🚀 Inspiration
Gone are the days of fiddling with a clunky music player, desperately searching for the perfect track, or constantly babysitting an algorithm to figure out what you like. We realized that the future of the music industry shouldn't be a screen, it should be a direct link to your subconscious. Driven by our deep passion for applied machine learning and an absolute obsession with winning Hack the 6ix, we wanted to build something that literally reads your mind to DJ your life.
🎧 What it does
Museic is a revolutionary, biometric-driven music recommendation engine. As you scroll through a highly engaging, TikTok-style feed of curated songs, Museic quietly captures your physiological signals in real-time using an EEG headband and your webcam.
- It measures your brain waves, facial expressions, and even the intensity of your head-bopping to gauge your precise emotional reaction.
- It correlates your arousal and valence to specific timestamps in the audio to map exactly what you love, down to the second the bass drops.
- It generates a deeply personalized, witty "Music Enjoyment Profile" and automatically exports custom, emotion-based playlists directly to your Spotify.
- It even features a "Music Compatibility Score" that overlays your emotional curve with a friend's curve side-by-side to see if your vibes truly match!
🛠️ How we built it
We engineered a highly robust, full-stack architecture capable of handling high-frequency sensor data:
- Frontend: We built a custom Electron app to house our feed UI while securing the necessary local system access for webcam sensors.
- Hardware & Sensors: We utilized a Muse 2 headband streaming raw EEG data over Bluetooth (BLE) alongside the Presage SDK to capture high-fidelity facial analysis, pulse rate, and heart rate variability (HRV).
- Backend & Database: We developed a lightning-fast FastAPI backend that pipes our high-frequency, time-series vectors directly into MongoDB Atlas. MongoDB's time-series bucketing was a perfect fit for our append-only, sensor-timestamped writes.
- AI & ML: We used the Gemini API, orchestrated through Backboard.io, to digest biometric spikes and audio metadata to generate our dynamic "Music Enjoyment Profiles".
- Identity & Integrations: We implemented Auth0's native-app flow with PKCE for secure login, utilizing Auth0 Token Vault to seamlessly vend tokens for our Spotify API export flow, allowing users to export to authenticate with BOTH the login of their choice + their Spotify account through the use of Auth0's "My Account API"
🧱 Challenges we ran into
Integrating bleeding-edge hardware and software in a single weekend was no small feat,
- Signal Lag: We had to normalize drastically different data rates. For instance, we had to balance 60-second rolling HRV windows against near-instant facial expression transitions and head movements to ensure our arousal graphs remained sharp and accurate.
- Hardware Instability: We had to manage the Muse headband's BLE range and stability in an incredibly dense, Wi-Fi-crowded hacker venue.
- Complex Auth: Wiring Auth0's native-app login flow via custom URI schemes (like museic://callback) and safely storing tokens within an Electron shell required significant trial and error, ESPECIALLY for the spotify integration - literally took 3.8k tokens
🏆 Accomplishments that we're proud of
We are incredibly proud of architecting a complete, sensor-agnostic pipeline that successfully translates raw biological signals into tangible music curation. We successfully bypassed subjective human inputs, proving that we can use raw alpha/beta band power and facial landmarks to perfectly map human emotion to musical features like tempo, key, and RMS energy. We didn't just build a software hack; we built a working physiological prototype!
📚What we learned
We gained massive hands-on experience in biological signal processing. We learned how to manually compute alpha/beta band power using scipy.signal.welch on short rolling windows of raw EEG data. We also discovered how phenomenally capable Large Language Models like Gemini are at acting as an analytics translation layer, turning raw JSON biometric spikes into a witty and readable breakdown of a user's subconscious tastes.
🔮 What's next for Museic
We are building toward a future where you never have to interact with a music UI again. Next, we plan to push deeper into applied machine learning by training a custom neural network (like a 1D-CNN) on our aggregated dataset to predict an individual's exact arousal curve directly from song feature extraction alone. We want to take Museic beyond our curated local library and turn it into the ultimate, mind-reading layer for all global streaming platforms!
Log in or sign up for Devpost to join the conversation.