Taha Y.

Your AI feature works in the demo and breaks in front of real users, or your Laravel app has become the thing nobody wants to touch. Both are the same job: find what is actually wrong, fix it at the cause, and leave tests so it stays fixed. 9+ years shipping production systems, now lead AI/ML engineer across three concurrent client engagements: LLM guardrails and evaluation gates for regulated users, and the Laravel, Node.js and Python backends underneath. WHERE I'M DIFFERENT Anyone can call an API. The hard part is making a model safe to put in front of a patient, a customer or a regulator: • Guardrails that hold structurally — a schema walker that fails the build if any registered model call site can return a number, plus an import fence making one module the only path to a model. The model may select, phrase or explain. It never computes. • LLM-as-reviewer gating every generated clinical report against a fixed rubric — groundedness, non-diagnostic safety, plain language, completeness — kept off the serving hot path and degrading safely when no provider is available. • A four-layer generation pipeline keeping raw patient features out of every final prompt, with per-layer token and reasoning budgets. • Multi-provider across Vertex AI (Gemini), OpenAI and Anthropic, with fallback. WHAT I BUILD • LLM features — RAG over your own data, AI agents, document and report generation • Evaluation and guardrail layers — rubric scoring, structured pass/fail, regression gates • ML models and inference APIs — FastAPI, Docker, GCP, with a versioned model registry and reproducible rollbacks • Data-protection engineering — GDPR and PDPL: data-subject rights, consent, retention, RoPA • The full stack around all of it — Node.js, Next.js, Laravel, PostgreSQL, Terraform, AWS/GCP RESULTS • Rankings endpoint 7s → 300ms — 23x on the product's hottest path • Infrastructure spend cut ~85% while holding availability through MLB and NFL seasonal peaks • EfficientNetB1 chest X-ray classifier at 96.1% test accuracy — public code, controlled ablation from 95.3% • A live fantasy-sports platform migrated Laravel 5.4 → 11 with a React SPA at zero downtime • 10,000+ concurrent users sustained; 99.8% uptime across a multi-service super app • 1,000+ scans processed on the versioned ML inference platform I built HOW I WORK 100% Job Success across 13 contracts and 1,100 hours. I tell you what AI can't do for your case before I tell you what it can. On ML work I want a metric and a threshold agreed before we start — without one, "done" is whatever gets decided later, and we both lose. Based in Abu Dhabi (UTC+4): a full working day of overlap with the EU and UK, and the US morning. Arabic native, English fluent. Two ways to start: 1. AI production-readiness review, USD 350 fixed, 3 days: a written report on what will break, what it costs to fix, and what to leave alone. 2. Laravel/PHP diagnosis and fix plan, USD 250 fixed, 2 days. Or send me what you're building and get a straight answer on whether it's feasible and how I'd approach it.