Every year, colleges and universities are overloaded with large volumes of applications for admissions. In response, prior studies have attempted to fully automate the review process; however, such systems may lead to undesirable outcomes for some of the applicants. In this work, we propose a human-in-the-loop evaluation framework that incorporates a machine learning model to automate the acceptance and rejection of unquestionable applicants, while proceeding borderline applicants to be passed onto the admissions committee for review. To deploy our framework in colleges, we additionally consider two critical cases: (i) applicants who may be qualified should not be rejected by our model before the human review, and (ii) the model mispredictions should not be biased towards a certain subgroup of applicants (e.g., female applicants more likely to be mispredicted than male counterparts, and vice versa). To quantify the bias properly, we also propose a novel fairness metric that measures the degree to which each subgroup is getting mispredicted by the model. We empirically demonstrate that our proposed model, which is built upon the state-of-the-art deep neural networks, surpasses the widely used baselines in terms of the number of mispredictions made and the bias of mispredictions.