This thesis examines whether AI assistance improves student performance in medical triage simulation training. Thirteen students completed three disaster scenarios twice each (76 simulator sessions). Six had a multimodal AI assistant that could see the simulator screen and answer spoken questions; seven controls narrated reasoning aloud without AI. Both groups improved significantly in score and accuracy (p < .05). Performance did not differ by condition (p = .98), indicating no added benefit from AI beyond practice, though the small sample size (n = 13) limits the power to detect such effects. The AI group reported higher subjective workload (d = 1.02, p = .11). Controls produced nearly twice as many question-like utterances per round (p = .18). The AI assistant ran on a custom multimodal AI pipeline connecting to OpenAI's Realtime API, grounded in triage protocols via retrieval-augmented generation and built to support multiple concurrent sessions.