Crane uses on-device machine learning to detect posture problems in real time. Here is exactly how it works — no black boxes, no hand-waving.
When you first launch Crane, you sit naturally in your normal working position. Crane captures your baseline posture using MoveNet, a state-of-the-art pose estimation model from Google. The model identifies key body landmarks — nose, left ear, right ear, left shoulder, and right shoulder — and records their relative positions.
This baseline represents your "good posture" reference point. Crane does not impose a universal posture standard — it learns what good posture looks like for your body and your setup. If you recalibrate later (new chair, new desk height), Crane adapts to the change.
Once calibrated, Crane continuously processes your webcam feed using TensorFlow.js running entirely in your browser process. Each frame is analyzed for deviations from your baseline posture.
Three layers of signal processing prevent false alerts:
EMA smoothing means that if you briefly lean forward to reach for your coffee, Crane does not react. The smoothing algorithm requires sustained deviation before adjusting the posture score. Confidence gating handles poor lighting, partial occlusion (hand in front of face), and other conditions where the model cannot reliably estimate pose.
The sigmoid scoring function maps raw deviation values into a smooth 0-to-1 posture grade. This avoids the "jitter zone" problem where small movements cause the score to flip rapidly between good and bad.
When your posture score degrades beyond the configured threshold for a sustained period, Crane displays a non-blocking notification. The notification appears as a translucent overlay on your screen — visible enough to notice, subtle enough to ignore if you are in the middle of something important.
Key design decisions:
The goal is to inform, not interrupt. You see the nudge, process it in your peripheral vision, and adjust your posture when you are ready — not when the app demands it.
Independent of posture detection, Crane runs an ergonomic cycle timer based on your selected profile. Each profile defines a repeating sequence of activities — sit, stand, walk, exercise, stretch — with durations based on occupational health research.
For example, the Deep Focus profile for developers:
When a transition is due (e.g., time to switch from sitting to standing), Crane sends a notification with the next activity. The timer continues regardless of whether you follow the suggestion — Crane guides but never forces.
Crane includes 15 profiles covering work styles from intensive coding to creative work to call-center roles. Three are free; all 15 are available with Pro.
Crane's privacy model is not a policy decision — it is an engineering decision. There is no cloud backend for posture data. There is no server that receives your webcam feed. Your posture data never leaves your machine. The only network activity is license validation for Pro users and app updates delivered through the store.
Your webcam feed is processed frame-by-frame in memory by TensorFlow.js. Each frame is analyzed, a posture score is computed, and the frame is immediately discarded. No images are saved to disk. No data is buffered, queued, or transmitted.
This is the difference between "we promise not to look at your data" and "there is no mechanism to look at your data." Crane chose the latter. Even if someone compromised the application, there is no data to steal — because no data is stored.