Serotonin Pod

The brief

The Exploratorium asked a team of five Minerva students for a pop-up photobooth concept that helps young adults practise emotional granularity: naming a specific emotion instead of a vague one. I built the web component. The full deliverable also included the booth exterior concept, a Figma user flow, and a printed takeaway card.

The prototype was delivered to the Exploratorium for internal review. It has not been installed for visitors.

How it works

1. Intro. A field of drifting clouds and a voice guide that welcomes the visitor.
2. Feeling wheel. Eight core emotions, each opening into three more specific ones, 24 in total. The visitor picks by voice or by tapping.
3. Pose studio. For every emotion outside the Fear family, the camera tracks the visitor’s body and draws a glowing “ghost” mannequin of the target pose next to their own skeleton. The skeleton changes colour as the two line up, and a hold timer fills once the match is close. Each session runs two one-minute poses.
4. Breathing studio. For the Fear family, a pulsing orb paces a two-minute box-breathing exercise while the voice guide talks the visitor through it.

The technical core

• Pose tracking. MediaPipe Pose runs in the browser and returns 33 body landmarks per frame.
• Alignment score. For ten key joints, the distance between the detected landmark and the target pose becomes a 0–100 score. Joints the camera cannot see are skipped.
• Voice. Three ElevenLabs conversational agents, one per stage. The agents can call functions in the browser: the intro agent selects the emotion the visitor names, and the pose agent can ask for the live alignment score.
• Overlay. Both skeletons are drawn with Three.js through React Three Fiber on a transparent canvas over the video, with smoothing between frames.
• Session flow. One state machine coordinates the camera, voice, timers, and the pose swap.

How it was built

I built the app with Lovable, an AI development tool, over many rounds of prompting, testing, and revision. The tool wrote code quickly. The work was deciding what it should build and checking that it did: translating the team’s visitor flow into screens and states, shaping how pose tracking, scoring, and the voice guide fit together, testing each iteration with a real camera and microphone, and reworking whatever broke or felt wrong.

When I published the repository I moved API identifiers into environment variables and added unit tests for the alignment score.

Limits

• It is a prototype, tested by the team and not with museum visitors.
• Pose targets are hand-set coordinates for a front-facing camera. They are not calibrated for body size or distance.
• The voice agents need an ElevenLabs account that is not included in the repository.