Problem Statement

Pilots listening to new music often bounce after the first 15 seconds. We built the Music Recommendation Platform to respond to that drop-off in real time by mixing content-based filtering with reinforcement feedback.

System Design

1. Intent Detection – A mini BERT classifier interprets natural-language requests such as "late-night coding" or "jazz for brunch" into seed vectors.
2. Candidate Generation – Spotify audio features feed a similarity search powered by FAISS.
3. Interactive Feedback – Each thumbs-up or skip updates a contextual multi-armed bandit that reprioritises candidates.

Feedback Loop

flowchart LR
    A[User Request] --> B(Intent Detection)
    B --> C(Candidate Pool)
    C --> D{Feedback?}
    D -- Like --> E(Update Positive Weights)
    D -- Skip --> F(Update Negative Weights)
    E --> C
    F --> C

Impact

• 27% increase in session length across beta testers.
• Average of 3.4 on-the-fly adaptations per listening session.
• Generated explainable recommendations by surfacing the acoustic features behind each pick.

What's Next

• Experiment with contextual embeddings from OpenAI's text-embedding-3-large model.
• Roll out collaborative playlists so multiple listeners train the policy together.