HalfBeat

Watch the 80-second demo: halfbeat-demo-v2.mp4 · the report in a browser, no install: memeshee.github.io/halfbeat

A camera-only interview trainer that scores the one thing nobody scores: how fast you come back down. Built with the Presage SmartSpectra SDK, which estimates pulse, breathing, facial expression and talking from a plain webcam.

Every speaking coach on the market scores the level — “your composure was 71”. HalfBeat scores the return: after each hard question, how long until your pulse was back inside your own calm band, and what the exponential half-life of that decay was. That number is the product, because it is the one that changes with practice.

npm install
echo 'YOUR_SMARTSPECTRA_KEY' > ~/.presage_api_key
npm start                     # live camera session → http://127.0.0.1:7777

No key? npm run replay runs the whole app on a synthetic signal source so you can see the interface and the analysis. It is labelled MOCK everywhere and is never used for a reported result.

What a session does

1. Consent gate — the camera is not opened until you press the button. Plain language, no pre-ticked anything, and the app states what is and is not stored.
2. Framing check — face, light and signal quality from the SDK's own validation codes, with the hint text the SDK supplies (kTooDark, kChestNotVisible, kExcessiveMotion, …).
3. Settle-in (45 s) — your personal calm baseline: median pulse and a band, measured in this room, on this camera, today. Every later number is compared to your baseline only.
4. Rounds — a real interview question, a 60 s answer out loud, then a 60 s recovery window where you stop talking and sit still. The SDK's pulse (12 s average), talking detection and expression probabilities all land on one timeline.
5. Report — questions ranked by how slowly you came down, peak pulse and delta, time back inside the baseline band, and the decay half-life. Run the same question set again and the report compares the two runs.

How the numbers are made (and what they can't do)

number	how	honest limit
pulse	SmartSpectra cardio.pulseRate (12 s average, confidence gated at ≥40)	resolution is one 12 s window; a “time to come back down” of 24 s means two windows, not 24.0 s of precision
failure to recover	seconds from the round's peak until pulse stays inside baseline ±band to the end of the window	if it never returns, we say never and show the residual, rather than extrapolating
half-life	least-squares fit of ln(pulse − baseline) after the peak	useful as a trend across rounds, not as a physiological constant
talking	face.talking.detected	visual, not audio — no microphone is opened
expression	face.expression dominant class	used as context, never as the headline claim
HRV	cardio.hrv (RMSSD)	needs a full 60 s clean window, so it is usually absent in a 60 s round; shown as “— (needs 60s)” rather than faked

The 12-second pulse window is the reason this app is a recovery trainer and not a reaction-time game. Pretending otherwise would be the easy way to make a demo look better and a product useless.

Consent and data

• Camera only. No microphone, no audio, no screen capture.
• No video is written. Frames are decoded by the SDK in memory and discarded; the app stores a JSON file of numbers, and only when you pass --out (or accept the default sessions/ path).
• Nothing is uploaded. The UI is served from 127.0.0.1 over a local WebSocket. There is no server component and no account.
• SDK telemetry is off by default in this build: the SDK is constructed with enableTelemetry: false (see src/source.mjs). Pass --telemetry if you want the vendor's aggregate, non-identifying session diagnostics back on.
• Egress is measured, not promised. scripts/egress.mjs watches every outbound TCP connection the app's process tree opens while a session runs and writes docs/EGRESS.md. The app itself never calls requestInsight() and never uploads a frame; the table is the evidence, including the connections the SDK's own licensing and metric authorisation open.
• Wellness only. Every figure here is a general-wellness estimate. The SDK is not an FDA-cleared device — Presage's cleared product is a separate thing — and nothing in this app is a diagnosis, a treatment, or a fitness-for-work assessment.
• Delete = delete. The session files are ordinary JSON in sessions/; remove the directory and the record is gone. There is no other copy.

Commands

npm start                                   live camera
npm run replay                              synthetic source (no camera, no key)
npm run file -- --file clip.mp4             run a recorded video through the SDK

--source camera|file|mock    --port 7777    --out sessions/x.json
--baseline 45 --answer 60 --recover 60      --rounds 6
--headless                                  one JSON line per sample, no UI
--telemetry                                 leave the SDK's aggregate telemetry on
--no-open                                   do not open a browser

web/replay.html renders any saved session with no server, no camera and no SDK. It is deployed statically at https://memeshee.github.io/halfbeat/ — open it and it renders the bundled sample session in the browser, with no install and nothing to trust but the JSON.

Layout

src/index.mjs      process entry, CLI, signal handling
src/source.mjs     camera / file / mock sources behind one interface
src/signals.mjs    protobuf metrics → one flat sample per tick
src/session.mjs    phases, rounds, the session artefact
src/recovery.mjs   baseline, peak, time-to-band, half-life, ranking (pure functions)
src/server.mjs     localhost HTTP + WebSocket
web/               the UI (no build step, no framework); report.js is shared with replay
scripts/verify.mjs end-to-end check, exits nonzero on failure
scripts/eval.mjs   runs the clip set and writes the numbers table in docs/EVAL.md

Requirements

Node ≥ 20 on Linux x64/arm64, macOS arm64 or Windows x64. The SDK ships prebuilt native runtimes per platform (npm install pulls all of them, ~400 MB). Linux needs glibc ≥ 2.35.

Credits

Vitals are computed by the Presage SmartSpectra SDK; the recovery analysis, the UI and the evaluation harness are this repository's work. Licence: MIT.