ProbaBaby

We live in Serbia as foreigners. Nobody told us which tests pregnancy and the months after birth call for, and our doctor warned us about a few we had already missed. So the app carries the checklist itself, offline: 51 pregnancy and newborn tests, and the vaccine schedules of 241 countries and regions, bundled in the app.

The language barrier shaped the doctor visit. The Doctor Visit and Pediatrician summaries switch to the doctor’s language, any of the 32, without changing the app’s. The doctor can scan a QR code and open the report on their own phone; the link opens once, expires after 24 hours, and the parent can revoke it. Or the parent hands over their phone: a full-screen mode shows one topic per swipe, up to ten pages, with buttons that make the text larger, because many of the older doctors I saw in Serbia read with glasses.

All of it is free, so no family pays to understand its own check-ups.

When one parent starts a feed, sleep, kick or contraction session, every other device in the family gets a read-only Lock Screen mirror through push-to-start, even with the app closed.

Making that dependable took a stack of separate fixes in June and July 2026:

• a database trigger that had lost its authentication after a platform key change;
• start tokens that were registered only after notification permission;
• a session row that never reached the server, now written directly;
• sandbox tokens labelled as production, now retried on the other push environment;
• zombie sessions, now prevented by a database index, age-aware timeouts and duration caps;
• mirrors that could never update, because their tokens were never observed.

What it learned is written down as a lifecycle contract that every Live Activity manager must meet.

Login froze because the first sync pulled the full history of about 34 tables on the main thread, and every import reloaded the screen. The fix imports in chunks of 50 with a pause between them, checks each chunk against edits still pending, and batches screen updates.

A variant that paused inside the bridge was rejected after a 60-agent, three-pass adversarial review showed that an edit made during a pause could be overwritten.

Offline writes were lost to three stacked causes: the queue had stopped draining after a refactor, an id lookup could never see the stored models, and passing network errors used up the retry budget. All three were fixed and checked against 25 adversarial scenarios.

The app ships 32 languages, Arabic and Hebrew right-to-left among them: 15,023 keys and about 480,000 localized cells. Five checks run on every commit, and a manual audit checks display width, counting a CJK character as two cells.

A repair in September 2026 fixed about 7,600 cells, among them Simplified Chinese that had been machine-converted from Traditional, and lists of warning signs that had shifted out of line. Read-only agents wrote the patches, and each was reviewed before it was applied.

Apple denied the Critical Alerts entitlement on 4 June 2026. A wake alarm on AlarmKit shipped the same day: a real system alarm that rings through the silent switch and Focus, with Snooze and Stop.

Backstop alarms that escalate every ten minutes until a feed starts, and cancellation when a partner logs the feed, came with it. The core alarm is verified on a device; those two are still on the device-test list.

Per-baby Home cards, Live Activities and widgets shipped in v1.0.3 in June 2026, with a short note about them in What’s New. From August, sign-ups from families expecting twins stood out, so v1.0.4 rebuilt the app twins-first: timers for each baby at once, delivery timelines adjusted for chorionicity, and corrected age. Sign-ups went from about 1.8 a day in the three weeks before it to about 10 a day after. The whole story is in Twins found us.

The optional prediction computes trait probabilities in code, with Punnett squares and no randomness, before any image is generated: eye colour from a two-gene model, hair colour with red hair modelled on its own, hair type, a three-gene skin model and five single-gene traits. Its confidence starts at 60% and grows with the data it is given, up to 95%, and the report’s text only narrates the numbers it is handed.