Bank statement extractor

The brief

Loan officers retype transactions from customers' PDF statements to check income and spending. Build an extractor that reads statements from different banks into validated JSON transactions with categories, and flags rows it is unsure about. Good looks like: field accuracy measured on hand-labelled statements, validation that catches impossible values, and a review list of unsure rows.

What you will build

1. Create sample statements in several layouts
2. Extract transactions with a vision-capable LLM and structured outputs
3. Validate dates, amounts and running balances
4. Measure accuracy per field on a labelled set
5. Flag unsure rows for review and report cost per statement

Free, open data to build it with

• CORD: receipt understanding dataset ↗ Clova AI (GitHub)

Check each dataset's licence before you publish your work. Your mentor gives your team its own version of the brief (data, scope or region), so every team's project is its own.