handle poorly processed reviews

[roworu]

if domain_model poorly classified text, we should retrain that model based on LLM response, so in future it could process similar texts better.

initial implementation plan:

1. if confidence of domain_model < threshold, store to db following:
   text, domain_model confidence, domain_model result (grade), LLM confidence, LLM result

2. prepare a scheduled job, that will take all failed classifications from DB, and retrain domain_model based on it