Sustainable groundwater management in semi-arid regions of North Africa is constrained by fragmented climatic records, which undermine recharge estimation and resilience planning. This study introduces a reproducible artificial intelligence pipeline that integrates machine learning and deep learning to reconstruct missing precipitation data and enable forecasting to support groundwater management. The novelty lies in the first application of the Self-Attention-based Imputation for Time Series (SAITS) model for filling precipitation data gaps in North Africa, integrated into a seamless imputation-forecasting workflow. SAITS was used to reconstruct a monthly precipitation record over the span of 74 years from three stations in Tunisia's Guenniche Catchment and was benchmarked against Random Forest and Support Vector Regression under synthetic missingness scenarios of 15-25%. SAITS achieved superior accuracy (RMSE = 16.88; MAE = 4.06; R-2 = 0.89; KGE = 0.91; ER95= 0.91 at 15% misingness) and remained robust under larger gaps. The reconstructed dataset was subsequently used to forecast precipitation to 2030 using Long Short-Term Memory (LSTM) and Transformer models. Forecasting performance varied by station: the Transformer captured long-term trends more effectively (El-Alia, R-2 = 0.82), whereas LSTM reproduced short-term variability more accurately (El-Nechrine, RMSE= 20.62). The Python-based pipeline provides a transparent and transferable framework for transforming incomplete records into actionable hydrological insights, reducing uncertainty in recharge estimates, guiding managed aquifer recharge planning, and informing allocation policies. Beyond Tunisia, the approach is applicable to other data-scarce, climate-vulnerable catchments, offering a practical tool for advancing water security toward Sustainable Development Goal 6.