I've implemented ML-KEM by the spec as an exercise recently (https://github.com/AlexanderYastrebov/mlkem) and here are related links that helped me understand the math:
* [Enough Polynomials and Linear Algebra to Implement Kyber](https://words.filippo.io/kyber-math/)
* [Basic Lattice Cryptography. The concepts behind Kyber (ML-KEM) and Dilithium (ML-DSA)](https://eprint.iacr.org/2024/1287.pdf)
* [A Complete Beginner Guide to the Number Theoretic Transform (NTT)](https://eprint.iacr.org/2024/585.pdf)
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).