Polaris is a safety-focused LLM constellation designed for real-time patient-AI healthcare conversations. Unlike prior healthcare LLMs focused on tasks like question answering, Polaris focuses on long, multi-turn voice conversations. The system consists of a primary agent driving patient-friendly conversations and several specialist agents handling healthcare tasks like medication management, lab analysis, and nutrition. A sophisticated training protocol enables iterative co-training of agents to optimize for diverse objectives. Models are trained on proprietary data, clinical care plans, and medical documents, and aligned to speak like medical professionals using organic and simulated healthcare conversations. The system is evaluated by over 1100 U.S. licensed nurses and 130 U.S. licensed physicians, showing performance comparable to human nurses on medical safety, clinical readiness, patient education, and bedside manner. Specialist agents outperform GPT-4 and LLaMA-2 70B in healthcare tasks. The system includes privacy and compliance, checklist, medication, labs and vitals, nutrition, and hospital policy specialists. Evaluation shows Polaris performs well in clinical tasks and patient interactions, with a focus on safety and accuracy. The architecture ensures real-time, voice-based conversations with built-in safety guardrails and human supervision. The system is designed for healthcare workforce augmentation, reducing staffing shortages and improving patient outcomes through safe, empathetic, and accurate AI interactions.Polaris is a safety-focused LLM constellation designed for real-time patient-AI healthcare conversations. Unlike prior healthcare LLMs focused on tasks like question answering, Polaris focuses on long, multi-turn voice conversations. The system consists of a primary agent driving patient-friendly conversations and several specialist agents handling healthcare tasks like medication management, lab analysis, and nutrition. A sophisticated training protocol enables iterative co-training of agents to optimize for diverse objectives. Models are trained on proprietary data, clinical care plans, and medical documents, and aligned to speak like medical professionals using organic and simulated healthcare conversations. The system is evaluated by over 1100 U.S. licensed nurses and 130 U.S. licensed physicians, showing performance comparable to human nurses on medical safety, clinical readiness, patient education, and bedside manner. Specialist agents outperform GPT-4 and LLaMA-2 70B in healthcare tasks. The system includes privacy and compliance, checklist, medication, labs and vitals, nutrition, and hospital policy specialists. Evaluation shows Polaris performs well in clinical tasks and patient interactions, with a focus on safety and accuracy. The architecture ensures real-time, voice-based conversations with built-in safety guardrails and human supervision. The system is designed for healthcare workforce augmentation, reducing staffing shortages and improving patient outcomes through safe, empathetic, and accurate AI interactions.