Preshanth Jagannathan
I am preshanth jagannathan and I work in the algorithms research and development group here at NRAO in socorro. I specialize in radio astronomy, software development, and data analysis. My current work focuses on developing and applying advanced radio interferometric data analysis techniques towards polarimetric imaging.
Sessions
Radio interferometric pipelines encode expert judgment as fixed heuristics: reasonable defaults for a "typical" dataset. When a dataset doesn't fit the mold, an expert inspects the diagnostics and steps in by hand. That doesn't scale as surveys push toward higher data volumes and less human oversight. LLMs can close that gap. They read the same diagnostics an expert would and reason about them per dataset. But letting a model both reason and act is risky. LLMs hallucinate, drift, and skip steps silently. We built an architecture that strictly separates measurement from reasoning. The model gets room to adapt, but it never touches the data path directly.
An MCP layer wraps the reduction software (CASA, here) and exposes each Measurement Set operation (metadata queries, instrument geometry, calibration, imaging) as an independent tool. Every tool returns structured data with explicit completeness and provenance. None of them interpret their own output or call each other. Reasoning lives outside the tools, in skills: version-controlled, plain-text documents that encode interferometric expertise. They're fed into the LLM's context stage by stage, so it can reason about what the tools hand back.
An orchestration layer walks the model through a sequence of stages that looks like a processing pipeline, except the parameters aren't fixed. Skill-based reasoning lets the model make an informed, per-dataset call instead of falling back on defaults. Every stage's state, its parameter choices, and the sequencing are written out as documentation and a reproducible script. Since reasoning lives outside the MCP layer, the orchestrator doesn't care which model is driving it. Cloud (Claude, Codex) or local and open (Gemma, Qwen): both work. A model equipped with real domain expertise, via skills, beats a one-size-fits-all pipeline, dataset by dataset.
We demonstrate end-to-end inspection and calibration on VLA, GMRT, and ALMA data. We discuss extending the approach to other instruments, facilities, and pipelines.
Faraday rotation along a line of sight is recovered either by Faraday synthesis, which transforms the Stokes Q and U spectra into Faraday depth space, or by QU-fitting, which fits parametric depolarization models directly to Q(λ²) and U(λ²). QU-fitting returns calibrated uncertainties and handles model comparison naturally, but it costs one full Bayesian inference per sightline, typically 10⁴ to 10⁶ forward-model evaluations under a sampler. At survey scale this is prohibitive.
VROOM-SBI removes the sampler. Simulation-based inference never evaluates a likelihood. We draw parameters from the prior, push them through the same forward model QU-fitting already uses (depolarization model, frequency coverage, noise realization), and keep the resulting parameter and spectrum pairs. A neural density estimator learns the mapping from the simulations themselves, so anything the simulator can generate can be inferred, including instrumental effects that are awkward to write into an analytic likelihood.
The specific method is neural posterior estimation. A conditional normalizing flow q(θ | Q, U) is trained on the simulated pairs by maximizing the log-density of the true parameters given their spectra, so the network targets the posterior directly rather than a likelihood or a summary-statistic distance. Training is paid once per depolarization model and frequency setup. Inference on a new sightline is then a single forward pass returning a full posterior in milliseconds. This amortization is the entire source of the 500x speedup over traditional QU-fitting. Our cost scales with the number of models trained, not the number of pixels.
Recovered posteriors agree with Faraday synthesis and classical QU-fitting carried out using RM-Tools, with coverage validated by Tarp testing and simulation based calibration, at roughly 500× lower per-sightline cost. We have released the code on github (https://github.com/skunkworks-ra/vroom-sbi) and trained models via HuggingFace (https://huggingface.co/skunkworks-ra/vroomsbi) currently for L-Band at the VLA.