An Agentic Machine Learning Pipeline for Drift-Aware Debris Flow Detection

Abstract

Monitoring destructive debris flows using seismic data and machine learning is a novel protective strategy. However, traditional supervised pipelines often fail during operational deployment because data distributions change over time, a phenomenon known as concept drift. Selecting a suitable model to overcome concept drift presents a significant challenge. This study introduces an Agentic Spec Kit framework that enables agentic AI to explore and select from various machine learning frameworks. We combine development driven by a spec kit with an agentic tree search to automatically design, implement, and evaluate a complete hazard detection stack. This work provides three advances: (1) a reproducible, high-performance pipeline that isolates dependencies; (2) a drift aware modeling protocol employing strict time-based splits and distributional shift monitoring; and (3) a program synthesizing search where an AI agent iteratively refines code to meet operational constraints. We demonstrate this system using the Illgraben, Switzerland seismic dataset, which features minute resolution data and a significant class imbalance of approximately 340 to 1. The framework automatically exports all artifacts, including metrics, plots, and the best runnable program. This process creates a ruggedized, field ready solution for environmental seismology.