The unconfined compressive strength (UCS) of a stabilised soil is a major mechanical parameter in understanding and developing geomechanical models, and it can be estimated directly by either lab testing of retrieved core samples or remoulded samples. However, due to the effort, high cost and time associated with these methods, there is a need to develop a new technique for predicting UCS values in real time. An artificial intelligence paradigm of machine learning (ML) using the gradient boosting (GB) technique is applied in this study to model the unconfined compressive strength of soils stabilised by cementitious additive-enriched agro-based pozzolans. Both ML regression and multinomial classification of the UCS of the stabilised mix are investigated. Rigorous sensitivity-driven diagnostic testing is also performed to validate and provide an understanding of the intricacies of the decisions made by the algorithm. Results indicate that the well-tuned and optimised GB algorithm has a very high capacity to distinguish between positive and negative UCS categories (‘firm’, ‘very stiff’ and ‘hard’). An overall accuracy of 0.920, weighted recall rates and precision scores of 0.920 and 0.938, respectively, were produced by the GB model. Multiclass prediction in this regard shows that only 12.5% of misclassified instances was achieved. When applied to a regression problem, a coefficient of determination of approximately 0.900 and a mean error of about 0.335 were obtained, thus lending further credence to the high performance of the GB algorithm used. Finally, among the eight input features utilised as independent variables, the additives seemed to exhibit the strongest influence on the ML predictive modelling.