I did the same. Something that helped me get my head around it was realising that NTT is mostly a performance optimization, a bit like montgomery form in RSA. You can conceptually implement ML-KEM without it, it'll just be slower (it also won't be interoperable because the wire format involves the NTT'd form - I think, it's been a while since I looked at it in detail).