OP here.. yeah, these are all great suggestions and we considered them at length (even validated some)... the problem is, before GPT, building a deployable (i.e. not just academic paper but something that actually works) NLP pipeline was fairly niche and hard and now the barriers to entry are almost zero (but over time it's definitely trending towards zero) .. we're a small bootstrapped company doing this on the side (until we can afford to switch over).. we can't afford to compete against 100x new companies doing this (some with way deeper pockets bc either they are VC or daddy funded).. so we're not going to.. also see my response about ms copilot etc above..
