A 12M-parameter GPT-style transformer trained on Reddit doctor-patient Q&A (~6,000 interactions). Uses BPE subword tokenization. Generates plausible-sounding medical dialogue — not real medical advice.