Unturtle

Inspiration

Hour three of an assignment. Your head is halfway to the screen and you haven't noticed. In a survey of 500 software professionals in Chennai, 69% reported a work-related musculoskeletal disorder in the past year. The neck was hit hardest, and only 34% took regular breaks.

Posture apps exist, and most of us have muted one. They repeat the same alert until you stop listening. In one study of clinical reminders, every extra reminder lowered acceptance by 30%. Plenty of people also don't want a camera app watching them at all.

Our first idea, SlouchLock, locked your screen until you sat up. We dropped the lock, because forcing people is how an app gets uninstalled. People call this hunch "turtle neck", so we named the fix Unturtle.

What it does

Unturtle uses the webcam you already have to notice when your head drifts toward the screen. It compares you with how you sit when you're comfortable, so there's no "ideal" posture to live up to.

20-second calibration sets your baseline.
Drift detection: if your head stays 12% below that baseline for 90 seconds, Unturtle steps in.
A ladder that backs off: a corner nudge, then one soft chime, then a suggested two-minute break. Dismiss the nudge and it won't repeat it; next time it offers a break and tells you why. Ignore everything and it stops at a suggestion, then goes quiet. It never locks your screen.
Breaks you confirm: the camera notices you stood up, left and came back. That only supports your answer. You confirm the break.
You stay in control: meeting mode, snooze, pause and quit are one click away.
Private by design: frames are processed on your laptop and dropped. If history is on, it keeps 5-minute event summaries, and you can delete them in one click.

How we built it

Pose: MediaPipe Pose Landmarker running in the browser (WebGL/WASM), at about 15 ms per frame on an Apple M4.
Features: head height and head–torso alignment, scaled by shoulder width so your distance from the camera doesn't matter, then smoothed over time.
Detection: an alert line at −12% and an all-clear only below −6%, a 90-second persistence window, and confidence gating. When the model isn't sure, Unturtle makes no claim.
Policy: plain rules instead of a black box. Every decision is logged with its reason, and that log powers the "Why am I seeing this?" button.
Privacy you can check: the status bar counts frames read, frames saved and network requests since the camera turned on, using the browser's own Performance API.

Challenges we ran into

False alerts. Single bad frames made the signal flicker. Smoothing, the persistence window and separate alert and all-clear thresholds stopped it.
People leaving the frame. When someone stood up and walked off, the pose tracker kept reporting a skeleton that wasn't there, then failed to find them when they sat back down. We added plausibility checks, so "not sure" means no alert, and a fresh tracker whenever the feed restarts.
Our own privacy counter caught us. Swapping in a fresh tracker re-downloaded the model, and the network counter ticked up. We now build the spare tracker before the camera turns on, and the count stays at zero.
Not becoming another nag. The hardest design question was when to stop.

Accomplishments that we're proud of

A working end-to-end prototype: calibration, drift detection, the adaptive ladder, the break flow, meeting mode and history delete.
In our recorded run it read 3,667 frames, saved 0 and made 0 network requests after the camera turned on.
Claims we can defend. We checked competitors' official pages and claim only what we couldn't find elsewhere: reminders that adapt to how you respond, and a break you confirm.

What we learned

Plenty of apps can already detect a slouch. The product is everything after that: whether people listen, whether they trust it, and whether it knows when to back off. We also learned that your own comfortable posture makes a better baseline than any textbook one.

What's next for Unturtle

Finale (2–3 Nov): a desktop tray app built on site, plus a Windows benchmark.
Pilot: one week, 15–30 volunteers, fixed reminders against adaptive ones. We'll measure alert acceptance, confirmed breaks, false alerts per hour, 7-day retention and annoyance, and publish what fails too.
Then: a pilot sponsored by a college, with opt-in, aggregate numbers only. No person-level reports to teachers or employers.

Unturtle is a work-habit aid, not a medical device.