This is a submission for the Presage SmartSpectra SDK Challenge
Draft notes for the author. Every
TODOmust be filled from a real captured session or removed before publishing. Delete this block and the "Before publishing" list at the end.
What I Built
Chess.com tells you what happened in a game. It does not tell you what your body was doing while you played.
ChessPulse is a local desktop app for reviewing a chess session after you finish it. While you play, the app reads your pulse from the webcam with the SmartSpectra SDK. Afterwards, you paste your games in. The app puts the pulse trace, the clock, the think time, and the moves on one timeline.
The idea came from a common bad evening: three losses in twenty minutes, then a rematch you should have skipped. Engine reviews are good at showing the bad move. They do not show that the bad stretch started while your pulse was well above normal.
The report gives you:
- One timeline with the pulse trace, the moves, and the games placed in order.
- One verdict sentence that names the stretch that mattered most.
- Accuracy for each pulse band. "Elevated" means at least 10 bpm above your own first-minute baseline. A band with fewer than five scored moves shows no percentage.
- A refusal, not a weak report, when the camera did not capture enough usable pulse. The cutoff is 70% coverage.
- A short note between games, if your pulse stays elevated after you mark a game as ended.
- An optional AI reading. It is labelled as inferred text. Every number in the report is computed by the app, not by the model.
ChessPulse is for blitz and rapid players, roughly 700 to 2000 rated, who already review their games. It is not a chess trainer, not an engine overlay, and not a medical device. It does not diagnose anything.
The fair-play rule is the core of the product. No engine, evaluation, or advice appears while a rated game is live. Sensing runs during play, but the app only shows measurements and analysis after the game ends.
Demo
TODO: Embed or link the demo video. It must show one real session from start to finish: consent, pulse during play, the game-end marks, the paste, and the report. The report at the end must come from the session shown earlier. Nothing in the video may be a fixture or a re-enactment.
TODO: Fill in from the real session: date, number of games, pulse coverage percentage.
TODO: Add a screenshot of the report from that session.
The project has no demo mode and no fixture path. A report can only come from a session captured on the machine.
Code
Repository: https://github.com/kernelKain/chesspulse
npm install cp .env.example .env # add your SmartSpectra API key to .env npm test npm run dev You need Node 22, a display, a webcam, and a SmartSpectra API key. The README covers camera framing, and it explains the --no-sandbox flag that some Linux systems need to start Electron.
TODO: Confirm the license name in the repository and the placement of the Stockfish GPL-3.0 notice before publishing.
How It Responds to Human Signals
The app uses pulse. The report is built from pulse readings and nothing else.
Here is how the SDK is wired in:
- The SDK runs in Electron's main process. The API key is read there from the environment, and the page never receives it.
- The app opens the camera through the SDK. The page shows a preview, which is a smaller copy of the SDK's video output. The page never opens the camera itself.
- Each pulse reading has a confidence score. A reading counts toward your baseline and coverage only when the SDK marks it stable and the confidence is at least 50 out of 100. Weaker readings stay in the raw file, but they are not treated as zero.
- The median of the usable readings in the first 60 seconds becomes your baseline.
- Each usable reading covers the time until the next one, up to one second. Coverage is the share of the session that has usable readings.
- While you play, the window draws the pulse strip as readings arrive. A gap longer than a second breaks the line. The SDK's framing feedback, such as "Place more of the chest in view," is shown on screen.
The response is not instant. The trace is smoothed, so it lags the moment that caused it. The between-games note is the only live-style response, and it appears after you press Game ended. If the median of the last minute of usable readings is at least 10 bpm above your baseline, one line appears with the measured numbers and suggests a break. Pressing Next game started removes it. The app does not call that number a loss or a mental state.
The app cannot find game boundaries on its own, so you mark them. That is one manual click per game.
TODO: Confirm from the real session that the pulse strip moves before the first move, and that the between-games note fires from live pulse. If either does not happen, say so here.
Prize Categories
I am entering Best Thoughtful Real-World Use.
The people I built this for are adult blitz and rapid players who already review their games and want to know whether a bad session was a state problem or a chess problem.
Design choices for real use:
- Local only. No account, no hosted page, no cloud sync.
- Post-game only. No engine or advice runs while a game is live.
- Explicit, revocable camera consent.
- A refusal when the trace is too thin to support a report.
- No medical or diagnostic claims in the app or this write-up.
TODO: Confirm whether entering this category also applies to the overall prizes. I have not checked the contest rules for that.
Consent and Data Handling
Camera. The camera stays off until you press Allow the camera. While sensing, the app shows a Sensing line and the pulse strip, and the camera's own light is on. You can revoke consent at any time. Revoking releases the camera and stops any open session.
What is stored. One folder per session, under $XDG_DATA_HOME/chesspulse or ~/.local/share/chesspulse. The folder holds:
-
meta.json: the session header. -
samples.ndjson: pulse readings with timestamps and confidence values. -
games.json: the PGNs you pasted. -
report.json: the computed results.
No video or image file is written. Frames are handled in memory for measurement and the preview.
Retention and deletion. The data stays until you delete the session folder. Deleting that folder deletes the session.
Who receives data.
- The SmartSpectra SDK does pulse computation on your machine. Its backend is contacted only to authenticate your API key and check your subscription. That authentication is tied to your SmartSpectra account, as with any SDK use. Presage's privacy policy governs that data.
- The app turns off SDK telemetry (
enableTelemetry: false). Telemetry would send session summaries and device readiness reports. It never sends frames or measured values, but it is off here. - Sensing and report building send nothing else anywhere.
- If you add an AI provider key and press Write reading, that provider receives your question, your PGNs, and summary numbers. It never receives video or the raw pulse trace. Without a key, no request is made.
Not medical. ChessPulse does not diagnose or treat anything. Pulse values are not presented as health information.
TODO: The README currently says sensing and report building "upload nothing." That is stronger than the SDK's behavior, because of the authentication call. Change the README to match this section before publishing.
Agent Session
I used Entire to record the AI coding sessions for this project. Entire saves each session as a checkpoint in local git refs.
What I set up:
- Entire CLI 0.12.3, enabled for this repository.
- Cursor hooks, so Cursor chats in this repository are recorded from the session-start hook onward.
- An OpenCode plugin, so OpenCode sessions are recorded too.
- Entire's telemetry off,
push_sessionsoff, and a manual checkpoint push policy. A normalgit pushdoes not upload session checkpoints. - Secret redaction at Entire's default setting.
What it does not cover:
- Sessions that ran before the hooks were installed. Earlier chat history was not imported, so those sessions are not recorded.
- Other agents. Claude Code, Codex, and Copilot are not hooked in this repository, so any work done with them would not be captured.
- Anything after the recording. Checkpoints do not make commits or publish a transcript.
DEV's agent session uploader supports Claude Code, Gemini CLI, Codex, GitHub Copilot CLI, and Pi. It does not list Cursor or OpenCode, so I have not embedded a transcript here.
TODO: Decide how to show the session record. Options are a short excerpt written by hand, or a plain description like this one. If you upload any transcript, check it for keys and other sensitive data first.