The reliability work nobody asks for

When Skilladder's assessment platform took on Nigeria's 3MTT programme, the work that mattered most was the work nobody asks for in a demo. The platform has now handled about 109,000 completed attempts from roughly 48,000 candidates.

Queues, not hope

During 3MTT, results and notifications went out one email at a time, and some never arrived. I rebuilt sending as a durable, database-backed queue with workers, so a failed send is retried instead of lost, and delivery no longer depends on a single attempt.

Every failure in one place

I wrapped all 128 backend services in an in-house monitoring stack with automatic error classification and an admin triage dashboard. Classifying errors means a spike arrives as one problem with a name, not hundreds of log lines.

Moving without breaking delivery

Moving the platform behind Cloudflare had to happen without disrupting email deliverability. A change like that, done carelessly, can push results into spam folders just when candidates are waiting for them.

Fairness is reliability too

A platform is not dependable if it fails particular candidates. When candidates were dropping off, an accessibility rebuild shipped in three days improved screen-reader support, keyboard navigation and use on low-end Android phones, and completion improved. A mobile fix restored fair retakes for 1,300+ affected candidates, read-aloud assessments use Nigerian accent voices, and extended time accommodations are enforced on the server so they hold on any device.

None of this shows up in a screenshot. All of it shows up when a national cohort relies on the platform.