The paper presents the VULMA project, as a machine learning framework for estimating a simplified seismic vulnerability index for existing buildings by exploiting photographs. In detail, VULMA, acronym of VULnerability analysis using MAchine learning, is characterized by four consecutive modules, organized to be part of a specific processing pipeline that allow to train, test and use the tool. In detail, the first module is Street VULMA, which allow to systematically download photographs from web services (e.g., Google Street View); the second module is Data VULMA, which consists in the tool for detecting structural features on the photographs and storing those in a database; the third module is Bi VULMA, which allow to train different machine-learning models on the data previously collected; the fourth module is In VULMA, which assign a vulnerability index to a building basing on the detected features. The methodology has been applied on an initial database of photographs regarding reinforced concrete and masonry buildings, showing to be a good and fast way to perform a first screening of existing building portfolios and providing an alternative and new method for developing risk mitigation strategies.