In many cases, the design of mechanical components is based on the experience and intuition of engineers and designers, which makes technology transfer difficult. With this work, we want to present a methodology for the automatic extraction of the know-how from a database of existing designs and the use of this know-how to create a first approximation to the design of a new mechanical component. A series of difficulties arose when developing this methodology, mainly the visualization and manipulation of high-dimensional data, such as those that define the geometries, the impossibility of interpolating between topologically different geometries, as well as the scarcity of data in the database. We use various tools to solve these difficulties, including Manifold Learning techniques, in particular the Locally Linear Embedding (LLE) technique, which allows us to reduce the dimensionality of the database. We also use Topological Data Analysis (TDA) tools, to topologically characterize each of the geometries. A final key technology was selected to compare the geometries of the database appropriately and to interpolate between them, the Wasserstein distance (instead of the Euclidean distance) which, among other benefits, reduces the number of artefacts resulting from interpolation and leads to coherent design.