This is a submission for the Presage SmartSpectra SDK Challenge
Status on Nov 19. The SDK measurements in this post were taken on my real webcam. The game itself has so far run end to end only on its simulator. My first real session and the demo video go in here before the deadline. Until then every picture below is from the simulator, and its banner says so.
What I Built
Feltbeat is a game with one instruction: sit still and count the heartbeats you can actually feel. No finger on your wrist. A webcam reads your pulse through the SmartSpectra SDK and keeps the real count to itself until you have typed yours.
Then it shows you both. Your tally marks in ink, the camera's in red underneath.
From the simulator. No number in this picture is a real reading.
A session is five steps:
- One screen before the camera. It says what the camera does and what leaves the computer. The camera stays off until you press the button.
- The app waits for a steady pulse reading. While it waits, it asks what you think your heart rate is.
- One tone. You count. Two tones. You type your number. Nothing on screen moves or shows time during a round, and you are never told how long it lasted.
- One round of counting seconds instead of heartbeats.
- "Show what the camera saw."
Why the extra questions. There are two ways to do well at this without feeling anything. You can count time, or you can multiply a heart rate you already know. The seconds round shows how well you keep time. The guess shows what plain arithmetic would have given. When your count matches either one, the results screen says so and does not pretend otherwise.
Why a camera. The task needs an instrument the player cannot feel. In a 2019 study, people wearing a hard finger clip scored 51.5% on this exact task against 38.3% with chest electrodes, because they could feel the pulse in the clipped finger. A camera touches nothing.
Where it comes from. Psychologists have used the heartbeat counting task since 1981. Told to count only the beats they actually feel, people in lab studies report about a third of them. Research groups have run the task online with webcam pulse reading since 2022. I found no game that does.
Demo
The demo video is not recorded yet. It will show a real round, with the player's window beside the observer window, and it goes here before the deadline.
Until then, these are the two windows during a round, on the simulator. The player sees only the left one. The person running the session sees the right one: the pulse rate, the pulse waveform, and the SDK's confidence, stable flag and validation code.
| The player's window | The observer window |
|---|---|
![]() | ![]() |
And a round the app would not count:
To try it with no key and no camera: clone the repo, then npm install and npm run sim. The whole game runs on a simulated signal, and every screen carries the banner "Simulated data. No camera is in use and these are not real readings." With a Presage key in .env, npm start uses the real camera. It runs from source on Windows. There is no installer.
Code
SoumyaEXE / feltbeat
Count your own heartbeats without touching anything. A webcam keeps the real count until you have answered. Built on the Presage SmartSpectra SDK.
Feltbeat
Count your own heartbeats without touching anything. A webcam keeps the real count and shows it only after you have typed yours.
Feltbeat is a desktop game for Windows, written in Electron and TypeScript. It reads your pulse through the camera with the Presage SmartSpectra SDK (@smartspectra/node-sdk). It is a solo entry to the DEV Presage SmartSpectra SDK Challenge.
Feltbeat is a game about perception. It is not a health check and cannot tell you anything about your heart. It gives no advice, and it never comments on a heart rate.
The picture above is from the simulator, as its banner says. No number in it is a real reading.
Status on Nov 19, 2026. Every screen has been run on the simulator, and the SDK measurements in study/results/gates.md were taken on the real camera with the recorder scripts. A full session on the real camera…
- Layout. All scoring is pure functions in
src/renderer/core/, with unit tests. One file imports the SDK. Everything else, including the simulator, sits behind one small interface. - Check my measurements.
study/results/gates.mdholds every SDK number in this post, with the scripts that produced them inscripts/gates/. They run on your own machine with your own key. - Licence. The repo's code is MIT. The SDK is Presage's, installed from npm and not vendored, and none of their sample code is copied.
How It Responds to Human Signals
Signals. Two: PULSE_RATE and ARTERIAL_PRESSURE_TRACE, the pulse waveform. Feltbeat does not request breathing, HRV, expressions or face landmarks. The game does not need them, and each one would be one more thing to explain before the camera turns on.
The response, in real time. Most apps respond to a vital sign by showing it. This one responds in four other ways.
- It withholds. During a session the player's window receives one fact from the sensing layer: the signal is steady, or it is not. No pulse value reaches that window before the reveal. The observer window, for the person running a session and for the video, is the one deliberate exception.
- It waits. No round starts until the pulse reading has held steady for 8 seconds. In three runs on this laptop the first pulse sample arrived about 19, 20 and 46 seconds after launch. The slow one had a second face flickering in frame.
- It refuses. After each round the app checks the SDK's own signals for that window. A round counts only if 90% of it carried a good validation code, no bad stretch ran past 3 seconds, no gap between pulse samples ran past 5 seconds, and 80% of the samples were marked stable. A round that fails is voided in plain words ("The camera lost its steady view of you for a few seconds, so that round doesn't count.") and can be tried again.
- It draws the reveal. The camera's count is the pulse rate added up over the round, shown with a margin: Presage states 3 beats a minute either way for a reading it marks stable, which comes to 1.75 beats over a 35-second round. The screen rounds that up and says "give or take 2".
Integration. Electron, with the SDK's renderer build in a sandboxed window and its main-process half bound only after consent. This is the whole set-up, from src/renderer/sensing/sdkSource.ts:
const sdk = new SmartSpectraSDK({ apiKey, requestedMetrics: [MetricType.PULSE_RATE, MetricType.ARTERIAL_PRESSURE_TRACE], enableTelemetry: false, enableAccumulatedOutput: false, }); sdk.on('metrics', (buf, timestampUs) => this.onMetrics(buf, timestampUs)); // Replies arrive that nobody requested. Count them. Never read what they carry. sdk.on('insight', () => { this.insightEvents++; }); Scoring is arithmetic and the result sentences are fixed templates. Feltbeat itself calls no language model.
What I measured before trusting it as a referee
One Windows 11 laptop, its built-in webcam, SDK 3.5.0, plain Node recorder scripts, one person (me), one evening.
| Measured | Value |
|---|---|
| Pulse-rate samples | one every 1.14 s, each arriving about 1.3 s after its own timestamp |
| Waveform samples | one every 32 ms, about 31 a second |
One spell of kExcessiveMotion | no pulse sample for 17.3 s |
| A second "face" behind my chair | validation flickered between kOk and kMultipleFacesFound for 14 s |
| A 3-minute sitting run, replayed through the rules above | 96 of 128 possible 25-second rounds would have counted, and 56 of 108 possible 45-second rounds |
Two things for the next person using the SDK. A quality check that reads validation codes alone would miss the 17-second hole, so check for gaps between samples too. And validation changes several times a second, so judge a window by fractions of time, not by single events.
The windows in the last row overlap, so read it as "how much of the run was usable", not as separate trials. Every round that failed overlapped that one movement. Short rounds survive it more often, which is why Quick mode uses 25 and 35 seconds.
Prize Categories
Most Creative or Delightful Use.
What inspired it: a psychology task that only works if the instrument cannot be felt, and an SDK that reads a pulse without touching anyone.
What makes it unexpected: a vital-signs SDK is built to show you a reading. Here it is the referee that keeps it. The question is not what your pulse is, but how much of it you can feel.
What makes it playful: the reveal. You have committed to your number before the second row of marks starts to draw.
One other entry, Body Radar by Ashan_Dev, also hides the reading until you commit, and it is worth playing. It asks you to guess a rate. Feltbeat asks you to count beats, then checks the count against a seconds round and your own guessed rate.
Consent and Data Handling
Consent. The camera stays off until you have read one screen and pressed the button. "Not now" leaves it off. A "Camera on" mark stays on the window while it is on. Stop, the Esc key or closing the window turns it off. The app is for adults.
What is measured. Pulse rate and the pulse waveform. Nothing else is requested from the SDK.
What leaves the computer.
- A check of the API key with Presage. The SDK needs a connection for the whole session.
- The pulse reading, about once a minute. I did not ask for this. See below.
- Camera pictures: Presage's documentation says "Raw video and facial imagery are never transmitted". I have not verified that independently, and the consent screen says so.
- From Feltbeat itself, nothing. It has no account, no server and no analytics.
What the SDK sent that I did not ask for. Presage documents an LLM Insights feature in which "the SDK automatically dispatches vitals snapshots" to Presage, "which forwards it to the LLM". Feltbeat requests two cardio metrics, turns telemetry off and never calls requestInsight. I registered a handler anyway and logged the process's connections once a second. In one 180-second run:
| Observed | Value |
|---|---|
insight events received | 3, at 27, 100 and 160 seconds after launch |
| One payload, decoded once, from an earlier run | 1,020 bytes: a timestamp, and written text about the pulse rate the SDK had just measured |
| Remote endpoints of the process | 1, on port 443. The address is one that Presage's API host resolved to that hour |
| A setting for this among the 3.5.0 constructor options | none found |
What this does not show: the log sees where connections go, not what is in them. That the reading goes on to a language model is Presage's documentation, not my measurement. I did not ask Presage whether there is a switch I missed, so the claim is "no setting found", not "it cannot be turned off". The log was taken with a recorder script, and I have not yet repeated it on the Electron app.
What Feltbeat does about it: the consent screen says the reading is sent, before the camera turns on. The handler only counts the events. The text is never shown, stored or logged, because it comments on a heart rate and this app does not. If your app requests cardio metrics, register the handler and look before you write your own consent text.
SDK telemetry. Off (enableTelemetry: false), with no setting in the app to turn it on.
AI services. Feltbeat calls none. The exception is the SDK's own, above.
Storage and retention. Nothing is written unless you press "Keep this result". A kept result is one JSON file in %APPDATA%\Feltbeat\sessions with your answers and the pulse numbers for each round. No picture or video is ever saved. The file stays until you delete it. What Presage keeps of the readings it receives, I do not know. Their insights page does not say.
Deletion. "Delete this result" is on the results screen. "Delete everything" is under "About your data".
Sharing. None by the app.
Not medical. Feltbeat is a game about perception. It says nothing about anyone's heart or health.
Sources
- The task: Schandry (1981), "Heart beat perception and emotional experience", Psychophysiology 18(4), 483–488. https://doi.org/10.1111/j.1469-8986.1981.tb02486.x. I read it through Brener and Ring (2016), https://pmc.ncbi.nlm.nih.gov/articles/PMC5062103/.
- The strict instruction, and the seconds and guessed-rate checks: Desmedt and colleagues (2020), Biological Psychology 154, 107904. https://doi.org/10.1016/j.biopsycho.2020.107904
- The finger clip: Murphy and colleagues (2019), "I feel it in my finger: Measurement device affects cardiac interoceptive accuracy", Biological Psychology 148, 107765. https://doi.org/10.1016/j.biopsycho.2019.107765
- Counting against detection: Ring and Brener (2018), Psychophysiology 55(9), e13084. https://doi.org/10.1111/psyp.13084
- Webcams in this research: Arslanova and colleagues (2022), Affective Science 3(4), 862–877. https://doi.org/10.1007/s42761-022-00151-4
- Presage: LLM Insights and Telemetry and Privacy.





