Contract Evaluator, Handshake AI (Project Escher)
Text
Red TeamingDesigned adversarial visual reasoning prompts to test multimodal AI model failure modes in chart, diagram, and infographic counting, color discrimination, and multi-region synthesis. Authored carefully constrained task submissions with determinate answers and strict output formatting to meet reviewer quality criteria. Produced structured rationales and evaluation outputs to inform model training signals based on observed error clusters and qualitative scoring dimensions. •Created adversarial prompt sets targeting specific weaknesses in multimodal visual understanding. •Ensured submissions followed format-pinned outputs and used copyright-cleared sources. •Performed evaluation across naturalness, conversational dynamics, and emotional attunement dimensions. •Delivered structured rationales and error analyses usable as training/evaluation feedback.