MEL Platform

01

The testing problem

MEL is not one app. The same feature has to behave identically in a browser, in a Tauri desktop build, and in a Tauri Android build that may be offline for hours — then reconcile when it syncs. Every screen exists in English, Hindi and Kannada. Every record is scoped by role: a field officer sees only their own centre, a State Head sees their states, an admin sees everything. A single candidate-status change can be correct on web and wrong on the offline path.

• Multiple role types with different visibility, each a separate data path to verify
• Three languages, two of which produce longer strings that break tight layouts
• Three runtime targets, one of which writes to a local store and syncs later
• Users are people with disabilities — accessibility is a functional requirement, not a polish item

02

How I set quality up

• Wrote the test suite module by module — Candidates, Trainings, Courses, Centres, Campaigns, Settlements, Workplace Support, Committees, Master Data, Reports and Notifications — to 1,028 cases with per-role variants.
• Ran a wireframe-versus-build comparison for every Master Data entity, sub-view by sub-view, so specification drift was caught before UAT rather than during it.
• Drove UAT cycles end to end and produced the sign-off packages, release notes and production readiness reports the client signs against.
• Built role-and-data-path smoke suites so a scope-visibility regression fails fast instead of surfacing in the field.
• Extended the same coverage onto Android through a self-contained Appium and emulator harness, with mobile UAT sheets mirroring the web ones.

03

Defects worth naming

The bugs I am proudest of are the quiet ones — the kind that pass a happy-path script and lose real data.

• Offline sync silently dropped rows: the scoped sync cursor keyed on a timestamp alone, so records written in the same tick were skipped forever. Fixed with a compound cursor.
• Duplicate-key failures surfaced as opaque 500s because the ORM wrapped the Postgres unique violation. Mapped to a proper 409 with a usable message.
• Future dates were accepted as a candidate's date of birth. Caught in UAT and tracked as a fail across two runs until it was fixed.
• A training could be archived while candidates were still on its roster — on the offline path as well as online.
• Editing a GSC centre did not prefill its pincodes, so saving quietly narrowed the centre's coverage area.

04

Where I crossed into the codebase

On a small team, handing over a one-line fix costs more than writing it. Over six months, 206 of my commits were merged across the frontend and backend — roughly 50,000 lines added. Several shipped as matched frontend and backend pairs on the same ticket.

• Settlement module and the 45- and 90-day follow-up scheduler, including sequence gating and draft persistence
• An application-wide centre and state scope switcher — API and UI, one ticket
• Server-side search, filtering, pagination and denormalised counts across every Master Data tab
• Committee PDF export, end to end
• English-only input validation enforced across the whole API and app
• URL-backed list filters, so a filtered view can be shared as a link
• The accessibility remediation sweep described in its own case study