KINTO stabilized an undocumented legacy payments platform and took failing user approvals to zero errors in the observation window, with alerts that fire within minutes.
Problem
A payments platform inherited from a previous developer had no documentation. Failures were silent: user sign-in broke, and account approval left accounts half-created. The company website also had to be migrated.
Signals
- Market: International payments in its segment are still handled with manual transfers, email and spreadsheets.
- Business: The platform depended on third-party services whose outages dragged entire processes down.
- User experience: Approved users could not use their dashboard.
- Support: Errors arrived as user reports; with low traffic and no alerts, failures went unnoticed.
How we understood it
- Cloud infrastructure audit (inventory, network, costs) with prioritized findings.
- Code inventory, verified identical to what ran in production.
- Dependency review: the platform carried fragile dependencies; stabilizing it came before evolving it.
- Root cause from data: targeted queries and telemetry showed that third-party outages propagated unchecked into sign-in and approval.
What we built and why
- Graceful degradation: approval always persists and is marked “pending” if a third party is down; accounts are recovered later through an idempotent, protected process. Providers are switchable by configuration.
- Sign-in fix without breaking access, plus front-end guards.
- Observability: logging turned on and alerts for the failures that matter.
- Runbook and idempotent deployment with backup, verification and rollback; new tests; documentation and a phased roadmap.
- Company website migrated.
Why this way: When a third party goes down, approval should stay “pending” and be recovered later by an idempotent process instead of leaving accounts half-made; and failures should raise their own alerts.
Transformation
-
Sign-in from error to correct response, validated end to end.
-
Failing user approvals to zero errors in the observation window; deployed on the first try.
-
From silent failures to alerts in 15–30 minutes.
-
From an undocumented platform to versioned, documented code.
-
user approval errors: failing → zero in the window (measured)
-
detection time with new alerts: 15–30 minutes (measured)
What we don’t know yet
- Validation with real users at close and the new site’s metrics: no evidence.
