With the rapid growth of construction project data and increasingly prominent multi-source heterogeneity, achieving ef?cient storage and agile computing has become a critical challenge in the post-evaluation of construction projects. This paper proposes an innovative method based on ef?cient storage and agent-based agile computing using autoencoder-based discrepancy analysis, aiming to address the integration, storage, and computation of heterogeneous engineering data and enhance the accuracy and ef?ciency of post-evaluation indicators. First, to tackle the multi-source heterogeneity of construction project data, hierarchical storage architecture is designed. This architecture integrates distributed databases with lightweight caching mechanisms, enabling high-ef?ciency data compression and fast retrieval while reducing storage overhead and ensuring data integrity. Furthermore, the study introduces a dual learning mechanism that combines autoencoder networks with discrepancy analysis. By leveraging deep generative models for feature extraction and data fusion, and quantifying data distribution discrepancies through a discrepancy metric function, the method identi?es key in?uencing factors and potential anomalies. This approach not only adapts to unstructured data (such as text, images, and sensor timeseries data) but also enables dynamic adjustment of model computation parameters through agent-based autonomous optimization strategies, enhancing analytical robustness. Experiments conducted on real-world construction project datasets demonstrate that the proposed method improves storage ef?ciency by over 4000% compared to traditional relational databases and reduces computation response time by 3500%. The reliability of post-evaluation indicator optimization is also signi?cantly improved. Additionally, the discrepancy analysis module successfully detects latent data con?icts that are often missed by conventional methods, validating its practical value in complex construction project scenarios.