Vansh Gadhia

← Back

Inception

Apr 2026 · Three.js, Python, FastAPI, TRIBE v2, scikit-learn, OpenBCI, MediaPipe, WebSockets

Inception is a browser-based 3D city builder where the world is shaped by how you feel. Emotion — arousal, valence, focus — together with mood and environment settings like sun elevation and fog density, determines what gets built and how it looks. You place objects on a grid by typing a prompt or by blinking, and Claude turns the object plus its emotional context into concrete Three.js material parameters: emissive glow, roughness, metalness, scale.

A second dimension runs through Meta's TRIBE v2 brain encoder. A prompt becomes an image via FLUX, the image becomes a short clip, TRIBE extracts a neural feature vector from it, and a scikit-learn classifier maps that vector onto a city element — bridge, lake, skyscraper, tree — which then spawns in the scene. The same vectors can be rendered as cortical surface maps with nilearn.

The most interesting mode is multi-brain. Two people wear EEG headsets — a g.tec BCICore-8 and an OpenBCI Cyton — and co-create in one shared world. The system streams each person's emotion at 10 Hz, watches for blinks to trigger hands-free placement, and computes symbiosis scores correlating the two signals. What shapes the city's atmosphere is the collective emotional state, not any one person's.

Around that sits a FastAPI backend and three WebSocket services: EEG decoding, MediaPipe iris tracking that steers the camera by gaze at 30 fps, and the TRIBE decoder. ElevenLabs narrates placements, and Perplexity or Sketchfab search pulls in external GLB models when the built-in asset kit doesn't have what you asked for.

Built in 48 hours at Global NeuroHack 2026, held at Frontier Tower in San Francisco, alongside teams from Oxford, Berkeley, Rice, and Imperial.


More projects