Over time, the soil's nutrient level gradually declines. Soil scientists recommend precise doses of fertilizers to compensate for soil nutrient loss. However, soil scientists are scarce, expensive, and inaccessible to marginal farmers in most countries. As a common practice, rural farmers use chemical fertilizers in blind doses without proper scientific knowledge. Such indiscriminate use of chemicals creates a nutrient imbalance and leads to huge crop losses. This work aims to provide a low-cost and near-expert-level recommendation system for three fertilizers, nitrogen (N), phosphorous (P), and potassium (K), for two major crops, paddy and potato, cultivated in the Gangetic alluvial plain of West Bengal, India. We designed the system using a light gradient boost regressor, one of the most preferred machine learning methods for solving various soil-related issues, to suggest the precise doses of N, P, and K. Our designed system recommends fertilizers based on the nutrient contents and other relevant parameters in the arable soil. Experimental results revealed that the system achieved the highest performance (with R2 = 0.9997 and RMSE = 0.9381). The proposed system provides an elegant alternative to the scarce and expensive soil scientists who recommend the precise dose of the appropriate fertilizer