Sonder was an experimental music‑driven dating‑profile and song‑recommendation project.
Users browsed “dating profiles” generated from songs, swiped right or left, and the system inferred
musical‑taste attributes from these swipes. A native clustering‑based recommendation algorithm then suggested new songs aligned to the user’s evolving profile.
⚠️ Deprecated: This project is archived and no longer functional due to changes in Spotify’s API that removed or restricted the data access the system relied upon.
Sonder explored a playful question:
What if songs had dating profiles—and what if your taste could be inferred by swiping on them?
The application generated a personality‑style profile for each track using metadata and descriptive features.
Users would swipe right (like) or left (dislike) on these profiles.
Each action updated a set of internal attributes describing the user’s preferences.
Those attributes powered a clustering‑based recommendation engine that surfaced new tracks the user was likely to enjoy.
-
Song Dating Profiles
Each song was assigned a personality‑style description generated from its musical features. -
Swipe‑Based Preference Modeling
Users interacted with songs through a simple familiar interface, making the experience intuitive and game‑like. -
Attribute‑Driven User Profiles
Every swipe updated a music‑taste vector representing the user. -
Clustering‑Based Recommendations
A lightweight machine‑learning approach grouped similar songs and used proximity to recommend music.
Sonder relied on Spotify’s Web API for essential metadata.
Recent changes—including more restrictive access policies, removed fields, and increased authentication requirements—broke the original data pipeline.
As a result:
- Song metadata cannot be fetched consistently
- Recommendation inputs fail to compute
- OAuth flows may no longer function
- The app cannot operate as intended
Because the project was a personal prototype, it is no longer maintained and is provided as is.
If you wish to revive or extend Sonder, possible directions include:
- Integrating with an alternative open music dataset (e.g., MusicBrainz + AcousticBrainz)
- Using embeddings (e.g., Spotify Annoy index dumps, if available externally)
- Expanding the concept to movies, podcasts, or books
- Rebuilding the recommender using modern vector search (FAISS, Milvus, Pinecone)
- Redesigning the swipe‑based interface into a general “taste‑discovery” tool
No license was originally specified.
If this project is revived, you may add one (MIT recommended for open projects).
Thanks to Spotify’s API ecosystem for originally making projects like this possible.
Though the app is deprecated, it remains a fun exploration of music, machine learning, and interaction design.
Repository maintained by @xpoes123. Project archived.