🎵 Mello

Rank your music taste. Discover what you actually love.

Mello uses head-to-head song battles and real ML inference to build a mathematically precise map of your music taste — then uses it to recommend songs you'll love and blend your taste with friends.

---

The Problem

Spotify has 100 million songs and recommends the same 20 to everyone. It tracks what you play, not what you love. There's no way to say "I'm obsessed with this" versus "it was just on in the background."

Mello fixes that.

---

How It Works

1. Elo Ranking

We show you two songs. You pick one. Under the hood, we run an Elo rating system — the same algorithm used to rank chess grandmasters — to build a ranked preference map from your choices. Not listening history. Actual ranked preference data.

2. GPU Embedding via Modal

Every song you rank gets converted into an embedding vector — a mathematical fingerprint of its sonic identity — computed live on Modal's GPU infrastructure using sentence-transformers. Real ML inference, not API calls.

3. Persistent Taste Profiles via Supermemory

Your taste vector is stored in Supermemory and evolves with every ranking session. Mello doesn't just know who you are today — it knows who you're becoming.

4. Cosine Similarity Recommendations

When you hit the Recs tab, we pull your Spotify top 50, embed them all, and find the songs closest to your taste vector. Every recommendation shows a real cosine similarity score and explains exactly why: "Matches the style of Cruel Summer by Taylor Swift."

5. Taste Codes & Vector Blending

Every user gets an 8-character taste code. Share it with a friend, they enter it in the app, and we blend your embedding vectors — literally average your taste signatures in high-dimensional space — and surface music that lives in the intersection. Songs you'd both love that neither of you would've found alone.

---

Tech Stack

Layer	Technology
Mobile	React Native (Expo)
Backend	Next.js (deployed on Vercel)
Auth	Spotify OAuth via custom auth server
ML Inference	Modal (GPU, sentence-transformers/all-MiniLM-L6-v2)
Memory	Supermemory (persistent taste profiles)
Music Data	Spotify Web API
Local Storage	Expo SecureStore

---

Project Structure

MELLO/
├── mobile/                  # React Native (Expo) app
│   ├── app/
│   │   ├── index.tsx        # Auth & onboarding
│   │   ├── add.tsx          # Head-to-head song ranking
│   │   ├── list.tsx         # Elo-ranked song list
│   │   ├── recs.tsx         # ML-powered recommendations
│   │   ├── friends.tsx      # Taste codes & blended recs
│   │   └── share.tsx        # Shareable taste card
│   ├── context/
│   │   └── userContext.tsx  # Global user state
│   ├── hooks/
│   │   └── useSpotifyAuth.ts # Token management & refresh
│   └── utils/
│       └── storage.ts       # Elo ranking logic & persistence
│
├── backend/                 # Next.js API (Vercel)
│   └── src/app/api/
│       ├── auth/spotify/    # OAuth + token refresh
│       ├── openai/recommend/ # Recommendation endpoint
│       └── memory/          # Supermemory read/write
│
└── modal/
    └── recommend.py         # GPU embedding inference

---

Getting Started

Prerequisites

• Node.js 18+
• Expo CLI (npm install -g expo-cli)
• Python 3.10+ (for Modal)
• A Spotify Developer account
• A Modal account
• A Supermemory account

1. Clone the repo

git clone https://github.com/your-username/mello.git
cd mello

2. Deploy the Modal inference function

cd modal
pip install modal sentence-transformers numpy
modal token new
modal deploy recommend.py

Copy the deployed endpoint URL — you'll need it for the backend.

3. Set up the backend

cd backend
npm install

Create a .env.local:

SPOTIFY_CLIENT_ID=your_spotify_client_id
SPOTIFY_CLIENT_SECRET=your_spotify_client_secret
SPOTIFY_REDIRECT_URI=your_redirect_uri
SUPERMEMORY_API_KEY=your_supermemory_key
MODAL_ENDPOINT=https://your-modal-endpoint.modal.run

npm run dev        # or deploy to Vercel

4. Run the mobile app

cd mobile
npm install
npx expo start --clear

Scan the QR code with Expo Go on your phone.

⚠️ If your Spotify app is in Development Mode, add your Spotify account email at developer.spotify.com/dashboard → Your App → Settings → User Management.

---

Key Features

• Head-to-head ranking — swipe-style battles that reveal true preference
• Elo scoring — mathematically rigorous ranking that converges fast
• Match percentages — real cosine similarity scores, not vibes
• Song-level explanations — "Similar energy to X by Y"
• Taste codes — 8-character codes for social sharing
• Vector blending — mathematical average of two taste profiles in embedding space
• Historical memory — taste profiles evolve across sessions
• Shareable cards — beautiful dark-mode taste card with your top 10

---

How the Recommendation Engine Works

User ranks songs via Elo
        ↓
Top 10 ranked songs → Modal GPU
        ↓
sentence-transformers embeds song names/artists
        ↓
Compute mean embedding vector (user taste vector)
        ↓
Spotify top 50 tracks → Modal GPU (embed all candidates)
        ↓
Cosine similarity: taste vector vs each candidate
        ↓
Sort by score, return top 10 with explanations

For friend blending:

User taste vector (50%) + Friend taste vector (50%)
        ↓
Blended vector → cosine similarity against candidates
        ↓
Recommendations both users would love

---

Team

Built at HackIllinois 2026.