Students in our AI/ML course submitted polished project analyses, but when asked in class to defend a single choice from their own work, many could not. The writing looked strong; the understanding often did not match it. Oral examinations retain an evidentiary link that written work has lost, yet they have never scaled: a 25-minute oral with two graders costs roughly 30 instructor and TA hours for 36 students, and is untenable at 100.
Voice AI and automated grading change that arithmetic. We built Viva by NYU Stern, which conducts a personalized oral exam and grades the transcript with a panel of three large language models that score independently, read each other’s assessments, and revise. Across two undergraduate cohorts (36 students, then 37), grading cost stayed under one dollar per exam; above a comparable subscription it budgets to about a dollar per ten graded minutes, which is practical for weekly assignments rather than finals alone.
The system also broke instructively. Question stacking, non-random topic selection, and an unexpectedly harsh cloned voice each yielded practical engineering lessons for any conversational system that delegates behavior to a language model: decompose into single-purpose modules, constrain behavior in code rather than prompts, keep randomization out of the model, grade with a multi-model panel whose members can disagree, and treat voice as a first-class design variable.