acceptodds
Under review as a conference paper at ICLR 2027

Modular Tooling and Adaptive Spatial MLOps: An MCP-Enabled Agentic Delivery of Natural Hazards Insights and Predictions to Multidisciplinary Stakeholders

Abstract

Extreme weather events expose human and economic frailty, of which is reliant upon a stable physical environment. To minimize the societal risk incurred from these natural bouts of instability, multi-disciplinary resilience efforts, spanning science, engineering, policymaking, and more, are essential. The issue, however, arises when the needed collaborative efforts are impeded by the inability to derive findings outside of one’s own expertise or to discern the outputs and outcomes of another. Hence, this work novelly proposes a shared space solution, an adaptable, flexible and automated agentic-enabled application, to facilitate collaboration among experts and stakeholders by the provision of practical climate resilience content. The application presently affords physical and social characteristics, extreme-event risk information, interactive and downloadable maps, predictive models, and resilience-planning tools, along a modular architecture designed for the scalable incorporation of additional capabilities. Operationally, via the interface, the end-user may either directly access an instrument or select an agent to orchestrate instruments via a Model Context Protocol (MCP) tool layer, where the Large Language Model (LLM) interprets requests, selects and invokes tools, and synthesizes outputs for the end-user. Further, two implemented capabilities are highlighted in this paper. First, a flood susceptibility prediction model, running via a Machine Learning Operations (MLOps) lifecycle, is demonstrated. The automation includes the weekly consumption of new radar and sensor observations, triggering the retraining and evaluations of graph convolutional neural network and logistic models. Models and sensors are then dynamically selected based on predefined performance criteria. At the daily scale, same- and next-day precipitation forecasts are ingested, producing localized 1km x 1km predictions. A second highlight is the LLM-enabled multi-risk tool, which integrates physical, extreme-event risk, and sociodemographic information, allowing users to identify areas exhibiting combinations of vulnerabilities relevant to their particular resilience inquiries. Ultimately, by providing the stakeholder facing application, the immediate aim is to assist New Yorkers in preparing for and mitigating the consequences of extreme weather events, while providing a foundational framework that may be extended to other urban environments.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Agentic AI, Model Context Protocol (MCP), MLOps, Natural Hazards, Extreme Events, Climate Resilience, Flood Forecasting, Spatial Machine Learning, Graph Neural Networks, Graph Convolutional Networks, Logistic Regression, Unsupervised Clustering, Geospatial Analysis, Multidisciplinary Collaboration, Decision Support

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_wawcOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work presents the NYC Resilience Application, a tool designed to provide the end-user with a unified interface for assessing or monitoring natural hazards from diverse data sources.

Critical strengths

The paper does not present any contributions suitable for a major machine learning conference, such as ICLR

Critical weaknesses

The paper is not suitable for publication at ICLR. It doesn't propose any contributions that would advance the application of machine learning. Instead, it is more of a web (or UI) application development project than an academic research paper.

Questions for authors

None. Paper not suitable for publication in an academic machine learning conference.

Reviewer_wmSvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a New York City climate-resilience application integrating geospatial and sociodemographic data, risk queries, planning tools, and flood prediction. It uses weekly retraining and per-sensor model selection, with selected capabilities exposed through MCP for LLM orchestration. The paper primarily provides operational workflows, interface examples, and model-selection and spatial-support metrics from one representative weekly cycle.

Critical strengths

Clear users and application needs. The application brings urban risk data, local forecasts, and planning tools into one interface for multidisciplinary participants and city stakeholders, providing a concrete application setting. A substantial operational modeling workflow. The paper describes observation ingestion, weekly retraining, eligibility screening, spatial-support assignment, and daily forecasting, with monitoring metrics listed in the appendix.

Critical weaknesses

Spatial generalization beyond supporting sensors needs evaluation. There are useful checks, but the grid-weighted support F1 is not an independently measured error at the 750 imputed grids out of 837. Please evaluate transfer to held-out sensors or areas and report directly supported and imputed predictions separately, to assess the spatial generalization used by the city-wide product. Systematic evaluation of the agentic service is still missing. The paper documents the deployed orchestration service and execution tracing, but explicitly leaves tool-selection, parameter-extraction, multi-tool reasoning, grounding, and response-quality evaluation to future work. Thus, implementation is demonstrated, while reliable completion of stakeholder requests is not yet systematically assessed. Please evaluate representative requests with specified expected outcomes and report tool-use correctness, task completion, and grounding.

Questions for authors

see weakness part.

Reviewer_Po7YOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

Systems paper about deployed NYC Resilience Application: dashboard with census and FEMA risk data, LLM tool that converts natural language to tract filters (Claude makes JSON, deterministic code does geo work), automated flood pipeline (FloodNet sensors + MRMS rain feeding GCN and logistic regression, weekly retraining, sensor eligibility gates, k-means grid imputation, daily forecasts), and MCP agent layer (FastAPI, Claude Haiku, five tool services, Phoenix tracing).

I have concerns about the paper contribution, outlined below.

Critical strengths
  • It is deployed and running. Weekly retraining with real governance on top, rare to see.
  • Correct design instincts: LLM never invents operational values, MRMS for training vs HRRR for operations, imputed grids kept out of retraining
Critical weaknesses
  • No methodological novelty. Standard parts (GCN, logistic regression, k-means, MLflow, off-the-shelf LLMs) put together. Related work comparison is mostly positioning. Good systems report, weak ICLR contribution.
  • The MCP agent layer is the title's selling point but has zero quantitative evaluation. No tool selection accuracy, no latency, no failure taxonomy.
  • Flood evidence is one anecdotal event plus one MLflow snapshot. Mean selected F1 0.492. No held-out event evaluation, no trivial baseline. And 750 of 837 grids (89%) are imputed, no direct vs imputed comparison.
  • Thresholds (precision 0.25, recall 0.30, K=5, k=4, 25.4 mm) not justified anywhere.
  • Reproducibility limited: Bedrock, ArcGIS, live APIs.
  • Writing needs cleanup (wass, coud, "novelly, etc)
  • System gives flood forecasts to public, but no uncertainty discussion, ethics statement is one line
Questions for authors
  1. Held-out event-level performance per sensor?
  2. Direct vs imputed grid accuracy, is this shown to users?
  3. Any logged numbers on agent layer: tool success rate, calls per request, latency?
  4. How stable is eligibility week to week?
  5. Why k=4 neighbors, K=5 clusters?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.