Does a matched filter really improve SNR?

Does the matched filter actually improve SNR or does it maximize it (making sure that it is not less than what it should be) ? improving SNR would not make sense to me as suppose a receiver has a certain NF that defines an SNR for a signal at a certain power level and a certain BW, improving SNR would mean reducing NF, and that wouldn't make sense, would it?

Assuming a pulse shaped BPSK signal with a symbol or bit rate =1/BW (a roll off almost equal to zero), the SNR at the matched filter output =Eb/No, multiplying the denominator and the numerator by BW we get: signal power/ noise power and noise power would of course be a function of NF, so in that case where is the improvement? or do you want to say that already assuming that signal BW=1/rate because of pulse shaping is the condition that did the job?