Zeitra

Zeitra plans nutrition, training and recovery around when a person is actually awake and working. Mainstream health apps assume a stable daytime routine, so their guidance fails nurses, drivers, first responders and anyone on a rotating roster, and it ignores how appetite and energy move across a menstrual cycle. Zeitra treats the schedule as the input rather than the exception.

A conversational coach called Ria turns that schedule into a concrete plan, and the app captures what keeps the plan honest: food by photo, barcode or search, workouts against a large demonstration library, sleep, and heart rate read through the phone camera.

There is no organisation tenancy, so isolation is per user: every request carries a signed token and each service scopes its queries to that token's subject, never to an identifier supplied in the request body. Exactly two paths cross a user boundary, a coach-to-client relation and a revocable read-only cycle share addressed by an opaque code, and both are explicit records rather than implicit trust.

The deepest subsystem is camera photoplethysmography. With the torch lit and a fingertip on the lens, frame brightness rises and falls with the blood volume pulse, so the frame stream is a noisy one-dimensional signal whose fundamental frequency is the heart rate. A worklet on the camera thread reduces each frame to a mean luma plus its minimum and maximum; the spread between them decides whether a finger is present at all. Frames that fail that test emit a sentinel rather than a sample.

The estimator then works purely in the sample domain: sample rate from the median gap between timestamps rather than an assumed frame rate, baseline drift removed with a long moving average, sensor noise with a short one, peaks detected above a dynamic threshold under a physiological refractory gap, and the rate taken from the median inter-beat interval. Confidence comes from beat regularity and beat count, and below a floor the function returns no value.

Two decisions make it hold. The estimator imports nothing, not the camera library, not the UI framework, so the whole algorithm runs under the standard test runner with no native mocks. And it refuses to answer: an estimate that quietly returns a plausible number from noise is worse than one that admits it failed, because a fabricated pulse is indistinguishable from a real one to the person holding the phone.