Speaker: Krishna Sekhar, NRAO

We present an architecture for LLM-driven radio interferometric data reduction. The overarching idea is to use an interface to the LLM (a "harness") which will drive the processing via CASA. We achieve this by defining a Model Context Protocol (MCP) layer, which is a thin Python layer that wraps CASA. The MCP exposes various Measurement Set operations - metadata queries, flagging, calibration and imaging as independent tools that return structured data with explicit completeness and provenance annotations. No tool interprets or chains it's outputs to another. The reasoning is supplied via SKILL documents - version controlled natural language domain documents that encode interferometric expertise. The skills provide the context necessary for the orchestrator to define pass/fail criteria for each stage of the processing in an adaptable manner. The orchestrator uses the MCP and skills to route the model through a workflow, allowing for human checkpoints at pre-defined points for inspection. The decisions, parameters and choices are all persisted for inspection and validation, we do not blindly trust the LLM output.

We demonstrate "radio-analyst" - a Claude Code based interface to the underlying MCPs that enable natural language driven processing of radio interferometric data. We show successful calibration runs on various EVLA datasets (3C391, G55). We discuss upcoming work to develop a custom harness that allows for more flexibility (to use open-weight models, and multi-dataset orchestration).
============================
1800 UTC; 12pm MT, 1pm CT, 2pm ET
Zoom info:
https://go.nrao.edu/soaudzoom
Audio-only: +1 646 876 9923
Meeting ID: 675 659 4969
Passcode: 787787