Younes T.

AI safety red teamer and bilingual model evaluator. I find the prompts that make frontier AI models fail — jailbreaks, harmful outputs, policy violations — in both Arabic and English, then document exactly how to reproduce them. 100% Job Success, 892+ hours. WHY BILINGUAL RED TEAMING MATTERS Most adversarial testing happens in English. But safety guardrails that hold in English often collapse in Arabic, or when an attack is framed inside a specific cultural or regional context. As a native Arabic speaker (MSA + Algerian dialect) with a Master's degree in English Language and Literature, I test both sides of that gap — and I write findings up in clear, precise English your team can act on. WHAT I DO AI Safety & Red Teaming - Adversarial prompting, jailbreak and prompt-injection testing - Multi-turn attack chains and harmful-content elicitation - Refusal and over-refusal analysis, safety policy application - Severity classification and structured vulnerability reporting Model Evaluation & Training Data - RLHF preference ranking and response comparison - Factuality and hallucination detection - Rubric-based scoring and instruction-following assessment - Prompt writing, data annotation, and annotation QA Arabic Language & Localization - English ↔ Arabic translation — technical, marketing, academic - Culturally grounded adversarial test case design - Localization review and transcreation Technical & Content Writing - Technical documentation and editorial review - SEO articles, blog posts, and long-form web content CURRENT WORK I red-team frontier models through Mercor and train and evaluate model outputs with Invisible Technologies. Before that: eight years of continuous contract work on Upwork evaluating AI-generated content, with a 100% Job Success Score and long-term repeat clients rather than one-off jobs. WHY CLIENTS KEEP ME - Native Arabic fluency paired with Master's-level English — a rare combination in AI evaluation - Findings written clearly enough to act on without follow-up questions - Fast response, reliable delivery, no missed deadlines across 892+ hours Send me a message with what you're building and what you need tested. I'll tell you honestly whether I'm the right fit.