Dating apps shape romantic opportunities but remain inaccessible to researchers due to commercial secrecy and privacy concerns. We developed an alternative approach using authentic user profiles and Large Language Model agents,powered by GPT-4o to simulate dating decisions in a controlled environment. These agents engaged in 80,000+ simulated interactions, generating both choices and explanations for their preferences. Our supervised machine learning analysis (regression models, decision trees) and unsupervised semantic clustering revealed consistent patterns: education emerged as the dominant predictor of dating desirability across demographic groups. Gender differences were particularly pronounced in evaluation hierarchies—female agents prioritized income in male profiles, while male agents emphasized age when assessing females. Control experiments demonstrated high behavioral fidelity, with agents consistently acting according to their assigned demographic attributes and producing reasoning patterns that closely mirror documented human mate selection criteria. This methodology offers a transparent believable method for studying bias patterns in social decision making typically hidden behind proprietary algorithms.