What I Built
Your breathing moves the crosshair, the way it moves a rifle. STEADY is a seated breath-timing trainer for rifle and pistol shooters. Your webcam reads your chest breathing trace through Presage SmartSpectra, and that trace physically offsets the crosshair: zero sway in your natural respiratory pause (the still point after a normal exhale), maximum sway at full inhale. Fire in the pause and you hit; fire mid-breath and you miss red.
Shooters are coached to release the shot in the natural respiratory pause. Beginners can't see whether they do it; they only feel unsteady. Today that feedback needs a coach watching or a chest strap. A webcam is something every shooter already owns.
A session is consent → 45 s calibration → a 60 s Guided round (pause shaded green) → a 60 s No-HUD round (no help) → a report with every shot placed on your breathing curve.
Demo
Code
https://github.com/sairammr/steady
npm install
cp .env.example .env # add PRESAGE_API_KEY
npm start
No key or camera? npm run replay -- recordings/real-session-1.jsonl runs the whole detector and report on a bundled recording (a real 171 s session of mine: breathing trace, pulse, shots; numbers only, no video).
How It Responds to Human Signals
Signals used: chest breathing trace (breathing.upperTrace, the core), breathing rate, pulse rate (calm/elevated vs your own baseline only), talking (face.talking) and SDK validation status.
-
Trace → sway. Crosshair offset =
clamp((x − trough)/A, 0, 1) × maxOffset, mostly vertical, drawn about a quarter second behind the newest sample (≈ 250 ms in my session, shown on the HUD), interpolated, because delivery is jittery. - Trace → pause → score. Pause opens after an exhale when the trace is flat for ≥ 300 ms in the bottom 35 % of the breath, closes when the inhale starts or after 5 s. Score = hit accuracy × 1.0 in the pause, × 0.3 mid-breath. Final scores are computed post-round over the final trace with the same function that colours shots live.
- Talking / weak signal → grey. Shots while talking (+ 2 s) or with an invalid signal are not scored.
Measured honesty (method in the README):
- Delivery delay
D(arrival − sample timestamp, no human in the loop): 105 ms median, p95 261 ms (jitter above the fastest sample). The offset between the SDK sample clock and the wall clock differed between my two recordings (−413 ms, +109 ms), so STEADY measures it at calibration instead of assuming one. - Timing offset
Lfrom a paced guide (3 cycles, scheduled timestamps): not reliable yet: 265 ± 265 ms from 2 usable cycles on my first recording, 0 on the second. STEADY does not depend on it. - Metrics cadence median 33 ms (p90 44 ms), 1 sample per event (~27.5 Hz), revisions: none (0 repeated timestamps).
- Detector: tuned on one real recording (17 pauses found vs 2 before tuning), then run unchanged on a second real session: 11 pause windows in 171 s. No held-out agreement score yet; two recordings is too few to claim one.
- My results: first full session: 0 % of 22 scored shots in the pause with the HUD (I was firing as fast as targets appeared), 33 % of 6 without it once I slowed down. Honest, small, and the point: you can't feel this without feedback.
- Failure cases I saw: 9.5 % of samples invalid:
ChestNotVisiblewhen I leaned in to aim,FrameRateTooLowin dim light,FaceNotForward. Early builds paused the round on each one; it now keeps running through short dropouts (1.5 s grace, pause only after 5 s) and greys out shots in weak moments instead.
Prize Categories
Best Overall Human-Sensing Experience · Most Creative or Delightful Use · Best Thoughtful Real-World Use.
Consent and Data Handling
- Consent first: a consent screen runs before the camera opens: what is measured (chest breathing motion, breathing rate, pulse, talking, face presence), that processing is on-device, what is stored and what leaves the device. The camera starts only after "I agree"; a Camera off button is always visible; consent can be withdrawn in Settings. The copy asks you to be alone in frame, because the SDK tracks one face and the app cannot detect bystanders.
- Stored (locally only): session summaries and raw signal recordings (JSONL: breathing trace, breathing rate, pulse, talking/validity flags, timestamps, shots), used for replay and post-round scoring. Never stored: video frames or images; the webcam picture is display only.
- Shared: with nobody. No analytics, no crash reporting, no cloud.
-
Leaves the device: Presage SDK key verification, usage metering and model download (key + device/licence info). SDK telemetry is disabled (
enableTelemetry: false); the SDK forwards the flag to its main process (js/main/index.js). I have not captured network traffic to prove it. See Presage's telemetry page. - AI use: no AI or LLM service runs in the app; the coaching insight is rule-based. Presage's on-device models compute the signals. The code was written with Claude Code.
- Not medical: a training aid for breath timing; not a medical device; not for diagnosis. Pulse is only calm/elevated relative to your own session baseline.
- Retention and deletion: kept until you delete it. Settings → Delete all data removes every session summary and every raw recording.
Built with an agent
Built with Claude Code: PRD, detector tuning on my real recordings, UI and this post's numbers all came out of agent sessions, with every number re-measured by npm run replay on the bundled recording.
Limitations
- Seated desk practice only: the SDK needs a still, silent, front-facing subject in even light. Not for a real shooting stance.
- The sway is drawn about a quarter second behind the newest sample; the HUD says so.
- About 2.75 live minutes per session, so ≈ 18 sessions per free key.
- Tested on Apple Silicon only.
- Rifle/pistol natural-pause technique; archery's partial-exhale hold is not modelled.
- Detector thresholds were tuned on 2 of my own recordings; other people may need retuning.
- Real-shooter feedback: none yet.



Top comments (2)
TE CITO: "Desfase de tiempo respecto a una guía de ritmo (3 ciclos, marcas horarias programadas): no fiable todavía: 265 ± 265 ms desde 2 ciclos utilizables en mi primera grabación, 0 en la segunda. STEADY no depende de ello.L
Cadencia mediana de las métricas 33 ms (p90 44 ms), 1 muestra por evento (~27,5 Hz), revisiones: ninguna (0 marcas temporales repetidas).
Detector: ajustado en una grabación real (17 pausas encontradas frente a 2 antes de afinar), luego ejecutado sin cambios en una segunda sesión real: 11 ventanas de pausa en 171 s. Aún no hay puntuación de acuerdo retenido; Dos grabaciones son demasiado pocas para reclamar una."
Considero que es importante destacar tu sinceridad y evaluación de ensayos, demuestra que tienes los cálculos de un inicio de calibración de respiración para el objetivo que trasmites y te comento: que es una gran idea 💡 de evolución
Some comments may only be visible to logged-in visitors. Sign in to view all comments.