Lenders live or die on how well they read a borrower’s bank data. We built a classifier that turns messy transaction feeds into a clean, reliable signal for credit-risk decisions.
The challenge
Raw bank-transaction data is messy and high-volume. Categorising millions of transactions by hand does not scale, and every mis-categorisation feeds straight through into a flawed credit decision. Lenders needed classification they could actually trust at production volumes.
What we built
A large random-forest model, served in production with sklearn and FastAPI, that categorises bank transactions at 0.97 precision. It turns raw statements into a dependable basis for affordability and credit-risk assessment, at the volumes real lending runs on.
Deployed to major banking clients, enabling accurate credit-risk assessment at scale.
The results
Lenders can assess affordability consistently and at scale, with categorisation accurate enough to rely on in live credit decisions rather than as a rough first pass.
Our role
We were the key technical ML lead on the build, working alongside one other ML engineer, from model design through to production deployment.