The system begins by collecting relevant information from the user, including age, ailments, medications, and concerns (e.g., cost sensitivity, access to pharmacies, maintaining continuity of care, or transportation challenges). Geolocation data, such as the user’s county of residence, identifies the available Medicare plans, which are further enriched with CMS data, including premium costs, Star ratings, enrollment statistics, and formulary information. This comprehensive dataset is indexed into a vector database, enabling efficient and semantic searches.

For each plan under consideration, an LLM is “”spawned”” to evaluate its suitability for the user. A “”referee”” LLM monitors for potential biases in recommendations, fostering an adversarial‑cooperative dialog among the LLMs. The agents collaboratively analyze the user’s specific needs, advocating for or against their assigned plans. Plans deemed unsuitable are discarded early, especially when mismatched with non‑negotiable criteria (e.g., premium costs or required coverage).

The final recommendations are presented to the user, highlighting how the selected plans address their concerns. For added transparency, the system optionally displays the dialog among LLMs, allowing users to trace the “”thought process”” that led to the recommendations. This design not only personalizes the selection process but also increases trust and understanding, empowering seniors to make informed healthcare decisions.

This innovative application of adversarial‑cooperative LLMs bridges the gap between complex data‑driven insights and user‑centric healthcare needs, offering a scalable solution to Medicare plan selection challenges. The proposed system demonstrates how AI can transform healthcare analytics, combining state‑of‑the‑art technologies with practical, user-focused outcomes.