Fine-tune an LLM for financial document extraction

Project Description

We process a few hundred vendor invoices and bank statements a week and extract line items by hand. We want that down to a review step rather than a data entry step. The documents are Indian formats — GST invoices, bank statements from six or seven banks, all PDF, some scanned. Layouts vary a lot. What matters most to us is that the model is honest about uncertainty. We would rather it flags 20% for review than silently gets 5% wrong, because a wrong figure in a reconciliation costs us far more than a manual check. Please share something comparable you have built. Accuracy claims without a test set do not help us judge.

Skills Required