Project Grant 2541610

The National Science Foundation Division of Information and Intelligent Systems awarded Brown University $447,980 on October 1, 2026, under the Computer and Information Science and Engineering program (CFDA 47.070) to develop runtime methods for efficient and observable compound AI systems.

The project creates programming models, compiler techniques, and resource management approaches that enable practitioners to deploy multi-step AI systems—which combine learned components such as language models with non-learned components including databases, document retrieval systems, web search, and code execution—without manual optimization or inefficient component duct-taping. The research pursues three thrusts: a streaming dataflow model that automatically identifies pipelining and parallelism opportunities in compound AI workflows; an efficient execution runtime; and resource management techniques that help developers observe and improve end-to-end system accuracy. The work focuses on exposing incremental processing capabilities in common AI components to reduce the expense of repeated model calls and tool invocations in complex use cases such as report synthesis, code production and testing, and evidence-based question answering.

The award period runs through September 30, 2031, with work performed in Providence, Rhode Island.