Context
App Tower is an internal console for the ~660 white-label school apps an ed-tech company ships on Android and iOS. I was product owner, tech lead and architect. The users are the internal team that builds, publishes and watches those apps.
The problem
Builds, releases and monitoring for the fleet lived in scripts and consoles. One bad build could reach every school.
The trigger was an incident where a bad config served 404s for about 10 months before anyone noticed. There was no single place to see fleet health.
What I did
I wrote the spec, 31 stories and about 40 API contracts before any code existed.
The console covers a fleet list with health filters, batch builds, and staged Google Play rollouts that can be raised, halted and resumed. Rollouts have per-school edit locks and an append-only audit trail.
Build dispatch is exactly-once, so a double click cannot start two builds. A crash-free monitor reads Sentry with a fixed time budget and a cache lock to stop stampedes.
Delivery was spec-first. AI agents wrote the code, a separate agent reviewed it adversarially, and only humans approved the gates.
Key decision
I wrote contracts and tests before code. Agents write fast, so the only way to keep review meaningful was to fix the target first and make the tests the judge.
The cost was front-loaded work on the spec and contracts before the first feature shipped. I accepted that so a fleet-wide tool could not reach production unreviewed.
Result
From kickoff to production took 10 days. The backend suite finished at 2,228 specs with 0 failing, and the browser suite at 345/345 end-to-end tests.
The bucket sync, which maps existing app configs into the database, ran first in production for 133 schools.
What I learned
Some gates after phase 1 were never formally approved. Next time the gate itself blocks the merge, instead of relying on me to notice.
Stack
- Rails
- GraphQL
- React
- Sidekiq
- Playwright
- Sentry
- Google Play API
- GCP
