9-10 years ago I was in a smaller structure and basically doing close to SOTA work in the field of NLP: contributions to `spacy`, some of the best word2vec models models for our language (french) and good english ones, at least models that crushed other word2vec models of the time on our large corporate benchmarks (both french and english), and were on par (english) and sometimes better (french) than the official benchmarks of the time. Mostly through minor model tweaks but major data sanitation and preparation work. And then custom transformer-based architecture for other NLP tasks that were also better than all off-the-shelf NLP solutions of the time. And even on my free time I could do some cool ML/NLP stuff no one else was really doing.
Now I am in a larger structure with more resources, yet I know that both in terms or resources and honestly to a degree in terms of skill I feel left out of the current NLP/AI SOTA landscape.
It sounds like you were already doing statistical NLP (the thing that has been clobbering everything rule-based in performance) but have seen that field itself start to require much greater resources than what you were working with.