Amr Adel
Elkhouly

A wallet serving a national user base had to hold up through seasonal transaction peaks. Capacity was assumed rather than measured, and the peak was not far away.

Modelled the workload from real production volumes across the four highest-traffic journeys, then drove load until the ceiling appeared. Test data was generated per virtual user so caching and record locking could not disguise real contention — a detail that changes what the results mean.

Constraints turned out to sit at four independent layers, not one. Remediation ran across application code, database logic, configuration and infrastructure together, and cut average response time by roughly 80% while reducing resource utilisation.

Point-of-sale traffic could not be load tested the usual way. It depends on terminal hardware and on a third-party switch, and it carries live sales — so the acceptable error rate was zero, not a percentile.

Built a JMeter harness that constructs and transmits raw ISO 8583 messages directly against the gateway tier, reproducing exactly what the downstream switch consumes. That removed the dependency on hardware and on third-party availability entirely.

Established the real per-node ceiling and the headroom before errors appear, then converted that measurement into a production sizing model. It was adopted as the deployed capacity.

Examination results for a national student population publish at a single announced moment. There is no ramp and no second chance: effectively the entire user base arrives within minutes, and a failure is visible to the whole country.

Built and executed a spike profile reaching 200,000 concurrent users, shaped around the publication event rather than a steady-state average — because the average was never the risk.

Scalability limits were identified and mitigated before publication day, and the platform was signed off against a workload that reflected what actually happens rather than what is convenient to test.

AI tooling could take real effort out of test authoring, but a regulated financial institution cannot send source code or test data to an external model. The usual answer — a hosted API — was not available.

Architected an agent-based platform on self-hosted inference running inside the network: a design agent generating structured test cases against the team’s existing test hierarchy, and an execution agent driving browser automation. Nothing left the internal environment.

Scripting effort dropped by roughly 40%. Equally important, explicit validation rules and guardrails govern every generated artifact — so the output stays reviewable and maintainable instead of becoming a black box the team cannot own.