At Open Train AI, I evaluated and ranked Arabic and English LLM outputs against rubrics for factuality, instruction following, tone, formatting, and safety. I also fact-checked claims against primary sources and wrote evidence-based rationales for model preference rankings and error classifications.
I designed complex prompts and gold-standard responses for supervised fine-tuning, and audited bilingual data for linguistic naturalness and cultural appropriateness. I also used adversarial prompts to test safety guardrails and report root causes of biased, unsafe, or policy-violating outputs.
