Zeyad Elganainy

I treat models as a powerful component, not the whole system. In my claims-review build, Claude makes the visual calls but a deterministic Python layer applies the rules and reaches the verdict, so the model never has the final say on money. AI earns its place inside clear boundaries; outside them, deterministic code and explicit rules stay in charge.

Anything can look good for five minutes. I care about the 44th claim, the timed-out API call, the mid-run crash. My claims pipeline writes results row-by-row so a failure never wipes finished work, and the job agent ships with 90+ tests and clear, provider-named errors. I build for the hundredth run, not the demo.

Automation should widen a person's reach, not take them out of the loop. My job agent scans, scores, and drafts on its own every day, but it never hits submit: the final apply is always a deliberate human click. When it cannot finish a role, it routes the job to a triage queue with the reason instead of failing silently.

These are not chores I bolt on at the end. At CentralSquare I used test-driven development to keep regressions out of production. The job agent falls back from Claude to Gemini automatically and never invents a fact that is not in its single source-of-truth profile. Guardrails, tests, and fallbacks ship as part of the product, not after it.