Is $L = \{ W_1W_2 \mid W_1,W_2 \in (a+b)^* , N_a(W_1) = N_b(W_2)\}$ context free? Can we construct an NPDA for the language?

There is a book here that claims $L$ is not CF (without any elaboration), but I think we can construct a NPDA that accepts the language. My guess is we can construct the language with an NPDA where after reading some $a$ and $b$ and pushing $A$ for each $a$ into the stack, makes a guess to jump to a new state and consumes the pushed $A$ with each $b$.