Predicting Smallholder Farmer Adoption of Agri-Fintech Services in Rural India: An Ensemble Machine Learning Approach Grounded in ICT4D and Capability-Theoretic Foundations

India has witnessed an unprecedented expansion of agri-fintech services since 2020, but sustained adoption among smallholder cultivators remains stubbornly uneven. Existing diffusion-of-innovation studies do not adequately capture this micro-level heterogeneity and rarely produce predictions that policy-makers or last-mile providers can act upon. The present study addresses both the theoretical and the methodological gap. We construct a predictive analytics framework that combines four established information-systems lenses the Unified Theory of Acceptance and Use of Technology (UTAUT2), Heeks’s ICT4D design-reality gap model, Sen’s Capability Approach, and the Diffusion of Innovations theory with an ensemble of supervised machine-learning algorithms (Logistic Regression, Random Forest, XGBoost, LightGBM, and a stacked meta-learner). Cross-sectional survey data were collected from 1,842 smallholder farming households across six agro-climatic zones of Maharashtra, Madhya Pradesh, and Karnataka between October 2024 and February 2025. The stacked ensemble achieved a held-out ROC-AUC of 0.891 and balanced accuracy of 82.3 per cent, substantially outperforming the best single classifier (XGBoost, ROC-AUC 0.864). SHAP-based interpretability identified effort-expectancy, social influence from village self-help groups, and the language-localisation dimension of the design-reality gap as the three strongest drivers of adoption. The paper contributes a reproducible predictive pipeline, a theoretically grounded feature-engineering rubric, and policy-actionable insights for cooperative banks, Farmer Producer Organisations, and the Department of Agriculture and Farmers Welfare