AMPLIFY: A Modular LLM Platform for Mission-Critical Retrieval and Reasoning

This memorandum describes AMPLIFY, NASA’s compound AI system, designed as a flexible, all-in-one platform for integrating, managing, and scaling large language models (LLMs). It supports composable workflows across NASA-hosted and third-party models, with optional use of external models for non-sensitive queries. The initial use case addresses the challenges that air traffic controllers face in retrieving information from large repositories of technical documentation in a short amount of time. The system applies a retrieval-augmented generation (RAG) architecture, provides observability through Langfuse tracing, and production-grade evaluation via VALOR. A novel element of AMPLIFY is its use of a Model Context Protocol combined with a Model Router, enabling seamless orchestration and dynamic selection of the most appropriate models for each query. Statistical metrics that measure the precision and reliability of our system show that AMPLIFY can deliver accurate and context-aware responses to user-entered queries. Future work will investigate a multi-agent, multi-step architecture to enhance modularity, interpretability, and governance.