Sepsis, a leading cause of hospital mortality, is characterized by substantial heterogeneity, hindering the development of effective and interpretable prognostic and stratification methods. To address this challenge, we developed an explainable prognostic model (SepsisFormer, a transformer-based deep neural network with an enhanced domain-adaptive generator) and an automated risk stratification tool (SMART, a scorecard consistent with medical knowledge). In a multi-center retrospective study of 12,408 sepsis patients, SepsisFormer achieved high predictive accuracy (AUC: 0.9301, sensitivity: 0.9346, and specificity: 0.8312). SMART (AUC: 0.7360) surpassed most established scoring systems. Seven coagulation-inflammatory routine laboratory measurements and patient age were identified to classify patients' four risk levels (mild, moderate, severe, dangerous) and two subphenotypes (CIS1 and CIS2), each with distinct clinical characteristics and mortality rates. Notably, patients with moderate /severe levels or CIS2 derive more significant benefits from anticoagulant treatment. Our work, therefore, offers a novel set of simple, real-time executable tools for sepsis heterogeneity, demonstrating considerable potential to significantly enhance sepsis clinical practice globally, particularly in resource-constrained healthcare settings.