Python: Invoice Entity learning from Human corrections

We have document reader based on a tesseract and scanning script that do ocr and some other stuff on a script. After the scan is done all strings and all results are sent to the user screen where it is shown to the user so the user can make corrections. This happens THE FIRST TIME ONLY. So the learning from the user in that screen then be sent back to the document reader which the next time is trained and has correctly learned for that invoice or document. The next time the document reader scan that document it has learned where those entities are located. So in short: the user corrections should then be fed back to the document reader. The human train the system on the fly in essence. We would also like to have some help with enhancing the precision to beyond 90-95% whilst scanning the document. We have tried most research reports which suggest bert,stanford ,and spacy flairNLP for NER etc. We need someone who is a SPECIALIST in this area. We have VERY experienced ML and AI developers in the area and you must be a SPECIALIST IN THIS area. Wwe do not want the collective to big here. You will not have a chance. Please add: I am a spcialist in this OCR area and can help you with directions and coding.