Deadend Deafchild

Deadend Deafchild is the artist site for an industrial-electronic music project. The 2026 build throws out the conventional artist-site shape — hero image, bio block, embedded Spotify iframe — and makes the player the entire site: a Winamp-inspired panel floating over a MilkDrop visualizer, with the full catalogue underneath it and a dock of platform links at the bottom.

The visual system

The background is Butterchurn, a WebGL reimplementation of MilkDrop, running a pool of 1,132 classic presets across two of the five decks in the visual mixer — the others carry the album artwork, cutaway stills, and silent per-track video loops, crossfaded rather than cut. It is lazy-loaded on first play — preset data that a visitor who never presses play never pays for. Preset changes are no longer on a timer: they land on downbeats every 4/8/16 bars, hard-cut on drops, and a section-aware signature effect ramps in on the build-up before a drop rather than reacting after it.

That is possible because the visualizer runs on precomputed analysis, not live guessing. An offline Python pipeline (ffmpeg → Essentia + librosa) analyses every master once — beat grid, tempo, key, section boundaries with lead-in times, multi-band envelopes, waveform peaks — and ships a small JSON per track that the runtime reads by audio.currentTime. The whole corpus takes about 14 minutes on a 4-core box; a track with no analysis file falls back to the original live-DSP path. A hidden, browser-local control room (no server-side admin surface) exposes live diagnostics, 84 tuning settings, the preset browser, and per-track preset pins, all persisted in localStorage.

The part worth stealing is the feedback loop: a Web Audio analyser drives a spectrum strip inside the player, and the UI samples its accent colors back out of the live visualizer. Play a track and the buttons, the seek fill, and the volume slider all drift toward whatever palette the current preset happens to be in. Four hand-tuned base themes sit underneath as a floor, so the page still reads as branded when the visualizer is dark.

One binary, no dependencies

The backend is Go with an empty dependency list — stdlib only, and the whole frontend (HTML, CSS, JS, fonts, icons, vendored visualizer) compiled in through embed.FS. Deploying is copying one file and restarting a systemd unit; there is no build step on the server, no node_modules, and nothing to drift out of sync. The server bundles and gzips CSS and JS at startup and sets its own security headers, which is most of what a framework would have been doing anyway.

A second small command walks the audio tree and generates the playlist manifest — durations, file sizes, formats, and the public storage URLs for both the streaming copy and the original master. Adding music is: drop files in, transcode, regenerate, redeploy.

The desktop build

Since August 2026 the same engine also ships as DEDC Studio, an Electron shell around the server binary, for the things a browser cannot do: a Spout sender and a clean capture window for Resolume and OBS, OSC and MIDI-learn control surfaces, a second-monitor console showing every deck and the next drop counting down, and the analysis pipeline frozen into a sidecar so it can analyse the user's own files locally. Every control is an automation lane: a performance is recorded as passes that layer on top of each other and replay to the millisecond. GitHub Actions builds and smoke-tests the Windows installer on each release tag, the installer is served from Google Cloud Storage, and the update feed is synthesised by the Go server from a committed release document, so a release the server does not know about cannot be announced. It is in private beta and not publicly downloadable — the latest tag is 0.5.1.

Audio delivery

Streams are ~256k AAC served from Google Cloud Storage with CORS and range requests enabled, so seeking is instant and the VPS never touches audio bandwidth. Every track also links its untouched master — 84 tracks across 8 albums, roughly 5¾ hours, free to download, mostly WAV, including eight tracks recovered from a Google Takeout archive via a cost-bounded, range-read recovery script. The playlist shows per-track artwork thumbnails. The MediaSession API is wired up, so the lock screen and Android Auto get real track metadata and transport controls instead of "unknown media."

What it replaced, and why

The previous version of this site was a hand-rolled PHP CMS with an admin panel, SQLite, visitor fingerprinting, and a system-info dashboard. It worked, but almost none of it was used: the catalogue changes a few times a year, and every edit still meant logging into a CMS to change text that could have been a constant. The rebuild deleted the admin panel, the database, and the auth layer outright and replaced them with a single config file. That trade — no runtime editing, in exchange for no attack surface and no moving parts — is the right one for a site whose content is a finite, slow-moving list.

One authenticated surface has since come back, for a job a config file cannot do: a passphrase-gated Release Desk at /admin that prepares tracks for distribution — release metadata, licensing, scheduling, export, and the inbox the contact form lands in. It is still file-backed; there is no database and no third-party Go dependency. The site also carries a public /about.html writeup of the stack, the roadmap, and a changelog generated straight from the repository — 197 entries at last count.

It also got faster by a lot. Lighthouse mobile lands at 99 performance, 100 accessibility, 100 best practices, 100 SEO — on a page whose whole premise is a continuously-rendering WebGL scene.