This is a submission for the Presage SmartSpectra SDK Challenge
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 pink underneath.
From the simulator, as the yellow banner says. No number in it is a real reading.
Where it stands today, Nov 21. Three plain facts, because the rest of this post depends on them.
- The game runs end to end on its simulator, and you can play it in two commands with no key and no camera.
- The SDK has read my pulse on this webcam. On Nov 16 a recorder script got 130 pulse samples in a 180-second run, with a median of 52.2 a minute. The demo video replays a real recording from that evening.
- Inside the app, on the real camera, no round has been played yet. Today I ran five sessions. The camera path works: in one of them the app forwarded 15,473 frames in under nine minutes, 29.8 a second. The SDK never sent a pulse sample, so the app waited on "Finding your pulse" and refused to start a round. That is what it is built to do. Why it happened is the most useful thing in this post, and it is in "What the real camera taught me" below.
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." Then, if you like, "Play it back": one soft tick for every beat the waveform held in that round.
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.
Demo
Watch the demo video (101 seconds, with narration)
What it shows, in order, and what each part is:
- A real recording, replayed. My pulse rate and pulse waveform as the SDK reported them from my webcam on Nov 16. Every number and every point of the line is a sample from the log.
- The game, on its simulator. Both windows side by side: the player's, which says "Counting." and nothing else, and the observer window, which shows what the SDK reads. A round the app refuses to count. The reveal. Play-back. Keeping and deleting a result. The simulator's banner is in every frame.
- The real camera, today. Two screenshots of the app waiting with the camera on and no reading.
My face is in none of it. The capture tool hides the camera preview for every screenshot and drops every video frame while the preview is on screen.
| Before the camera | During a round, the player sees only this | A round that does not count |
|---|---|---|
![]() |
![]() |
![]() |
| The Play tab | The reveal, with "Play it back" | History |
|---|---|---|
![]() |
![]() |
![]() |
All six are from the simulator and carry its banner.
To try it with no key and no camera: clone the repo, then npm install and npm run sim. 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.
Both pictures are from the simulator, as the banner says. No number in them is a real reading.
Demo video (101 seconds): media/feltbeat-demo.mp4. It replays a real recording from the webcam, shows the game on its simulator, and shows what the app did on the real camera.
Status on Nov 21,…
-
Layout. All scoring is pure functions in
src/renderer/core/, with 345 unit and integration tests across the app. 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/andscripts/dev/. 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 five 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. Today it kept the game shut through three sessions of several minutes each.
- 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".
- It plays the round back. A second count comes from the peaks of the waveform. When that count holds up, "Play it back" ticks once for each of those beats at their real spacing, while the camera's marks light up one by one. It is the closest the game gets to letting you hear what you sat through.
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), the evening of Nov 16.
| 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 |
The windows in the last row overlap, so read it as "how much of the run was usable", not as separate trials.
What the real camera taught me on Nov 21
Today was the first time the app itself ran on the camera. It never reached a round, and it took five sessions to find out why. Both reasons were invisible on Nov 16.
| Observed today | Value |
|---|---|
| First session | the SDK stopped after about 8 seconds: error 14, frame rate too low |
| The same webcam with no SDK, room light, automatic exposure | 14 to 16 frames a second, while reporting 30 |
| The same, exposure forced short | 29 to 30 frames a second, picture brightness 7 to 10 out of 255 |
| The same, with a lamp in front | 29 to 31 frames a second, brightness 107 to 115 |
| Third session, with the lamp | 15,473 frames in under nine minutes, no error, no pulse sample |
| What the SDK reported during sessions three to five |
kChestNotVisible, in every sample I took |
Two things for the next person using the SDK.
Your webcam may be lying about its frame rate. In a dim room mine kept the picture bright by doubling the exposure and delivering half the frames. getSettings() still said 30. The SDK ends a session when the rate stays low, so the symptom looks like a crash eight seconds in. Count the frames yourself before you blame your code. npm run camera:check in my repo does that with no SDK and no key.
"Chest not visible" is not only about breathing. I request no breathing metric, so I had treated that code as harmless, and my setup screen said "The picture looks right." With that code reported for minutes on end, the SDK sent 15,473 validation events in one session and not one pulse sample. On Nov 16, when the code was about 2% of the time, pulse samples flowed. I cannot prove the code is the cause from three sessions on one person. I can say what I changed: the app now counts it as a picture to fix and tells you to sit further back, so your shoulders and chest are in view.
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. Then you can hear the round you just sat through.
Two other entries work nearby, and both are worth your time. Body Radar by Ashan_Dev hides the reading until you guess a rate. Practical by Samarth Kumble includes heartbeat counting as one of four experiments, with the pulse hidden, and reports sessions from two people, which is more than I have. What Feltbeat adds to the same task is the seconds round, the guessed-rate check, and a player's window that is told nothing but "steady" or "not steady".
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 with the recorder script:
| 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 |
On Nov 21 I repeated the connection log on the app itself, across five sessions. The only remote endpoints were two addresses on port 443, the same two Presage's API host had resolved to. No insight event arrived in those sessions, which fits: no pulse reading was produced, so there was none to send.
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". I am not the only one to notice the sending: Rehearse by Emanuele Miranda and beats to breathe to by Geovana Sousa report it too.
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.
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, from the History tab or with "Delete everything" under "Your data". What Presage keeps of the readings it receives, I do not know. Their insights page does not say.
Sharing. None by the app.
Not medical. Feltbeat is a game about perception. It says nothing about anyone's heart or health.
A reader's objection, and my answer (Nov 21)
In the comments, Adam Lewis wrote that he does not trust the voiding rules to hold over a whole session, and that 96 of 128 comes from replaying one sitting on one laptop with one face. If the thresholds were picked against that same run, he said, they need a fresh session or two before the void rate means anything.
He is right about the number. It describes one run. On his second point, the thresholds were written on Nov 10, before any run, and the replay did not change them. But this repository's history starts on Nov 19, so you cannot check that, and I will not ask you to take it on trust.
So I ran fresh sessions today, and the honest void rate from them is not a rate at all: no round started. The rules did not throw rounds out. The step before them, waiting for a steady pulse, never passed. That is a weaker result than the one he asked for, and it is the one I have.
Agent Session
I built Feltbeat with Claude Code. It proposed the idea from my brief, wrote most of the code and drafted this post. I chose what to build and what to cut, sat for every camera run, and I answer for every claim here.
What went wrong, and what this can't tell you
- No round has been scored on the real camera. Everything about counting, the seconds round and the reveal has been exercised on the simulator only. The camera's side is real up to the moment a pulse reading should arrive.
- I lost most of the window. Nothing was built between Nov 10 and Nov 16. Cut to fit: a round where you tune a ticking tone to your own heartbeat, a longer mode, the installer, and a planned comparison across ten people. No other person has played.
- My check of the camera's accuracy is weak. I counted my pulse at the wrist for one 60-second window and got 54 where the camera counted 52.0. When I replayed the log, the camera had given no pulse sample for about 38 of those 60 seconds, and the 52.0 comes from bridging the holes. The app's own rules would have thrown that window out.
- I guessed wrong about the SDK, twice. That leaving breathing out would keep it from sending anything, and that "chest not visible" would not matter without breathing. Both were wrong on this machine.
- The app asks the camera for 640 by 480. I lowered it from the SDK's default of 1280 by 720 before I found that light was the real cause of the low frame rate. I have not gone back to test the default through the app.
- It cannot tell you anything about your heart or your health, or how well you sense your body in general. Counting tasks and beat-detection tasks do not agree with each other, and researchers dispute what this one measures.
- It may not work equally for everyone. Camera pulse reading is known to work less well for some people and in some light. I tested one face, and today it did not work for that one.
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.










Top comments (2)
I don't trust the voiding rules to hold over a whole session. Honestly, I think that's the weak spot. A 17-second gap and 14 seconds of flickering validation say the camera loses you easily. The 96-of-128 figure comes from replaying one 3-minute sitting on one laptop with one face. If the thresholds were picked against that same run, they'd need a fresh session or two from someone else before the void rate means anything.
Counting your own heartbeats while the camera checks you is a clever way to make the measurement interactive. Nice entry for the challenge.