PowerUp

A model reading a blood pressure monitor is reading a clinical number. Silently writing a misread value into someone's health record is a far worse failure than not reading it at all, because the user has no way of knowing it happened — and in a product an employer has asked its staff to trust, one bad number that surfaces later costs more credibility than the feature ever earned.

So extraction commits nothing. It pre-fills the form. The user sees the photo they just took with the parsed values in editable fields beneath it, and nothing saves until they confirm. Practically, that turns data entry into a glance-and-tap, which is the actual win. The model isn't there to be autonomous. It's there to remove keystrokes.

That confirmation step is also the most valuable thing in the system, and treating it as merely a safety net would waste it. When a user edits a value the model got wrong, the edit is a labelled example: a real photograph, taken by a real user, on real hardware, in real lighting, paired with ground truth. No synthetic dataset gives you that distribution.

So corrections are captured and the model retrains on them bimonthly. It improves specifically where its users struggle — the device models, screen types and lighting found in the field, not the ones that happened to be in the original training set. The safety mechanism and the improvement mechanism are the same mechanism.

The cheap route is a general-purpose OCR or vision API. We trained a model instead, because this is a narrow, closed visual domain — segmented LCD digits in fixed positions on a small set of consumer device layouts, photographed at arm's length. Narrow domains are where a purpose-trained model beats a general one on both accuracy and cost per call. It also means the model belongs to the client and improves on the client's own data, rather than the feature's quality being pinned to whatever a third party ships next.

Background health sync is expensive in the two currencies that matter for a wellness app: battery, and the permissions you have to ask for. Both platforms treat persistent background access to health data as a serious grant — and an employee being asked by their employer's app for always-on access to their health records is a conversation the product doesn't need to have.

Sync-on-open gives the user current data every time they actually look, at no battery cost, behind a permission prompt that's much easier to say yes to. The tradeoff — data fresh as of last open rather than continuously — is invisible in a product whose interaction model is already daily check-in rather than live monitoring.