Youssef El Mounabih
From: United States of America
Location: Albuquerque
On Useme since today
About me
I build backend systems and I verify them, at production standard.
Six systems shipped solo in 2026, each with automated tests, CI and written architecture decisions. The one I show engineers: a ledger that is append only, enforced by a Postgres trigger and not application code, every entry signed and hash chained, covered by 52 unit tests and 13 end to end tests for forgery, replay, crash resume and chain integrity.
The other half is verification: 260+ paid evaluations of AI generated code at the highest difficulty tier for frontier AI labs, in repos I had never opened, across Python, TypeScript, Java, C++ and Go. I reproduced each environment in Docker, wrote my own test patches instead of trusting the reference solution, and separated a real defect from a harness artefact.
Stack: Python, TypeScript, Node.js, React, Next.js, FastAPI, NestJS, PostgreSQL, Docker, Linux, Bash, Git, REST APIs.
Rabat, UTC+1. Invoicing through Prestilink LLC, a US entity. Live code in the portfolio below.
Categories
CV / Résumé
Mar 2026 - Sep 2026
AI Code Evaluation Engineer
Mercor
260+ paid evaluations of AI generated code at the highest difficulty tier for frontier AI labs, in production repos across Python, TypeScript, Java, C++ and Go. Reproduced each environment in Docker and wrote my own test patches.
Oct 2023 - Now
Independent Software Engineer
Prestilink LLC
Backend systems shipped solo: REST APIs, Postgres schemas with constraints and triggers, Docker, CI pipelines. Six systems in 2026, each with automated tests and written architecture decisions. Public code at github.com/SpowZy
Portfolio
Every model and tool call gets a signed, hash chained receipt. Append only is enforced by a Postgres trigger and not by application code, so nothing can rewrite its own history. 52 unit and 13 end to end tests cover forgery and replay.
LambdaMART learning to rank model with an A/B harness using bootstrap confidence intervals. NDCG@5 at 0.94 in simulation, reported as simulation because that is what it is. Python, multi objective scoring, off policy evaluation.
Double booking is impossible by design: a Postgres range exclusion constraint rather than application logic, proven by 25 concurrent attempts where exactly one wins. Next.js and TypeScript on Vercel, 11 integration tests green in CI.