Designing the PulseSync Mix Mode
How we built emotion-aware playback from segment graphs, user tags, and local-only metadata.
The first time we played an emotion-tagged song in a completely different mood, we knew the idea was worth the engineering effort. This post walks through the key design decisions behind PulseSync Mix Mode.
Starting with a segment graph
The core interaction is a timeline where users mark “feel moments” inside a track. Each segment carries a mood label — calm, energetic, nostalgic — and a playback weight. When a user picks a mood, the player searches the local metadata store for matching segments and constructs a mix on the fly.
Early versions scanned metadata on every playback request. That was fast enough in development but unacceptable on real devices. We moved the index into Room as a prepopulated table and keep it in sync when tags change. This one move cut mixer startup from seconds to milliseconds.
Offline-only by design
We never planned to upload audio or segment data. The segment graph is stored in the same local database as the rest of the music library, and playback happens through the app’s existing ExoPlayer flow. This avoids a second consent boundary, a second network failure mode, and a second privacy review.
Why local inference instead of a server model
A mood classifier trained on acoustic features would have sounded impressive in a pitch deck. It would also have required shipping audio off-device, processing in a cloud function, and caching the result with no guaranteed offline fallback. We chose local inference because the inputs we care about — duration, tempo, tags, loudness — are already on the device.
What still needs work
The emotion vocabulary is still small. We tag only six moods today, and the UI exposes them as chips rather than a richer mixer. We plan a future editor that lets users combine moods and weight segments visually, similar to how the existing PulseSync Editor works for manual mixes.
If you are interested in these topics, the PulseSync documentation in the Features section has the current model definition and the tag schema we use across the music library.
Explore the Pulse Editor for marking segments and the Pulse Mixer for mix settings.