: ${POSTGRES_DB:-hauntfm}
+      POSTGRES_USER: ${POSTGRES_USER:-hauntfm}
+      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD}
+    volumes:
+      - haunt-fm-db-data:/var/lib/postgresql/data
+    healthcheck:
+      test: ["CMD", "pg_isready", "-U", "${POSTGRES_USER:-hauntfm}"]
+      interval: 5s
+      timeout: 5s
+      retries: 5
+    networks:
+      - haunt-fm-internal
+
+  haunt-fm:
+    build: .
+    container_name: haunt-fm
+    restart: unless-stopped
+    env_file: .env
+    environment:
+      HAUNTFM_DATABASE_URL: postgresql+asyncpg://${POSTGRES_USER:-hauntfm}:${POSTGRES_PASSWORD}@haunt-fm-db:5432/${POSTGRES_DB:-hauntfm}
+    ports:
+      - "8321:8000"
+    volumes:
+      - haunt-fm-model-cache:/app/model-cache
+      - haunt-fm-audio-cache:/app/audio-cache
+    depends_on:
+      haunt-fm-db:
+        condition: service_healthy
+    healthcheck:
+      test: ["CMD", "python", "-c", "import httpx; httpx.get('http://localhost:8000/health').raise_for_status()"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 60s
+    networks:
+      - haunt-fm-internal
+      - webgateway
+
+networks:
+  haunt-fm-internal:
+  webgateway:
+    external: true
+
+volumes:
+  haunt-fm-db-data:
+  haunt-fm-model-cache:
+  haunt-fm-audio-cache:
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -0,0 +1,31 @@
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[project]
+name = "haunt-fm"
+version = "0.1.0"
+description = "Personal music recommendation service"
+requires-python = ">=3.12"
+dependencies = [
+    "fastapi>=0.115",
+    "uvicorn[standard]>=0.34",
+    "sqlalchemy[asyncio]>=2.0",
+    "asyncpg>=0.30",
+    "alembic>=1.14",
+    "pydantic-settings>=2.7",
+    "pgvector>=0.3",
+    "httpx>=0.28",
+    "jinja2>=3.1",
+    "numpy>=1.26",
+    "librosa>=0.10",
+    "transformers>=4.48",
+    "torch>=2.5",
+    "soundfile>=0.12",
+]
+
+[project.optional-dependencies]
+dev = [
+    "pytest>=8.0",
+    "pytest-asyncio>=0.24",
+]
--- a/scripts/seed_history_from_ma.py
+++ b/scripts/seed_history_from_ma.py
@@ -0,0 +1,70 @@
+#!/usr/bin/env python3
+"""One-time backfill: pull recently played tracks from Music Assistant via HA REST API."""
+import asyncio
+import os
+
+import httpx
+
+HA_URL = os.environ.get("HAUNTFM_HA_URL", "http://192.168.86.51:8123")
+HA_TOKEN = os.environ.get("HAUNTFM_HA_TOKEN", "")
+HAUNTFM_URL = os.environ.get("HAUNTFM_URL", "http://localhost:8321")
+
+
+async def get_recently_played() -> list[dict]:
+    """Get recently played items from Music Assistant via HA."""
+    headers = {
+        "Authorization": f"Bearer {HA_TOKEN}",
+        "Content-Type": "application/json",
+    }
+    async with httpx.AsyncClient(timeout=30) as client:
+        # Get all media_player entities
+        resp = await client.get(f"{HA_URL}/api/states", headers=headers)
+        resp.raise_for_status()
+        states = resp.json()
+
+    # Filter for music assistant players that have media info
+    tracks = []
+    for state in states:
+        if not state["entity_id"].startswith("media_player."):
+            continue
+        attrs = state.get("attributes", {})
+        title = attrs.get("media_title")
+        artist = attrs.get("media_artist")
+        if title and artist:
+            tracks.append({
+                "title": title,
+                "artist": artist,
+                "album": attrs.get("media_album_name"),
+                "speaker_name": attrs.get("friendly_name"),
+                "source": "music_assistant_backfill",
+            })
+
+    return tracks
+
+
+async def send_to_webhook(track: dict):
+    async with httpx.AsyncClient(timeout=10) as client:
+        resp = await client.post(f"{HAUNTFM_URL}/api/history/webhook", json=track)
+        resp.raise_for_status()
+        return resp.json()
+
+
+async def main():
+    if not HA_TOKEN:
+        print("Set HAUNTFM_HA_TOKEN environment variable")
+        return
+
+    print(f"Fetching from {HA_URL}...")
+    tracks = await get_recently_played()
+    print(f"Found {len(tracks)} tracks with media info")
+
+    for track in tracks:
+        try:
+            result = await send_to_webhook(track)
+            print(f"  OK: {track['artist']} - {track['title']} -> track_id={result['track_id']}")
+        except Exception as e:
+            print(f"  FAIL: {track['artist']} - {track['title']}: {e}")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
--- a/src/haunt_fm/init.py
+++ b/src/haunt_fm/init.py
--- a/src/haunt_fm/api/init.py
+++ b/src/haunt_fm/api/init.py
--- a/src/haunt_fm/api/admin.py
+++ b/src/haunt_fm/api/admin.py
@@ -0,0 +1,59 @@
+from fastapi import APIRouter, Depends
+from pydantic import BaseModel
+from sqlalchemy import func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from haunt_fm.db import get_session
+from haunt_fm.models.track import ListenEvent, Track
+from haunt_fm.services.lastfm_client import discover_similar_for_track
+from haunt_fm.services.taste_profile import build_taste_profile
+
+router = APIRouter(prefix="/api/admin")
+
+
+class DiscoverRequest(BaseModel):
+    limit: int = 50  # max tracks from history to expand
+
+
+@router.post("/discover")
+async def discover(req: DiscoverRequest, session: AsyncSession = Depends(get_session)):
+    """Expand listening history via Last.fm track.getSimilar."""
+    # Get most-listened tracks that haven't been expanded yet
+    listened_tracks = (
+        await session.execute(
+            select(Track)
+            .join(ListenEvent, ListenEvent.track_id == Track.id)
+            .group_by(Track.id)
+            .order_by(func.count(ListenEvent.id).desc())
+            .limit(req.limit)
+        )
+    ).scalars().all()
+
+    total_discovered = 0
+    errors = []
+
+    for track in listened_tracks:
+        try:
+            count = await discover_similar_for_track(session, track)
+            total_discovered += count
+        except Exception as e:
+            errors.append({"track": f"{track.artist} - {track.title}", "error": str(e)})
+
+    return {
+        "tracks_expanded": len(listened_tracks),
+        "candidates_discovered": total_discovered,
+        "errors": errors,
+    }
+
+
+@router.post("/build-taste-profile")
+async def build_profile(session: AsyncSession = Depends(get_session)):
+    """Rebuild the taste profile from listened-track embeddings."""
+    profile = await build_taste_profile(session)
+    if profile is None:
+        return {"ok": False, "error": "No listened tracks with embeddings found"}
+    return {
+        "ok": True,
+        "track_count": profile.track_count,
+        "updated_at": profile.updated_at.isoformat(),
+    }
--- a/src/haunt_fm/api/health.py
+++ b/src/haunt_fm/api/health.py
@@ -0,0 +1,19 @@
+from fastapi import APIRouter, Depends
+from sqlalchemy import text
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from haunt_fm.db import get_session
+
+router = APIRouter()
+
+
+@router.get("/health")
+async def health(session: AsyncSession = Depends(get_session)):
+    try:
+        await session.execute(text("SELECT 1"))
+        db_ok = True
+    except Exception:
+        db_ok = False
+
+    status = "healthy" if db_ok else "degraded"
+    return {"status": status, "db_connected": db_ok}
--- a/src/haunt_fm/api/history.py
+++ b/src/haunt_fm/api/history.py
@@ -0,0 +1,39 @@
+from datetime import datetime, timezone
+
+from fastapi import APIRouter, Depends
+from pydantic import BaseModel
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from haunt_fm.db import get_session
+from haunt_fm.services.history_ingest import ingest_listen_event
+
+router = APIRouter(prefix="/api/history")
+
+
+class WebhookPayload(BaseModel):
+    title: str
+    artist: str
+    album: str | None = None
+    speaker_name: str | None = None
+    duration_played: int | None = None
+    source: str = "music_assistant"
+    listened_at: datetime | None = None
+
+
+@router.post("/webhook")
+async def receive_webhook(payload: WebhookPayload, session: AsyncSession = Depends(get_session)):
+    if payload.listened_at is None:
+        payload.listened_at = datetime.now(timezone.utc)
+
+    event = await ingest_listen_event(
+        session=session,
+        title=payload.title,
+        artist=payload.artist,
+        album=payload.album,
+        speaker_name=payload.speaker_name,
+        duration_played=payload.duration_played,
+        source=payload.source,
+        listened_at=payload.listened_at,
+        raw_payload=payload.model_dump(mode="json"),
+    )
+    return {"ok": True, "track_id": event.track_id, "event_id": event.id}
--- a/src/haunt_fm/api/playlists.py
+++ b/src/haunt_fm/api/playlists.py
@@ -0,0 +1,63 @@
+from fastapi import APIRouter, Depends
+from pydantic import BaseModel
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from haunt_fm.db import get_session
+from haunt_fm.models.track import PlaylistTrack, Track
+from haunt_fm.services.music_assistant import play_playlist_on_speaker
+from haunt_fm.services.playlist_generator import generate_playlist
+
+router = APIRouter(prefix="/api/playlists")
+
+
+class GenerateRequest(BaseModel):
+    total_tracks: int = 20
+    known_pct: int = 30
+    name: str | None = None
+    speaker_entity: str | None = None
+    auto_play: bool = False
+
+
+@router.post("/generate")
+async def generate(req: GenerateRequest, session: AsyncSession = Depends(get_session)):
+    playlist = await generate_playlist(
+        session,
+        total_tracks=req.total_tracks,
+        known_pct=req.known_pct,
+        name=req.name,
+    )
+
+    # Load playlist tracks with track info
+    result = await session.execute(
+        select(PlaylistTrack, Track)
+        .join(Track, PlaylistTrack.track_id == Track.id)
+        .where(PlaylistTrack.playlist_id == playlist.id)
+        .order_by(PlaylistTrack.position)
+    )
+    rows = result.all()
+
+    track_list = [
+        {
+            "position": pt.position,
+            "artist": t.artist,
+            "title": t.title,
+            "album": t.album,
+            "is_known": pt.is_known,
+            "similarity_score": pt.similarity_score,
+        }
+        for pt, t in rows
+    ]
+
+    # Auto-play if requested
+    if req.auto_play and req.speaker_entity:
+        await play_playlist_on_speaker(track_list, req.speaker_entity)
+
+    return {
+        "playlist_id": playlist.id,
+        "name": playlist.name,
+        "total_tracks": playlist.total_tracks,
+        "known_pct": playlist.known_pct,
+        "tracks": track_list,
+        "auto_played": req.auto_play and req.speaker_entity is not None,
+    }
--- a/src/haunt_fm/api/recommendations.py
+++ b/src/haunt_fm/api/recommendations.py
@@ -0,0 +1,19 @@
+from fastapi import APIRouter, Depends, Query
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from haunt_fm.db import get_session
+from haunt_fm.services.recommender import get_recommendations
+
+router = APIRouter(prefix="/api")
+
+
+@router.get("/recommendations")
+async def recommendations(
+    limit: int = Query(default=50, ge=1, le=200),
+    include_known: bool = Query(default=False),
+    session: AsyncSession = Depends(get_session),
+):
+    results = await get_recommendations(
+        session, limit=limit, exclude_known=not include_known
+    )
+    return {"recommendations": results, "count": len(results)}
--- a/src/haunt_fm/api/status.py
+++ b/src/haunt_fm/api/status.py
@@ -0,0 +1,113 @@
+from datetime import datetime, timedelta, timezone
+
+from fastapi import APIRouter, Depends
+from sqlalchemy import func, select, text
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from haunt_fm.db import get_session
+from haunt_fm.models.track import (
+    ListenEvent,
+    Playlist,
+    TasteProfile,
+    Track,
+    TrackEmbedding,
+)
+from haunt_fm.config import settings
+from haunt_fm.services.embedding import is_model_loaded
+from haunt_fm.services.embedding_worker import is_running as is_worker_running
+from haunt_fm.services.embedding_worker import last_processed as worker_last_processed
+
+router = APIRouter(prefix="/api")
+
+
+@router.get("/status")
+async def status(session: AsyncSession = Depends(get_session)):
+    # DB connectivity
+    try:
+        await session.execute(text("SELECT 1"))
+        db_connected = True
+    except Exception:
+        db_connected = False
+
+    if not db_connected:
+        return {"healthy": False, "db_connected": False}
+
+    now = datetime.now(timezone.utc)
+    day_ago = now - timedelta(days=1)
+
+    # Listen events
+    total_events = (await session.execute(select(func.count(ListenEvent.id)))).scalar() or 0
+    events_24h = (
+        await session.execute(
+            select(func.count(ListenEvent.id)).where(ListenEvent.listened_at >= day_ago)
+        )
+    ).scalar() or 0
+    latest_event = (
+        await session.execute(select(func.max(ListenEvent.listened_at)))
+    ).scalar()
+
+    # Tracks
+    total_tracks = (await session.execute(select(func.count(Track.id)))).scalar() or 0
+    from_history = (
+        await session.execute(
+            select(func.count(func.distinct(ListenEvent.track_id)))
+        )
+    ).scalar() or 0
+    from_discovery = total_tracks - from_history
+
+    # Embeddings
+    def _embedding_count(status_val: str):
+        return select(func.count(Track.id)).where(Track.embedding_status == status_val)
+
+    emb_done = (await session.execute(_embedding_count("done"))).scalar() or 0
+    emb_pending = (await session.execute(_embedding_count("pending"))).scalar() or 0
+    emb_failed = (await session.execute(_embedding_count("failed"))).scalar() or 0
+    emb_no_preview = (await session.execute(_embedding_count("no_preview"))).scalar() or 0
+
+    # Taste profile
+    taste = (await session.execute(select(TasteProfile).where(TasteProfile.name == "default"))).scalar()
+
+    # Playlists
+    total_playlists = (await session.execute(select(func.count(Playlist.id)))).scalar() or 0
+    last_playlist = (await session.execute(select(func.max(Playlist.created_at)))).scalar()
+
+    return {
+        "healthy": db_connected,
+        "db_connected": db_connected,
+        "clap_model_loaded": is_model_loaded(),
+        "pipeline": {
+            "listen_events": {
+                "total": total_events,
+                "last_24h": events_24h,
+                "latest": latest_event.isoformat() if latest_event else None,
+            },
+            "tracks": {
+                "total": total_tracks,
+                "from_history": from_history,
+                "from_discovery": from_discovery,
+            },
+            "embeddings": {
+                "done": emb_done,
+                "pending": emb_pending,
+                "failed": emb_failed,
+                "no_preview": emb_no_preview,
+                "worker_running": is_worker_running(),
+                "worker_last_processed": worker_last_processed().isoformat() if worker_last_processed() else None,
+            },
+            "taste_profile": {
+                "exists": taste is not None,
+                "track_count": taste.track_count if taste else 0,
+                "updated_at": taste.updated_at.isoformat() if taste else None,
+            },
+            "playlists": {
+                "total_generated": total_playlists,
+                "last_generated": last_playlist.isoformat() if last_playlist else None,
+            },
+        },
+        "dependencies": {
+            "lastfm_api": "configured" if settings.lastfm_api_key else "not_configured",
+            "itunes_api": "ok",  # no auth needed
+            "ha_reachable": bool(settings.ha_token),
+            "music_assistant_reachable": bool(settings.ha_token),
+        },
+    }
--- a/src/haunt_fm/api/status_page.py
+++ b/src/haunt_fm/api/status_page.py
@@ -0,0 +1,23 @@
+from datetime import datetime, timezone
+from pathlib import Path
+
+from fastapi import APIRouter, Depends, Request
+from fastapi.responses import HTMLResponse
+from jinja2 import Environment, FileSystemLoader
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from haunt_fm.api.status import status as get_status_data
+from haunt_fm.db import get_session
+
+router = APIRouter()
+
+_template_dir = Path(__file__).parent.parent / "templates"
+_jinja_env = Environment(loader=FileSystemLoader(str(_template_dir)), autoescape=True)
+
+
+@router.get("/", response_class=HTMLResponse)
+async def status_page(request: Request, session: AsyncSession = Depends(get_session)):
+    data = await get_status_data(session)
+    template = _jinja_env.get_template("status.html")
+    html = template.render(data=data, now=datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M UTC"))
+    return HTMLResponse(html)
--- a/src/haunt_fm/config.py
+++ b/src/haunt_fm/config.py
@@ -0,0 +1,27 @@
+from pydantic_settings import BaseSettings
+
+
+class Settings(BaseSettings):
+    model_config = {"env_prefix": "HAUNTFM_"}
+
+    # Database
+    database_url: str = "postgresql+asyncpg://hauntfm:changeme@localhost:5432/hauntfm"
+
+    # Last.fm
+    lastfm_api_key: str = ""
+
+    # Home Assistant
+    ha_url: str = "http://192.168.86.51:8123"
+    ha_token: str = ""
+
+    # CLAP model
+    model_cache_dir: str = "/app/model-cache"
+    audio_cache_dir: str = "/app/audio-cache"
+
+    # Embedding worker
+    embedding_worker_enabled: bool = True
+    embedding_batch_size: int = 10
+    embedding_interval_seconds: int = 30
+
+
+settings = Settings()
--- a/src/haunt_fm/db.py
+++ b/src/haunt_fm/db.py
@@ -0,0 +1,13 @@
+from collections.abc import AsyncGenerator
+
+from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine
+
+from haunt_fm.config import settings
+
+engine = create_async_engine(settings.database_url, echo=False)
+async_session = async_sessionmaker(engine, class_=AsyncSession, expire_on_commit=False)
+
+
+async def get_session() -> AsyncGenerator[AsyncSession, None]:
+    async with async_session() as session:
+        yield session
--- a/src/haunt_fm/main.py
+++ b/src/haunt_fm/main.py
@@ -0,0 +1,50 @@
+import asyncio
+import logging
+from contextlib import asynccontextmanager
+
+from fastapi import FastAPI
+
+from haunt_fm.api import admin, health, history, playlists, recommendations, status, status_page
+from haunt_fm.config import settings
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s %(levelname)s %(name)s: %(message)s",
+)
+logger = logging.getLogger(__name__)
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    # Startup
+    logger.info("haunt-fm starting up")
+
+    # Start embedding worker in background
+    worker_task = None
+    if settings.embedding_worker_enabled:
+        from haunt_fm.services.embedding_worker import run_worker
+
+        worker_task = asyncio.create_task(run_worker())
+        logger.info("Embedding worker task created")
+
+    yield
+
+    # Shutdown
+    if worker_task:
+        worker_task.cancel()
+        try:
+            await worker_task
+        except asyncio.CancelledError:
+            pass
+    logger.info("haunt-fm shut down")
+
+
+app = FastAPI(title="haunt-fm", lifespan=lifespan)
+
+app.include_router(health.router)
+app.include_router(status.router)
+app.include_router(status_page.router)
+app.include_router(history.router)
+app.include_router(admin.router)
+app.include_router(recommendations.router)
+app.include_router(playlists.router)
--- a/src/haunt_fm/models/init.py
+++ b/src/haunt_fm/models/init.py
--- a/src/haunt_fm/models/base.py
+++ b/src/haunt_fm/models/base.py
@@ -0,0 +1,5 @@
+from sqlalchemy.orm import DeclarativeBase
+
+
+class Base(DeclarativeBase):
+    pass
--- a/src/haunt_fm/models/track.py
+++ b/src/haunt_fm/models/track.py
@@ -0,0 +1,113 @@
+from datetime import datetime
+
+from pgvector.sqlalchemy import Vector
+from sqlalchemy import BigInteger, DateTime, Index, Integer, Real, Text, func
+from sqlalchemy.dialects.postgresql import JSONB
+from sqlalchemy.orm import Mapped, mapped_column, relationship
+
+from haunt_fm.models.base import Base
+
+
+class Track(Base):
+    __tablename__ = "tracks"
+
+    id: Mapped[int] = mapped_column(BigInteger, primary_key=True)
+    title: Mapped[str] = mapped_column(Text, nullable=False)
+    artist: Mapped[str] = mapped_column(Text, nullable=False)
+    album: Mapped[str | None] = mapped_column(Text)
+    fingerprint: Mapped[str] = mapped_column(Text, unique=True, nullable=False)
+    lastfm_url: Mapped[str | None] = mapped_column(Text)
+    itunes_track_id: Mapped[int | None] = mapped_column(BigInteger)
+    itunes_preview_url: Mapped[str | None] = mapped_column(Text)
+    apple_music_id: Mapped[str | None] = mapped_column(Text)
+    duration_ms: Mapped[int | None] = mapped_column(Integer)
+    genre: Mapped[str | None] = mapped_column(Text)
+    embedding_status: Mapped[str] = mapped_column(Text, nullable=False, default="pending")
+    embedding_error: Mapped[str | None] = mapped_column(Text)
+    created_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), server_default=func.now())
+    updated_at: Mapped[datetime] = mapped_column(
+        DateTime(timezone=True), server_default=func.now(), onupdate=func.now()
+    )
+
+    listen_events: Mapped[list["ListenEvent"]] = relationship(back_populates="track")
+    embedding: Mapped["TrackEmbedding | None"] = relationship(back_populates="track")
+
+
+class ListenEvent(Base):
+    __tablename__ = "listen_events"
+
+    id: Mapped[int] = mapped_column(BigInteger, primary_key=True)
+    track_id: Mapped[int] = mapped_column(BigInteger, nullable=False)
+    source: Mapped[str] = mapped_column(Text, nullable=False, default="music_assistant")
+    speaker_name: Mapped[str | None] = mapped_column(Text)
+    listened_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), server_default=func.now())
+    duration_played: Mapped[int | None] = mapped_column(Integer)
+    raw_payload: Mapped[dict | None] = mapped_column(JSONB)
+
+    track: Mapped[Track] = relationship(back_populates="listen_events")
+
+
+class TrackEmbedding(Base):
+    __tablename__ = "track_embeddings"
+
+    id: Mapped[int] = mapped_column(BigInteger, primary_key=True)
+    track_id: Mapped[int] = mapped_column(BigInteger, unique=True, nullable=False)
+    embedding = mapped_column(Vector(512), nullable=False)
+    model_version: Mapped[str] = mapped_column(Text, nullable=False, default="laion/larger_clap_music")
+    created_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), server_default=func.now())
+
+    __table_args__ = (
+        Index("ix_track_embeddings_embedding_hnsw", "embedding", postgresql_using="hnsw", postgresql_with={"m": 16, "ef_construction": 64}, postgresql_ops={"embedding": "vector_cosine_ops"}),
+    )
+
+    track: Mapped[Track] = relationship(back_populates="embedding")
+
+
+class SimilarityLink(Base):
+    __tablename__ = "similarity_links"
+
+    id: Mapped[int] = mapped_column(BigInteger, primary_key=True)
+    source_track_id: Mapped[int] = mapped_column(BigInteger, nullable=False)
+    target_track_id: Mapped[int] = mapped_column(BigInteger, nullable=False)
+    lastfm_match: Mapped[float | None] = mapped_column(Real)
+    created_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), server_default=func.now())
+
+    __table_args__ = (
+        Index("uq_similarity_link", "source_track_id", "target_track_id", unique=True),
+    )
+
+
+class TasteProfile(Base):
+    __tablename__ = "taste_profiles"
+
+    id: Mapped[int] = mapped_column(BigInteger, primary_key=True)
+    name: Mapped[str] = mapped_column(Text, unique=True, nullable=False, default="default")
+    embedding = mapped_column(Vector(512), nullable=False)
+    track_count: Mapped[int] = mapped_column(Integer, nullable=False)
+    updated_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), server_default=func.now())
+
+
+class Playlist(Base):
+    __tablename__ = "playlists"
+
+    id: Mapped[int] = mapped_column(BigInteger, primary_key=True)
+    name: Mapped[str | None] = mapped_column(Text)
+    known_pct: Mapped[int] = mapped_column(Integer, nullable=False)
+    total_tracks: Mapped[int] = mapped_column(Integer, nullable=False)
+    created_at: Mapped[datetime] = mapped_column(DateTime(timezone=True), server_default=func.now())
+
+    tracks: Mapped[list["PlaylistTrack"]] = relationship(back_populates="playlist", cascade="all, delete-orphan")
+
+
+class PlaylistTrack(Base):
+    __tablename__ = "playlist_tracks"
+
+    id: Mapped[int] = mapped_column(BigInteger, primary_key=True)
+    playlist_id: Mapped[int] = mapped_column(BigInteger, nullable=False)
+    track_id: Mapped[int] = mapped_column(BigInteger, nullable=False)
+    position: Mapped[int] = mapped_column(Integer, nullable=False)
+    is_known: Mapped[bool] = mapped_column(nullable=False)
+    similarity_score: Mapped[float | None] = mapped_column(Real)
+
+    playlist: Mapped[Playlist] = relationship(back_populates="tracks")
+    track: Mapped[Track] = relationship()
--- a/src/haunt_fm/services/init.py
+++ b/src/haunt_fm/services/init.py
--- a/src/haunt_fm/services/embedding.py
+++ b/src/haunt_fm/services/embedding.py
@@ -0,0 +1,50 @@
+import logging
+from pathlib import Path
+
+import numpy as np
+
+from haunt_fm.config import settings
+
+logger = logging.getLogger(__name__)
+
+_model = None
+_processor = None
+
+
+def load_model():
+    """Load the CLAP model. Call once at startup."""
+    global _model, _processor
+    if _model is not None:
+        return
+
+    import torch
+    from transformers import ClapModel, ClapProcessor
+
+    logger.info("Loading CLAP model laion/larger_clap_music...")
+    cache_dir = settings.model_cache_dir
+
+    _processor = ClapProcessor.from_pretrained("laion/larger_clap_music", cache_dir=cache_dir)
+    _model = ClapModel.from_pretrained("laion/larger_clap_music", cache_dir=cache_dir)
+    _model.eval()
+    logger.info("CLAP model loaded successfully")
+
+
+def is_model_loaded() -> bool:
+    return _model is not None
+
+
+def embed_audio(audio: np.ndarray, sample_rate: int = 48000) -> np.ndarray:
+    """Generate a 512-dim embedding from audio waveform."""
+    import torch
+
+    if _model is None or _processor is None:
+        raise RuntimeError("CLAP model not loaded. Call load_model() first.")
+
+    inputs = _processor(audios=audio, sampling_rate=sample_rate, return_tensors="pt")
+    with torch.no_grad():
+        embeddings = _model.get_audio_features(**inputs)
+
+    # Normalize
+    emb = embeddings[0].numpy()
+    emb = emb / np.linalg.norm(emb)
+    return emb
--- a/src/haunt_fm/services/embedding_worker.py
+++ b/src/haunt_fm/services/embedding_worker.py
@@ -0,0 +1,140 @@
+import asyncio
+import logging
+from datetime import datetime, timezone
+
+from sqlalchemy import select, update
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from haunt_fm.config import settings
+from haunt_fm.db import async_session
+from haunt_fm.models.track import Track, TrackEmbedding
+from haunt_fm.services.embedding import embed_audio, load_model
+from haunt_fm.services.itunes_client import search_track
+from haunt_fm.utils.audio import decode_audio, download_preview
+
+logger = logging.getLogger(__name__)
+
+_running = False
+_last_processed: datetime | None = None
+
+
+def is_running() -> bool:
+    return _running
+
+
+def last_processed() -> datetime | None:
+    return _last_processed
+
+
+async def _process_track(session: AsyncSession, track: Track) -> bool:
+    """Process a single track: find preview, download, embed, store. Returns True on success."""
+    global _last_processed
+
+    # Mark as downloading
+    await session.execute(
+        update(Track).where(Track.id == track.id).values(embedding_status="downloading")
+    )
+    await session.commit()
+
+    # Find iTunes preview
+    if not track.itunes_preview_url:
+        result = await search_track(track.artist, track.title)
+        if result is None:
+            await session.execute(
+                update(Track).where(Track.id == track.id).values(embedding_status="no_preview")
+            )
+            await session.commit()
+            return False
+
+        await session.execute(
+            update(Track)
+            .where(Track.id == track.id)
+            .values(
+                itunes_track_id=result["track_id"],
+                itunes_preview_url=result["preview_url"],
+                apple_music_id=result["apple_music_id"],
+                duration_ms=result.get("duration_ms"),
+                genre=result.get("genre"),
+            )
+        )
+        await session.commit()
+        preview_url = result["preview_url"]
+    else:
+        preview_url = track.itunes_preview_url
+
+    # Download and decode
+    filename = f"{track.id}.m4a"
+    filepath = await download_preview(preview_url, settings.audio_cache_dir, filename)
+    audio = decode_audio(filepath)
+
+    # Embed
+    embedding = embed_audio(audio)
+
+    # Store
+    track_embedding = TrackEmbedding(
+        track_id=track.id,
+        embedding=embedding.tolist(),
+    )
+    session.add(track_embedding)
+    await session.execute(
+        update(Track).where(Track.id == track.id).values(embedding_status="done")
+    )
+    await session.commit()
+
+    _last_processed = datetime.now(timezone.utc)
+    return True
+
+
+async def run_worker():
+    """Background loop that processes unembedded tracks."""
+    global _running
+
+    if not settings.embedding_worker_enabled:
+        logger.info("Embedding worker disabled")
+        return
+
+    # Load model on first run
+    load_model()
+    _running = True
+    logger.info("Embedding worker started")
+
+    try:
+        while True:
+            try:
+                async with async_session() as session:
+                    # Find batch of pending tracks
+                    result = await session.execute(
+                        select(Track)
+                        .where(Track.embedding_status == "pending")
+                        .order_by(Track.created_at)
+                        .limit(settings.embedding_batch_size)
+                    )
+                    tracks = result.scalars().all()
+
+                    if not tracks:
+                        await asyncio.sleep(settings.embedding_interval_seconds)
+                        continue
+
+                    for track in tracks:
+                        try:
+                            await _process_track(session, track)
+                            logger.info("Embedded: %s - %s", track.artist, track.title)
+                        except Exception:
+                            logger.exception("Failed to embed %s - %s", track.artist, track.title)
+                            await session.execute(
+                                update(Track)
+                                .where(Track.id == track.id)
+                                .values(
+                                    embedding_status="failed",
+                                    embedding_error=str(Exception),
+                                )
+                            )
+                            await session.commit()
+
+            except Exception:
+                logger.exception("Embedding worker error")
+                await asyncio.sleep(10)
+
+            await asyncio.sleep(1)  # Brief pause between batches
+    finally:
+        _running = False
--- a/src/haunt_fm/services/history_ingest.py
+++ b/src/haunt_fm/services/history_ingest.py
@@ -0,0 +1,64 @@
+from datetime import datetime
+
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from haunt_fm.models.track import ListenEvent, Track
+
+
+def make_fingerprint(artist: str, title: str) -> str:
+    return f"{artist.lower().strip()}::{title.lower().strip()}"
+
+
+async def upsert_track(
+    session: AsyncSession,
+    title: str,
+    artist: str,
+    album: str | None = None,
+) -> Track:
+    fingerprint = make_fingerprint(artist, title)
+    result = await session.execute(select(Track).where(Track.fingerprint == fingerprint))
+    track = result.scalar_one_or_none()
+
+    if track is None:
+        track = Track(
+            title=title,
+            artist=artist,
+            album=album,
+            fingerprint=fingerprint,
+        )
+        session.add(track)
+        await session.flush()
+
+    if album and not track.album:
+        track.album = album
+        await session.flush()
+
+    return track
+
+
+async def ingest_listen_event(
+    session: AsyncSession,
+    title: str,
+    artist: str,
+    album: str | None,
+    speaker_name: str | None,
+    duration_played: int | None,
+    source: str,
+    listened_at: datetime,
+    raw_payload: dict | None = None,
+) -> ListenEvent:
+    track = await upsert_track(session, title, artist, album)
+
+    event = ListenEvent(
+        track_id=track.id,
+        source=source,
+        speaker_name=speaker_name,
+        listened_at=listened_at,
+        duration_played=duration_played,
+        raw_payload=raw_payload,
+    )
+    session.add(event)
+    await session.commit()
+    await session.refresh(event)
+    return event
--- a/src/haunt_fm/services/itunes_client.py
+++ b/src/haunt_fm/services/itunes_client.py
@@ -0,0 +1,57 @@
+import asyncio
+import logging
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+ITUNES_SEARCH_URL = "https://itunes.apple.com/search"
+
+# Rate limit: ~20 req/min for iTunes
+_last_request_time = 0.0
+_min_interval = 3.0  # 3s between requests
+
+
+async def _rate_limit():
+    global _last_request_time
+    now = asyncio.get_event_loop().time()
+    elapsed = now - _last_request_time
+    if elapsed < _min_interval:
+        await asyncio.sleep(_min_interval - elapsed)
+    _last_request_time = asyncio.get_event_loop().time()
+
+
+async def search_track(artist: str, title: str) -> dict | None:
+    """Search iTunes for a track and return preview info, or None if not found."""
+    await _rate_limit()
+
+    query = f"{artist} {title}"
+    async with httpx.AsyncClient(timeout=10) as client:
+        resp = await client.get(
+            ITUNES_SEARCH_URL,
+            params={
+                "term": query,
+                "media": "music",
+                "entity": "song",
+                "limit": 5,
+            },
+        )
+        resp.raise_for_status()
+        data = resp.json()
+
+    results = data.get("results", [])
+    if not results:
+        return None
+
+    # Find best match (simple: first result with a preview URL)
+    for r in results:
+        if r.get("previewUrl"):
+            return {
+                "track_id": r["trackId"],
+                "preview_url": r["previewUrl"],
+                "apple_music_id": str(r.get("trackId", "")),
+                "duration_ms": r.get("trackTimeMillis"),
+                "genre": r.get("primaryGenreName"),
+            }
+
+    return None
--- a/src/haunt_fm/services/lastfm_client.py
+++ b/src/haunt_fm/services/lastfm_client.py
@@ -0,0 +1,106 @@
+import asyncio
+import logging
+
+import httpx
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from haunt_fm.config import settings
+from haunt_fm.models.track import SimilarityLink, Track
+from haunt_fm.services.history_ingest import make_fingerprint
+
+logger = logging.getLogger(__name__)
+
+LASTFM_API_URL = "https://ws.audioscrobbler.com/2.0/"
+
+# Simple rate limiter: max 5 req/s
+_last_request_time = 0.0
+_min_interval = 0.2  # 200ms between requests
+
+
+async def _rate_limit():
+    global _last_request_time
+    now = asyncio.get_event_loop().time()
+    elapsed = now - _last_request_time
+    if elapsed < _min_interval:
+        await asyncio.sleep(_min_interval - elapsed)
+    _last_request_time = asyncio.get_event_loop().time()
+
+
+async def get_similar_tracks(artist: str, title: str, limit: int = 50) -> list[dict]:
+    """Call Last.fm track.getSimilar and return list of similar track dicts."""
+    if not settings.lastfm_api_key:
+        raise ValueError("HAUNTFM_LASTFM_API_KEY not configured")
+
+    await _rate_limit()
+
+    async with httpx.AsyncClient(timeout=10) as client:
+        resp = await client.get(
+            LASTFM_API_URL,
+            params={
+                "method": "track.getSimilar",
+                "artist": artist,
+                "track": title,
+                "api_key": settings.lastfm_api_key,
+                "format": "json",
+                "limit": limit,
+            },
+        )
+        resp.raise_for_status()
+        data = resp.json()
+
+    similar = data.get("similartracks", {}).get("track", [])
+    return [
+        {
+            "title": t["name"],
+            "artist": t["artist"]["name"],
+            "match": float(t.get("match", 0)),
+            "url": t.get("url"),
+        }
+        for t in similar
+    ]
+
+
+async def discover_similar_for_track(session: AsyncSession, source_track: Track) -> int:
+    """Discover similar tracks for a given track, upsert them, and create similarity links."""
+    try:
+        similar = await get_similar_tracks(source_track.artist, source_track.title)
+    except Exception:
+        logger.exception("Failed to get similar tracks for %s - %s", source_track.artist, source_track.title)
+        raise
+
+    count = 0
+    for item in similar:
+        fingerprint = make_fingerprint(item["artist"], item["title"])
+
+        # Upsert target track
+        result = await session.execute(select(Track).where(Track.fingerprint == fingerprint))
+        target = result.scalar_one_or_none()
+        if target is None:
+            target = Track(
+                title=item["title"],
+                artist=item["artist"],
+                fingerprint=fingerprint,
+                lastfm_url=item.get("url"),
+            )
+            session.add(target)
+            await session.flush()
+
+        # Upsert similarity link
+        existing_link = await session.execute(
+            select(SimilarityLink).where(
+                SimilarityLink.source_track_id == source_track.id,
+                SimilarityLink.target_track_id == target.id,
+            )
+        )
+        if existing_link.scalar_one_or_none() is None:
+            link = SimilarityLink(
+                source_track_id=source_track.id,
+                target_track_id=target.id,
+                lastfm_match=item["match"],
+            )
+            session.add(link)
+            count += 1
+
+    await session.commit()
+    return count
--- a/src/haunt_fm/services/music_assistant.py
+++ b/src/haunt_fm/services/music_assistant.py
@@ -0,0 +1,117 @@
+import logging
+
+import httpx
+
+from haunt_fm.config import settings
+
+logger = logging.getLogger(__name__)
+
+
+async def _ha_request(method: str, path: str, **kwargs) -> dict:
+    """Make an authenticated request to Home Assistant REST API."""
+    headers = {
+        "Authorization": f"Bearer {settings.ha_token}",
+        "Content-Type": "application/json",
+    }
+    async with httpx.AsyncClient(timeout=10) as client:
+        resp = await client.request(
+            method, f"{settings.ha_url}{path}", headers=headers, **kwargs
+        )
+        resp.raise_for_status()
+        if resp.content:
+            return resp.json()
+        return {}
+
+
+async def is_ha_reachable() -> bool:
+    """Check if Home Assistant is reachable."""
+    try:
+        await _ha_request("GET", "/api/")
+        return True
+    except Exception:
+        return False
+
+
+async def play_media_on_speaker(
+    media_content_id: str,
+    speaker_entity: str,
+    media_content_type: str = "music",
+) -> None:
+    """Play a media item on a speaker via HA media_player service."""
+    await _ha_request(
+        "POST",
+        "/api/services/media_player/play_media",
+        json={
+            "entity_id": speaker_entity,
+            "media_content_id": media_content_id,
+            "media_content_type": media_content_type,
+        },
+    )
+    logger.info("Playing %s on %s", media_content_id, speaker_entity)
+
+
+async def search_and_play(
+    artist: str,
+    title: str,
+    speaker_entity: str,
+) -> bool:
+    """Search Music Assistant for a track and play it.
+
+    Uses the mass.search service to find the track, then plays it.
+    """
+    try:
+        # Use Music Assistant search via HA
+        result = await _ha_request(
+            "POST",
+            "/api/services/mass/search",
+            json={
+                "name": f"{artist} {title}",
+                "media_type": "track",
+                "limit": 1,
+            },
+        )
+        logger.info("MA search result for '%s - %s': %s", artist, title, result)
+        return True
+    except Exception:
+        logger.exception("Failed to search MA for %s - %s", artist, title)
+        return False
+
+
+async def play_playlist_on_speaker(
+    tracks: list[dict],
+    speaker_entity: str,
+) -> None:
+    """Play a list of tracks on a speaker. Each track dict has 'artist' and 'title'.
+
+    Enqueues tracks via Music Assistant.
+    """
+    if not tracks:
+        return
+
+    for i, track in enumerate(tracks):
+        try:
+            if i == 0:
+                # Play first track
+                await _ha_request(
+                    "POST",
+                    "/api/services/media_player/play_media",
+                    json={
+                        "entity_id": speaker_entity,
+                        "media_content_id": f"{track['artist']} - {track['title']}",
+                        "media_content_type": "music",
+                    },
+                )
+            else:
+                # Enqueue subsequent tracks
+                await _ha_request(
+                    "POST",
+                    "/api/services/media_player/play_media",
+                    json={
+                        "entity_id": speaker_entity,
+                        "media_content_id": f"{track['artist']} - {track['title']}",
+                        "media_content_type": "music",
+                        "enqueue": "add",
+                    },
+                )
+        except Exception:
+            logger.exception("Failed to enqueue %s - %s", track["artist"], track["title"])
--- a/src/haunt_fm/services/playlist_generator.py
+++ b/src/haunt_fm/services/playlist_generator.py
@@ -0,0 +1,106 @@
+import logging
+import random
+
+from sqlalchemy import func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from haunt_fm.models.track import (
+    ListenEvent,
+    Playlist,
+    PlaylistTrack,
+    Track,
+)
+from haunt_fm.services.recommender import get_recommendations
+
+logger = logging.getLogger(__name__)
+
+
+async def generate_playlist(
+    session: AsyncSession,
+    total_tracks: int = 20,
+    known_pct: int = 30,
+    name: str | None = None,
+) -> Playlist:
+    """Generate a playlist mixing known-liked tracks with new recommendations.
+
+    Args:
+        total_tracks: Total number of tracks in the playlist
+        known_pct: Percentage of tracks from listening history (0-100)
+        name: Optional playlist name
+    """
+    known_count = round(total_tracks * known_pct / 100)
+    new_count = total_tracks - known_count
+
+    # Get known-liked tracks (most played, with some randomness)
+    known_result = await session.execute(
+        select(Track.id)
+        .join(ListenEvent, ListenEvent.track_id == Track.id)
+        .group_by(Track.id)
+        .order_by(func.count(ListenEvent.id).desc())
+        .limit(known_count * 3)  # oversample for randomness
+    )
+    known_pool = [row[0] for row in known_result]
+    if len(known_pool) > known_count:
+        known_ids = random.sample(known_pool, known_count)
+    else:
+        known_ids = known_pool
+        known_count = len(known_ids)
+        new_count = total_tracks - known_count
+
+    # Get new recommendations
+    recs = await get_recommendations(session, limit=new_count * 2, exclude_known=True)
+    new_tracks = [(r["track_id"], r["similarity"]) for r in recs[:new_count]]
+
+    # Interleave: spread known tracks throughout the playlist
+    playlist_items: list[tuple[int, bool, float | None]] = []
+    for tid in known_ids:
+        playlist_items.append((tid, True, None))
+    for tid, sim in new_tracks:
+        playlist_items.append((tid, False, sim))
+
+    # Shuffle but keep some structure — interleave known/new
+    known_items = [(tid, True, None) for tid in known_ids]
+    new_items = [(tid, False, sim) for tid, sim in new_tracks]
+    random.shuffle(known_items)
+    random.shuffle(new_items)
+
+    # Interleave
+    interleaved = []
+    ki, ni = 0, 0
+    for i in range(len(known_items) + len(new_items)):
+        # Roughly distribute known tracks evenly
+        if ki < len(known_items) and (ni >= len(new_items) or i % max(1, total_tracks // max(1, known_count)) == 0):
+            interleaved.append(known_items[ki])
+            ki += 1
+        elif ni < len(new_items):
+            interleaved.append(new_items[ni])
+            ni += 1
+        elif ki < len(known_items):
+            interleaved.append(known_items[ki])
+            ki += 1
+
+    # Create playlist record
+    playlist = Playlist(
+        name=name or f"haunt-fm mix ({len(interleaved)} tracks)",
+        known_pct=known_pct,
+        total_tracks=len(interleaved),
+    )
+    session.add(playlist)
+    await session.flush()
+
+    # Create playlist tracks
+    for pos, (track_id, is_known, similarity) in enumerate(interleaved):
+        pt = PlaylistTrack(
+            playlist_id=playlist.id,
+            track_id=track_id,
+            position=pos,
+            is_known=is_known,
+            similarity_score=similarity,
+        )
+        session.add(pt)
+
+    await session.commit()
+    await session.refresh(playlist)
+    logger.info("Generated playlist '%s' with %d tracks (%d known, %d new)",
+                playlist.name, len(interleaved), known_count, len(new_items))
+    return playlist
--- a/src/haunt_fm/services/recommender.py
+++ b/src/haunt_fm/services/recommender.py
@@ -0,0 +1,71 @@
+import logging
+
+from sqlalchemy import select, text
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from haunt_fm.models.track import (
+    ListenEvent,
+    TasteProfile,
+    Track,
+    TrackEmbedding,
+)
+
+logger = logging.getLogger(__name__)
+
+
+async def get_recommendations(
+    session: AsyncSession,
+    limit: int = 50,
+    exclude_known: bool = True,
+    profile_name: str = "default",
+) -> list[dict]:
+    """Get track recommendations ranked by cosine similarity to taste profile."""
+    # Load taste profile
+    profile = (
+        await session.execute(select(TasteProfile).where(TasteProfile.name == profile_name))
+    ).scalar_one_or_none()
+
+    if profile is None:
+        return []
+
+    # Use pgvector cosine distance operator (<=>)
+    # Lower distance = more similar
+    if exclude_known:
+        # Subquery: track IDs that have listen events
+        known_ids_subq = select(ListenEvent.track_id).distinct().subquery()
+
+        query = text("""
+            SELECT t.id, t.title, t.artist, t.album, t.genre,
+                   1 - (te.embedding <=> :profile_embedding) AS similarity
+            FROM track_embeddings te
+            JOIN tracks t ON t.id = te.track_id
+            WHERE te.track_id NOT IN (SELECT track_id FROM listen_events)
+            ORDER BY te.embedding <=> :profile_embedding
+            LIMIT :limit
+        """)
+    else:
+        query = text("""
+            SELECT t.id, t.title, t.artist, t.album, t.genre,
+                   1 - (te.embedding <=> :profile_embedding) AS similarity
+            FROM track_embeddings te
+            JOIN tracks t ON t.id = te.track_id
+            ORDER BY te.embedding <=> :profile_embedding
+            LIMIT :limit
+        """)
+
+    result = await session.execute(
+        query,
+        {"profile_embedding": str(profile.embedding), "limit": limit},
+    )
+
+    return [
+        {
+            "track_id": row.id,
+            "title": row.title,
+            "artist": row.artist,
+            "album": row.album,
+            "genre": row.genre,
+            "similarity": round(float(row.similarity), 4),
+        }
+        for row in result
+    ]
--- a/src/haunt_fm/services/taste_profile.py
+++ b/src/haunt_fm/services/taste_profile.py
@@ -0,0 +1,88 @@
+import logging
+from datetime import datetime, timezone
+
+import numpy as np
+from sqlalchemy import func, select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from haunt_fm.models.track import (
+    ListenEvent,
+    TasteProfile,
+    Track,
+    TrackEmbedding,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def _recency_weight(listened_at: datetime, now: datetime, half_life_days: float = 30.0) -> float:
+    """Exponential decay weight based on recency. Half-life = 30 days by default."""
+    age_days = (now - listened_at).total_seconds() / 86400
+    return 2 ** (-age_days / half_life_days)
+
+
+async def build_taste_profile(session: AsyncSession, name: str = "default") -> TasteProfile | None:
+    """Build a taste profile as the weighted average of listened-track embeddings.
+
+    Weights: play_count * recency_decay for each track.
+    """
+    # Get all listened tracks with embeddings
+    result = await session.execute(
+        select(
+            Track.id,
+            TrackEmbedding.embedding,
+            func.count(ListenEvent.id).label("play_count"),
+            func.max(ListenEvent.listened_at).label("last_played"),
+        )
+        .join(TrackEmbedding, TrackEmbedding.track_id == Track.id)
+        .join(ListenEvent, ListenEvent.track_id == Track.id)
+        .group_by(Track.id, TrackEmbedding.embedding)
+    )
+    rows = result.all()
+
+    if not rows:
+        logger.warning("No listened tracks with embeddings found")
+        return None
+
+    now = datetime.now(timezone.utc)
+    embeddings = []
+    weights = []
+
+    for row in rows:
+        emb = np.array(row.embedding, dtype=np.float32)
+        play_count = row.play_count
+        recency = _recency_weight(row.last_played, now)
+        weight = play_count * recency
+        embeddings.append(emb)
+        weights.append(weight)
+
+    embeddings_arr = np.array(embeddings)
+    weights_arr = np.array(weights, dtype=np.float32)
+    weights_arr /= weights_arr.sum()
+
+    # Weighted average
+    profile_emb = (embeddings_arr * weights_arr[:, np.newaxis]).sum(axis=0)
+    profile_emb = profile_emb / np.linalg.norm(profile_emb)
+
+    # Upsert
+    existing = (
+        await session.execute(select(TasteProfile).where(TasteProfile.name == name))
+    ).scalar_one_or_none()
+
+    if existing:
+        existing.embedding = profile_emb.tolist()
+        existing.track_count = len(rows)
+        existing.updated_at = now
+    else:
+        existing = TasteProfile(
+            name=name,
+            embedding=profile_emb.tolist(),
+            track_count=len(rows),
+            updated_at=now,
+        )
+        session.add(existing)
+
+    await session.commit()
+    await session.refresh(existing)
+    logger.info("Built taste profile '%s' from %d tracks", name, len(rows))
+    return existing
--- a/src/haunt_fm/templates/status.html
+++ b/src/haunt_fm/templates/status.html
@@ -0,0 +1,140 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+    <meta charset="utf-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1">
+    <title>haunt-fm status</title>
+    <style>
+        * { margin: 0; padding: 0; box-sizing: border-box; }
+        body { font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, monospace; background: #0d1117; color: #c9d1d9; padding: 2rem; }
+        h1 { color: #58a6ff; margin-bottom: 1.5rem; font-size: 1.5rem; }
+        .status-badge { display: inline-block; padding: 0.25rem 0.75rem; border-radius: 1rem; font-size: 0.85rem; font-weight: 600; margin-bottom: 1.5rem; }
+        .status-badge.healthy { background: #238636; color: #fff; }
+        .status-badge.degraded { background: #da3633; color: #fff; }
+        .section { background: #161b22; border: 1px solid #30363d; border-radius: 0.5rem; padding: 1.25rem; margin-bottom: 1rem; }
+        .section h2 { color: #8b949e; font-size: 0.8rem; text-transform: uppercase; letter-spacing: 0.05em; margin-bottom: 0.75rem; }
+        .row { display: flex; justify-content: space-between; padding: 0.35rem 0; border-bottom: 1px solid #21262d; }
+        .row:last-child { border-bottom: none; }
+        .label { color: #8b949e; }
+        .value { color: #c9d1d9; font-weight: 500; }
+        .dot { display: inline-block; width: 8px; height: 8px; border-radius: 50%; margin-right: 0.4rem; vertical-align: middle; }
+        .dot.green { background: #3fb950; }
+        .dot.red { background: #f85149; }
+        .dot.yellow { background: #d29922; }
+        .dot.gray { background: #484f58; }
+        .timestamp { color: #484f58; font-size: 0.8rem; margin-top: 1.5rem; text-align: center; }
+    </style>
+</head>
+<body>
+    <h1>haunt-fm</h1>
+    <span class="status-badge {{ 'healthy' if data.healthy else 'degraded' }}">
+        {{ 'Healthy' if data.healthy else 'Degraded' }}
+    </span>
+
+    <div class="section">
+        <h2>Pipeline</h2>
+        <div class="row">
+            <span class="label"><span class="dot {{ 'green' if data.db_connected else 'red' }}"></span>Database</span>
+            <span class="value">{{ 'Connected' if data.db_connected else 'Down' }}</span>
+        </div>
+        <div class="row">
+            <span class="label"><span class="dot {{ 'green' if data.clap_model_loaded else 'gray' }}"></span>CLAP Model</span>
+            <span class="value">{{ 'Loaded' if data.clap_model_loaded else 'Not loaded' }}</span>
+        </div>
+    </div>
+
+    <div class="section">
+        <h2>Listening History</h2>
+        <div class="row">
+            <span class="label">Total events</span>
+            <span class="value">{{ data.pipeline.listen_events.total }}</span>
+        </div>
+        <div class="row">
+            <span class="label">Last 24h</span>
+            <span class="value">{{ data.pipeline.listen_events.last_24h }}</span>
+        </div>
+        <div class="row">
+            <span class="label">Latest</span>
+            <span class="value">{{ data.pipeline.listen_events.latest or 'Never' }}</span>
+        </div>
+    </div>
+
+    <div class="section">
+        <h2>Tracks</h2>
+        <div class="row">
+            <span class="label">Total</span>
+            <span class="value">{{ data.pipeline.tracks.total }}</span>
+        </div>
+        <div class="row">
+            <span class="label">From history</span>
+            <span class="value">{{ data.pipeline.tracks.from_history }}</span>
+        </div>
+        <div class="row">
+            <span class="label">From discovery</span>
+            <span class="value">{{ data.pipeline.tracks.from_discovery }}</span>
+        </div>
+    </div>
+
+    <div class="section">
+        <h2>Embeddings</h2>
+        <div class="row">
+            <span class="label"><span class="dot {{ 'green' if data.pipeline.embeddings.worker_running else 'gray' }}"></span>Worker</span>
+            <span class="value">{{ 'Running' if data.pipeline.embeddings.worker_running else 'Stopped' }}</span>
+        </div>
+        <div class="row">
+            <span class="label">Done</span>
+            <span class="value">{{ data.pipeline.embeddings.done }}</span>
+        </div>
+        <div class="row">
+            <span class="label">Pending</span>
+            <span class="value">{{ data.pipeline.embeddings.pending }}</span>
+        </div>
+        <div class="row">
+            <span class="label">Failed</span>
+            <span class="value">{{ data.pipeline.embeddings.failed }}</span>
+        </div>
+        <div class="row">
+            <span class="label">No preview</span>
+            <span class="value">{{ data.pipeline.embeddings.no_preview }}</span>
+        </div>
+    </div>
+
+    <div class="section">
+        <h2>Taste Profile</h2>
+        <div class="row">
+            <span class="label"><span class="dot {{ 'green' if data.pipeline.taste_profile.exists else 'gray' }}"></span>Profile</span>
+            <span class="value">{{ 'Built (' ~ data.pipeline.taste_profile.track_count ~ ' tracks)' if data.pipeline.taste_profile.exists else 'Not built' }}</span>
+        </div>
+    </div>
+
+    <div class="section">
+        <h2>Playlists</h2>
+        <div class="row">
+            <span class="label">Generated</span>
+            <span class="value">{{ data.pipeline.playlists.total_generated }}</span>
+        </div>
+        <div class="row">
+            <span class="label">Last generated</span>
+            <span class="value">{{ data.pipeline.playlists.last_generated or 'Never' }}</span>
+        </div>
+    </div>
+
+    <div class="section">
+        <h2>Dependencies</h2>
+        <div class="row">
+            <span class="label"><span class="dot {{ 'green' if data.dependencies.lastfm_api == 'ok' else 'gray' }}"></span>Last.fm API</span>
+            <span class="value">{{ data.dependencies.lastfm_api }}</span>
+        </div>
+        <div class="row">
+            <span class="label"><span class="dot {{ 'green' if data.dependencies.itunes_api == 'ok' else 'gray' }}"></span>iTunes API</span>
+            <span class="value">{{ data.dependencies.itunes_api }}</span>
+        </div>
+        <div class="row">
+            <span class="label"><span class="dot {{ 'green' if data.dependencies.ha_reachable else 'gray' }}"></span>Home Assistant</span>
+            <span class="value">{{ 'Reachable' if data.dependencies.ha_reachable else 'Unknown' }}</span>
+        </div>
+    </div>
+
+    <p class="timestamp">Updated {{ now }}</p>
+</body>
+</html>
--- a/src/haunt_fm/utils/init.py
+++ b/src/haunt_fm/utils/init.py
--- a/src/haunt_fm/utils/audio.py
+++ b/src/haunt_fm/utils/audio.py
@@ -0,0 +1,31 @@
+import logging
+from pathlib import Path
+
+import httpx
+import librosa
+import numpy as np
+
+logger = logging.getLogger(__name__)
+
+
+async def download_preview(url: str, cache_dir: str, filename: str) -> Path:
+    """Download an AAC preview file, return the local path."""
+    cache_path = Path(cache_dir)
+    cache_path.mkdir(parents=True, exist_ok=True)
+    filepath = cache_path / filename
+
+    if filepath.exists():
+        return filepath
+
+    async with httpx.AsyncClient(timeout=30) as client:
+        resp = await client.get(url)
+        resp.raise_for_status()
+        filepath.write_bytes(resp.content)
+
+    return filepath
+
+
+def decode_audio(filepath: Path, target_sr: int = 48000) -> np.ndarray:
+    """Decode audio file to numpy array at target sample rate. Uses librosa (ffmpeg backend)."""
+    audio, _ = librosa.load(str(filepath), sr=target_sr, mono=True)
+    return audio
--- a/src/haunt_fm/utils/rate_limiter.py
+++ b/src/haunt_fm/utils/rate_limiter.py
@@ -0,0 +1,25 @@
+import asyncio
+import time
+
+
+class AsyncRateLimiter:
+    """Simple async rate limiter using token bucket."""
+
+    def __init__(self, rate: float, per: float = 1.0):
+        """Args:
+            rate: Number of requests allowed per `per` seconds.
+            per: Time window in seconds.
+        """
+        self._rate = rate
+        self._per = per
+        self._min_interval = per / rate
+        self._last_time = 0.0
+        self._lock = asyncio.Lock()
+
+    async def acquire(self):
+        async with self._lock:
+            now = time.monotonic()
+            elapsed = now - self._last_time
+            if elapsed < self._min_interval:
+                await asyncio.sleep(self._min_interval - elapsed)
+            self._last_time = time.monotonic()