As much as I love simple deterministic things, this is a classic example of where NLP is better than hardcoding a list of keywords. Trying to guess every set of quantity keywords with various spelling, punctuation and how they interact ("1dzn box, two pack" is actually 24 balls) seems more brittle than an LLM.

You're probably right. I can see how LLM is the better way to approach this. I haven't looked into AI usage fees or anything, but I would think the amount of queries I'd send wouldn't be that expensive. I'll look into it.