"Cloud Conversations" — an AI assistant that builds research clouds by chatting (NSF AI grant $2.54M)

The NSF awarded about $2,540,265 to "Cloud Conversations: AI-Augmented Interfaces to Research Infrastructure" (NSF Award 2609115; program: Software Institutes [Frameworks / collaborative research]; Illinois; starting October 2026).

1Why configuring cloud environments is hard

Per the abstract, cloud computing is essential to a growing number of science use cases, but configuring scientific environments for the cloud is challenging — requiring specialized knowledge in system administration, networking, and security and involving numerous configuration settings.

Many scientific workloads also depend on tightly coupled virtual clusters, specialized hardware, fast interconnects, accelerators, and custom drivers. Managing this complexity consumes time and attention researchers could otherwise devote to science.

The project develops an AI-based conversational assistant for configuring scientific computing environments: researchers describe what they need in everyday language, and it creates the environment and verifies its integrity on shared research cloud infrastructure — increasing productivity, lowering the cost of using cloud computing, and enabling practical reproducibility of computational experiments.

2Built on Chameleon and Jetstream2

Technically, it designs and deploys an AI-based agent framework that can plan, provision, and validate scientific computing environments on open research computing infrastructure such as Chameleon and Jetstream2.

The framework combines large language models running on open, high-performance academic hardware with software tools exposed through standard interfaces — cloud-based provisioning services, hardware/software environment templates, correctness checks, and a validation benchmark suite.

Key components include planning modules with built-in checks on resource limits, timing, and hardware compatibility; state- and error-handling modules that track multi-step workflows and summarize system events; and search pipelines that organize information from documentation, logs, help-desk tickets, and environment artifacts into a searchable knowledge base.

3Swapping what is asked of the researcher

Building a scientific computing environment in the cloud demands knowledge of systems administration, networking and security, and the handling of a great many configuration settings. This project does not simplify the environment; it swaps the side doing the demanding.

Tightly coupled virtual clusters, specialised hardware, fast interconnects, accelerators, custom drivers — scientific workloads depend on such conditions. The heavier the left column, the more of the time meant for science it takes.