Senior Data Scientist, AI Evaluation & Alignment (OpenAI Human Data Project via Aligned Labs)
TextI designed adversarial multi-step biomedical reasoning benchmarks to serve as expert training and evaluation data for frontier RLHF pipelines. These benchmarks probed failure modes of advanced AI models and were directly used in AI model training and evaluation. I consistently produced scientific reasoning problems that state-of-the-art models were unable to solve, supporting research in human data and model alignment. • Created original question datasets targeting model weaknesses in biomedical reasoning. • Labeled and evaluated multi-step answer derivations from AI systems using domain-specific criteria. • Developed failure-mode probes and adversarial benchmarks for RLHF (Reinforcement Learning from Human Feedback). • Used proprietary tooling and/or internal evaluation platforms for data labeling and benchmark administration.