How I Built
TOP SELECTA

It started with a feeling

I was obsessed with Semantle and Wordle variants. But every music version I played was the same: guess the Taylor Swift album, guess the pop song. Nothing that spoke to where I came from or the music I actually loved.

So I built my own. A daily puzzle game where the artist pool is the entire African diaspora — Afrobeats, grime, jazz, soul, dancehall, hip-hop, highlife, Afro-Brazilian, drill, soca. Music of the diaspora. 5,000+ artists, one hidden every day, and you find them through musical proximity rather than exact naming.

The similarity engine — how it actually works

This is the core of the game. When you guess an artist, the engine computes how musically similar they are to the hidden answer and returns a rank from 1 to however many artists are in the pool.

The engine uses 7 dimensions, each weighted based on how much signal they carry:

The mood vector — the most interesting bit

Each artist has a 6-dimensional "mood vector" — a set of numbers representing how their music feels. Things like energy, darkness, danceability, spirituality. These aren't just tags — they're continuous values, so you can compute actual mathematical distance between artists using cosine similarity.

This is why the game feels "right" even when two artists share no genre tags. A FROZEN Afrobeats producer might still be musically adjacent to a UK soul vocalist if their mood vectors align.

// Cosine similarity between two 6-dimension mood vectors
function cosineSim(a, b) {
  const dot  = a.reduce((s, v, i) => s + v * b[i], 0);
  const magA = Math.sqrt(a.reduce((s, v) => s + v*v, 0));
  const magB = Math.sqrt(b.reduce((s, v) => s + v*v, 0));
  return dot / (magA * magB);
}

// Full similarity formula
function artistSim(a, b) {
  return cosineSim(a.mood, b.mood)         * 0.33
       + (a.cat === b.cat ? 1 : 0)         * 0.19
       + collab(a, b)                       * 0.16
       + jaccard(a.era, b.era)             * 0.13
       + (a.region === b.region ? 1 : 0)   * 0.08
       + jaccard(a.type, b.type)           * 0.06
       + jaccard(a.i, b.i)                * 0.05;
}

Temperature tiers

The similarity score gets translated into a temperature — a human-readable indicator that tells you whether you're burning hot or ice cold.

Temperature also determines how much detail the feedback gives you. A SCALDING guess might say "They've worked together directly. Both Atlanta trap — heavy 808s, melodic flow. Active 2010s–2020s." A FROZEN guess gives you one bare directional fact. Reveals skill, rewards curiosity.

Building a 5,000+ artist database from scratch

There's no existing dataset of Black music artists with the fields I needed. I had to build it. Here's the full pipeline — from generation to enrichment to deployment:

Coverage lessons from real API data

I went in assuming MusicBrainz would have everything. It doesn't. Some real coverage numbers that surprised me:

Field                MusicBrainz   Wikidata
─────────────────────────────────────────────
Genre tags           ~70%          —
Origin city/country  ~65%          —
Associated acts      ~50%          —
Instruments          ~40%          —
Record labels        ~5%           ~70% (P264)
Influences           ~5%           ~35% (P737)

The label relationship issue was a database design quirk — MusicBrainz stores "this artist founded this label" not "this artist was signed to this label." A rapper's label-rels return nothing unless they own their own label. Wikidata's P264 property has this right. Lesson: always test coverage with a sample before you architect around an API.

The overnight expansion

The initial database had 2,404 artists. To hit 5,000+ and improve representation of under-served regions, I set up an overnight scheduled task. By morning: 5,002 artists — 2,598 new additions across East Africa, Francophone West Africa, contemporary Afrobeats, UK drill, Caribbean dancehall, and Latin Black music.

Every feature — and why it's there

The rank title ladder — a small thing that landed well

After posting your score, you get a leaderboard rank and a rank title. Top 6 are celebratory. From #7 down they get progressively more insulting:

What the share card looks like

Every game produces a shareable text card. The emoji trail is the game's "spoiler-safe fingerprint" — you can see exactly how someone played without knowing the answer:

The > prefix makes WhatsApp render the squares as a blockquote — visually distinct from the rest of the message. Pipe separators keep short games readable. Longer games pack into rows of 10 automatically.

How it came together — the real timeline

This wasn't a straight line. Here's the actual sequence of what got built and what broke along the way:

The stack — deliberately boring

I didn't reach for a framework. The game is a single HTML file, some Vercel serverless functions, and Vercel KV for the leaderboard. That's it. No build step. No npm install. Deploy on push.

This is one corner of something larger

TOP SELECTA isn't a standalone project. It's one of several things I'm building under the banner of Black Music Multiverse — a connected world of interactive experiences celebrating the music, stories, and culture of the African diaspora.

Think: a 3D world where each district is a different experience. You walk into a tavern and it's a strategy game. You enter a shrine and hear an audio story set in an Afro-utopia. You find a crypt and it's an art cinema. TOP SELECTA sits in this world as the music discovery game you stumble into off the street.

Everything is free. No ads, no paywalls, no data collection. Just things I care about, made with care, shared with anyone who wants them.