Your phone camera is a sensor, not just a photo tool

Most apps treat the camera as a way to produce a file. You point, you tap, you get a photo, and the interesting part happens somewhere else — an editor, a share sheet, a server. But a camera is really a sensor reporting a flood of measurements every second, and once you treat it that way, a different class of app becomes possible.

A camera is a measurement device

Point a phone at someone doing push-ups and the pixels are not the interesting part. The positions of their shoulders, elbows, hips and knees are. Modern phones can estimate those joint positions from a live camera feed continuously, on the device, without a frame ever leaving it.

That is the entire idea behind Form Coach. You set the phone on the floor, it draws a skeleton over your body in the camera preview, counts your reps, and scores each rep on form. You never tell the app what you did. It watched you do it.

Notice what changes about the relationship with the camera. Nothing is being saved. Frames arrive, get measured, and are discarded. What persists is small and structured: rep seven, a form score, a session summary you can look at later.

The measurement doesn't always need a model

Camera-as-sensor thinking isn't only about machine learning. GhostFrame composites yesterday's photo over today's live preview at an opacity you choose, so you can line your chin up with your chin before you press the shutter. There is no inference anywhere in that. It's the camera used as an alignment instrument rather than a capture button, and it's the difference between a photo-a-day project whose frames actually stack and one that jitters.

The general pattern: the camera gives you a live signal, and the app's job is to turn that signal into feedback fast enough to act on. Sometimes that takes a pose model. Sometimes it takes a compositing trick and an honest understanding of what the person is trying to do.

Why this belongs on the device

Once you see the camera as a sensor, sending the readings to a server starts to look strange. A rep counter that needs a network is a rep counter that quits in a basement gym. Feedback that has to make a round trip arrives after the rep it was describing. And the frames themselves are the most personal data most people generate — your living room, your body, your kids in the background.

A camera pointed at you in your own home is the most sensitive sensor you own, and the safest thing an app can do with those frames is measure them and throw them away.

On-device analysis solves all three at once. No latency, no outage, and no copy of the video anywhere to leak, subpoena, or accidentally train on. The constraint it imposes is real: you have to make the thing work on hardware you don't control, without a bigger model to hide behind.

The next wave of camera apps won't be about better filters. It'll be about phones that can watch a task and tell you something useful about it while you're still doing it — free, no subscriptions, ever, and private because the video never moved. You can see how we're applying that idea across our apps.