Computer numbers are usually limited, both in range and in precision. As a consequence, a careful certification has to be performed for applications that compute with these sets of numbers. Unfortunately, performing such a certification by hand is error-prone. Formal methods can ensure that the certification is correct, but making use of them is usually long and tedious, even for experts.This thesis aims at improving the availability of these methods to developers by automatizing their implementation. The key concepts are the use of interval arithmetic, a database of theorems on computer arithmetics, and a system for rewriting expressions in order to compute tight bounds on rounding errors.This approach has led to the development of the Gap...