Aura — A Consent-First Garden That Blooms With Your Real Pulse

This is a submission for the Presage SmartSpectra SDK Challenge.

What I Built

Aura is a small generative garden that wakes up only when Presage SmartSpectra delivers a real human signal. A pulse waveform changes the flowers’ glow, sway and bloom; pulse relative to the session baseline shifts the palette. When the signal stops, the flowers freeze rather than performing a fake heartbeat.

I wanted to make sensing feel less like surveillance and less like another score to optimize. Aura is a one-minute creative wellness break: no streak, diagnosis, account, feed, or microphone. Just a patch of light that responds to you and disappears when you ask it to.

Demo

Watch the 69-second demo video (MP4)

The video shows the consent gate, the recorded-camera replay, the first delayed pulse estimates, changing confidence values, the live relative waveform and the reactive garden. This is not a prerecorded metrics animation: the run observed 251 SmartSpectra metric packets, ending with an accumulated 80.1 BPM at 53% confidence. The actual two-second DOM sample log is public.

For reproducibility without a webcam, the app includes a transformed synthetic rPPG avatar research clip from UCLA, with attribution. It is not footage of a user. The same app also supports a live webcam through custom RGBA frames. Recorded-file mode demonstrates the real SDK pipeline, but I do not present it as proof of accuracy on a live person.

Code

Repository: github.com/jeffreyturov-dev/aura-garden

The README includes installation, a free Presage API-key setup, unit tests, real-SDK measurement commands, Playwright demo reproduction, and an optional root-free CPU Vulkan setup for headless Linux.

How It Responds to Human Signals

Aura uses @smartspectra/node-sdk 3.4.0 in two genuine SDK input modes:

// Live browser camera → local WebSocket → SDK custom input
sdk.useCustomInput(FrameTransform.kNone);
sdk.sendFrame(rgba, width, height, width * 4, PixelFormat.kRGBA, timestampUs);

// Reproducible recorded camera clip
sdk.useFile(demoPath, { interframeDelayMs: 33 });
sdk.start();

The server requests breathing and cardio bundles and decodes each real metrics buffer:

sdk.on('metrics', (buffer, timestampUs) => {
  const metrics = decodeMetrics(buffer);
  const state = metricsState.ingest(metrics);
  broadcast({ type: 'metrics', timestampUs, ...state });
});

The visual mapping uses:

• cardio.pulseRate: BPM, confidence and stability display, plus gradual palette/bloom changes relative to a median session baseline;
• cardio.arterialPressureTrace: a relative waveform only and the flowers’ instantaneous energy/sway;
• breathing.upperTrace and breathing.rate: additional movement only when a real value exists.

One detail mattered: SDK windows do not all contain every metric. null must remain null. Aura retains the last real sample per signal but never invents a missing pulse or breathing value. If packets become stale, the rendering freezes.

The practical issue I hit

The official docs made custom input and file mode look straightforward, but my first cardio run stopped after two seconds with:

kSetupError: Vulkan is not available on this system.

The fix was not a mock adapter. I ran the SDK with Mesa lavapipe, extracted from pinned Debian packages into a user-writable directory with SHA-256 checks. npm run setup:vulkan and npm run start:lavapipe reproduce that root-free setup on Linux x86_64. Native Vulkan is still preferred where available.

The second timing detail is important for judges: the first pulse can take 12–15 seconds. Aura says so in the UI instead of filling the wait with random numbers.

Prize Categories

Most Creative or Delightful Use

Aura turns an invisible physiological trace into a small living place rather than a chart. The garden’s visual “breath” comes from actual SDK packets; its restraint is part of the design. There is no simulated garden performing behind the interface while the camera is off.

Best Thoughtful Real-World Use

The intended use is a brief, opt-in creative pause for adults who want an ambient experience without creating an account or storing a health history. The controls use plain language, explain the initial delay, expose confidence, and provide one obvious End session & erase action. It deliberately avoids stress labels, diagnoses and medical decisions.

Consent and Data Handling

• Consent: neither source can start until the user reads the disclosure and ticks an explicit consent checkbox. Starting a source is a second, separate action.
• Camera: live mode requests getUserMedia({ video, audio: false }). There is no microphone request, background capture or covert monitoring.
• Processing: SmartSpectra performs metric computation on-device. Aura sends live RGBA frames only from the browser to its local Node process via the local WebSocket.
• Network: the SDK requires network access for API-key authentication, subscription validation and encrypted model artifact delivery. Aura does not upload raw camera frames to Presage or any other third party.
• Telemetry: the SDK is initialized with enableTelemetry: false.
• Storage and retention: the application has no account, cookie, analytics, database or cloud app backend. Live metrics remain in browser/server memory. Stop releases all media tracks and clears metrics, waveform and garden state.
• Sharing and deletion: there is no automatic export or sharing. End session & erase clears the session. Closing/reloading also releases local media.
• Developer recording: the separate Playwright script intentionally saves the public demo and its run log. It is never triggered by a normal user session.
• Optional AI services: none are used at runtime.
• Safety: Aura is a creative/general-wellness experience for adults. It makes no diagnosis or treatment claim and does not support medical decision-making. The waveform is explicitly labelled relative and never as a blood-pressure reading.

Validation

npm test checks real SDK payload handling, null preservation, erasure, bounded visual mappings and consent-gated controls. The Playwright run is an end-to-end test against the actual SDK and fails unless it sees more than 100 metric packets and a numeric pulse. There were no browser console errors in the checked-in final run.

Thanks to Presage for the SDK and to UCLA’s Vision Lab for publishing the synthetic rPPG avatar research demonstration used for reproducible testing.