First, let us remark that there exist several hundred read mappers, most of which have been even published (see, e.g., pages 25-29 of this thesis). Developing a new mapper probably makes sense only as a programming exercise. Whereas developing a quick proof-of-concept read mapper is usually easy, turning it into a real competitor of existing and well-tuned mappers can last for years.
It is not clear from the provided description how long is your reference, how many alignments you need to compute, etc. In certain situations, it may be useful to write a wrapper over existing mappers (e.g., using subprocess.Popen for running the mapper and PySam for parsing the output); while in some other situations, standard dynamic programming may be sufficient (e.g., using SSW with its Python binding).
Let assume that you want to develop a toy read mapper. Most of read mappers are based on a so called seed-and-extend paradigm. Simply speaking, first you detect candidates for alignments, usually as exact matches between a read and the reference (using either a hash table or some full-text index – e.g., BWT-index). Then you would need to compute alignments for these candidates, typically using some algorithm based on dynamic programming, and report the best obtained alignments (e.g., the ones with the highest alignment score).
There exist two big, powerful and well debugged libraries implementing BWT-indexes which can be easily used for building a read mapper: