In this paper, we illustrate a methodology to measure discrimination in educational contexts. In India, we ran an exam competition through which children compete for a large financial prize. We recruited teachers to grade the exams. We then randomly assigned child "characteristics" (age, gender, and caste) to the cover sheets of the exams to ensure that there is no systematic relationship between the characteristics observed by the teachers and the quality of the exams. We find that teachers give exams that are assigned to be lower caste scores that are about 0.03 to 0.09 standard deviations lower than exams that are assigned to be high caste. The effect is small relative to the real differences in scores between the high and lower caste children. Low-performing, low caste children and top-performing females tend to lose out the most due to discrimination. Interestingly, we find that the discrimination against low caste students is driven by low caste teachers, while teachers who belong to higher caste groups do not appear to discriminate at all. This result runs counter to the previous literature, which tends to find that individuals discriminate in favor of members of their own groups.