Ayuda Health, rebuilt release by release.

The Flutter app, 442 Dart files, had no working build environment and a stale main branch. Within a day the toolchain was bootstrapped without admin rights, the real release branch found, and a hard-coded version stamp in the Xcode project fixed after it caused duplicate-upload rejections. Version 3.4.1 went to both stores with health sync, Apple Health branding, a production environment switch and the "Ayuda Health: Pill Reminder" listing rebrand.

Google Play's Health Connect rejection was root-caused and a corrected build resubmitted the same evening at a staged rollout. 3.4.2 added the Android 16 target and a region-aware paywall and went live on iOS; 3.4.3 was held on TestFlight under a new rule: nothing is submitted without explicit sign-off.

The Health Connect flow re-prompted on every screen, requested roughly 68 Android permissions while the manifest declared 6, and every HealthKit call failed on real iPhones. The permission set was cut to a read-only list mapped to visible features, a previous store-rejection cause removed, and every read isolated per type so one failure cannot blank the dashboard. A 24-hour auto-sync was built on both platforms with catch-up on open and instant sync on connect; sleep goals auto-complete from wearable sleep behind a consent popup; an HRV tile was added; and iOS was rebranded "Apple Health" across 10 languages.

The app called Stripe directly from phones with a secret key inside the binary. Six payment endpoints were built first as a standalone TypeScript module with a 15-test suite, then ported into the backend following its conventions, with response shapes preserved so the app change was one repository file. The Stripe SDK was pinned to the API version the app's generated models expect.

A 30-agent adversarial review confirmed 12 findings, five fixed in code and re-verified. 92 of 92 end-to-end checks passed against real Stripe test mode, the first secret-free purchase completed on an emulator, and the swap shipped in 3.4.1. Production was unblocked by a supervised deploy after the auto-deploy runner was found dead.

The CEO's per-country pricing document was audited against the app, the backend and the live Stripe account: users had no country field, the backend served the first USD price to everyone, and existing subscribers had to be protected. A 25-tier map covers all 249 ISO country codes plus Kosovo with fallback countries flagged for review, IP-based detection ignores spoofed forwarding headers, and 50 Stripe prices were created under lookup keys by an idempotent, dry-run-first sync tool. Phase one went live first, so the already-shipped app showed regional prices immediately.

When "local pricing" turned out to mean local currencies, currency options were attached for 128 currencies at FX-derived, charm-rounded amounts, a Stripe-minimum bug that let sub-floor plans activate for free was fixed, and the app gained a region-aware paywall and correct currency signs.

An independent four-reviewer end-to-end review passed 13 of 13 requirements and also surfaced credential leaks on unauthenticated endpoints of the live API. They were closed in a single day along with stack-trace leaks, per-IP and per-email rate limiting, bcrypt password hashing with transparent migration on login, and an OTP-gated password reset. Stripping fields that shipped apps still parsed broke every login; the hotfix restored the contract within the hour while keeping the leaks closed. The API then moved from a raw IP address to https://api.ayudahealth.com with a Let's Encrypt certificate and TLS 1.2 and 1.3 only, with the old address kept alive for installed builds.

The monthly bill was reconciled to the cent across three regions with Cost Explorer at usage-type level and 14 days of CloudWatch metrics per resource. An unnamed Windows server at 0.4% CPU was 29% of the bill, an RDS Proxy had zero connections for its entire life, and the dev database had shown no connections for 14 days. Verification before cutting reclassified two "idle" servers as live APIs. With the owner's go-ahead the Windows box was snapshotted and stopped, and the proxy and dev database deleted behind final snapshots.

A same-day firewall tightening cut off connections the wide-open rule had hidden. It was rolled back three minutes after the report, and the lockdown was redone using VPC Flow Logs, which proved which server was the real production backend.

The owner asked which languages to add without growing the app, how to remind free users without hurting retention or store compliance, and how to bring wearable activity and calories into goals. A 14-agent analysis with readers, competing proposals, policy-checking judges and a critic found that the ten existing language files add only about 0.18 MB compressed, that the "trial ended" popup fired on every launch forever, and a zero-new-permission design for an energy loop. It published Ayuda-branded mockups built from the app's real design tokens.

With automatic translation chosen, all 931 strings were drafted into eleven languages with one translator and one independent reviewer per language, placeholder integrity checks and clinical back-translation. Ten language files were written for the next release; Traditional Chinese was held for a further review.