What it does
An underwriting pipeline for a lender. Statements in PDF or scanned form are read into transactions, every line is categorised, and income streams, rent, loan repayments and bounced payments are grouped. Flags mark large unexplained deposits, gaps in salary and cash withdrawals above a threshold. The underwriter sees one summary with the evidence behind each number and makes the decision.
The problem
Reviewing twelve months of statements by hand took hours per file, totals were added up in spreadsheets, and the unusual transactions that mattered were easy to miss.
Approach
Statements are read once into clean transactions, and totals are always recalculated from them so they tie back to the bank's own balances. Categories and flags follow the lender's rules. The summary never decides: each figure links to the transactions behind it, and the underwriter makes the call.
- Input
- AI step
- Rules or code
- Person decides
- Result
Screens
Swipe sideways to see the whole screen.
Client work is shown under confidentiality agreements: no client names, and screens are recreated with sample data.
Stack
- Document reading (Gemini API)
- Python
- Lender rule engine
- Underwriting dashboard
Impact
Typical for systems like this.
