Esther Eshiet

Evaluate AI-generated responses for factual accuracy, regulatory correctness, safety, and practical usefulness across government and non-profit domains. Write realistic prompts reflecting how professionals and consumers seek domain-specific guidance in public sector and development contexts. Identify fabricated claims, incorrect references, and misleading reasoning across model outputs, providing structured written justifications with specific evidence. Score and rank multiple model responses using structured rubrics across quality, accuracy, and safety dimensions to support frontier model training.

Reviewed AI-generated outputs for safety, bias, accuracy, and instruction adherence, applying structured ethical and compliance criteria to each evaluation. Detected fabricated claims, hallucinated references, and misleading reasoning in model outputs, flagging content that could be harmful, inaccurate, or inappropriate. Authored realistic domain-specific prompts and provided structured written feedback on policy-related and safety issues across diverse content types. Participated in calibration and QA processes, identifying edge cases and bias risks to improve model alignment with human values.

Classified and evaluated 1,000+ AI outputs against fairness, accuracy, and compliance criteria, maintaining high consistency across large-volume review cycles. Identified non-compliant, misleading, and potentially harmful outputs across diverse topics, providing evidence-based written justifications for each decision. Applied rule-based evaluation standards to detect outputs that violated safety and ethical guidelines, directly supporting model alignment efforts. Flagged inconsistencies and edge cases to strengthen dataset reliability and improve model performance across sensitive content areas.