Posts

“Today’s AI systems accomplish far more than providing a single prediction given an input,” says Brown CS faculty member Deepti Raghavan. “They’re typically compound, multi-step systems, and interactions between models and external tools enable complex use cases such as synthesizing reports, producing and testing code, or answering questions based on a corpus of evidence. However, deploying compound AI systems is difficult and resource-intensive.” 

To address this problem, she’s just received a National Science Foundation (NSF) CAREER Award to develop methods to serve and train compound AI systems efficiently. The ultimate goal is to produce new systems that enable practitioners to focus solely on creating and improving the accuracy of modern AI systems, without having to worry about details related to efficient execution and resource management. CAREER Awards are given in support of outstanding junior faculty teacher-scholars who excel at research, education, and integration of the two within the context of an organizational mission.

The project will pursue scientific advancements around three thrusts: 

1. Creating programming models to represent modern compound AI systems, particularly paying attention to pipelining and parallelism opportunities.

2. Contributing techniques to flexibly manage GPU compute and memory resources across compound AI components under dynamic load.

3. Helping practitioners observe and improve the accuracy of compound AI systems.

“I want this project,” Deepti tells us, “to result in software tools that allow practitioners to create, deploy, observe, and control compound AI systems more easily, and with fewer resources, enabling advances across the sciences.” The project will also train undergraduate and graduate students to build and reason about performance, resource consumption, and accuracy tradeoffs in modern machine learning systems.

A member of the Brown CS faculty since 2024, Deepti’s research is devoted to machine learning systems, networked systems, and operating systems. In particular, she focuses on developing new systems abstractions that enable new opportunities for performance, particularly by modifying APIs that affect when, how, and where data is moved. Her recent teaching includes CSCI 1390 Systems for Machine Learning, CSCI 0300 Fundamentals of Computer Systems (co-taught with Malte Schwarzkopf), and CSCI 2690 Cloud and Datacenter Operating Systems. CSCI 1390, a class she created, is one of the first undergraduate-level courses to offer hands-on assignments for students to explore tradeoffs through the modern machine learning systems stack. Her latest research includes “ALTO: An Efficient Network Orchestrator for Compound AI Systems”, “Towards Practically-Secure Tools for AI Agents”, and “Cornflakes: Zero-Copy Serialization for Microsecond-Scale Networking”.

Deepti joins numerous previous Brown CS winners of the award, including (in the past year alone) Nikos Vasilakis, Ellis Hershkowitz, Serena Booth, and Chen Sun.

For more information, click the link that follows to contact Brown CS Communications Manager Jesse C. Polhemus.