This sounds so familiar. I have worked in several startups that were commercializing NLP-based solutions for different problems.
The approach of management (although relatively educated in the field, but obviously not knowledgeable in everything) was essentially expecting from their research engineers:
- look into state-of-the-art related to our problem
- think very hard about how to apply the academic state-of-the-art to our problem
- stitch a fully fletched software solution together
The problem obviously emerged reliably in the last step, which was supposed to take negligible time and should be done by more junior team members (because they were not smart enough for step 2).
The result was a bunch of prototypes running in production, and huge costs for keeping the whole thing somehow running.
- look into state-of-the-art related to our problem - think very hard about how to apply the academic state-of-the-art to our problem - stitch a fully fletched software solution together
The problem obviously emerged reliably in the last step, which was supposed to take negligible time and should be done by more junior team members (because they were not smart enough for step 2).
The result was a bunch of prototypes running in production, and huge costs for keeping the whole thing somehow running.