VineSentinel: Implementing Case-Based XAI for Expert Scrutiny in Precision Viticulture

In precision viticulture, the effective deployment of predictive systems can be hindered by the opacity of model outputs, which may challenge the active scrutiny required by domain experts in high-stakes decision-making. Drawing on Human-Centered Explainable AI principles, specifically scrutiny and actionability, we present VineSentinel, a web application designed to assist expert verification on the task of intra-season vineyard yield forecasts. The system integrates a LightGBM predictive core with two interface mechanisms tailored for situated reasoning. To enable the validation of model logic, an explainable AI (XAI) module retrieves and displays example-based explanations, aiming to align the interface with the cognitive process of analogical reasoning. Furthermore, to connect diagnostic insight with field intervention, the system leverages a large language model to generate context-aware recommendations, synthesizing available ecological constraints into prescriptive support. Hence, this work proposes a hybrid intelligence architecture designed to support expert agency in smart farming.