Fintech

Questions fintech teams bring

Can we trust today's portfolio-at-risk number?

Only if the pipeline proves it before the dashboard shows it: contracts on the inputs, reconciliation against the source, and lineage that shows where every figure came from.

How do we add a new lending partner without building a new pipeline?

Standardise the model once, then make each partner a configuration: which door it uses (CDC, extract or dump), its schedule and its field mapping.

Which features can the credit model use in production, and how fresh are they?

The ones defined in the feature store with an online path and a refresh schedule. Anything computed only in a notebook is not a production feature yet.

How do we join on personal data without exposing it to analytics?

Join on a keyed hash, keep the raw identifier encrypted in a restricted schema, and publish masked marts. Analysts get the joins, not the identities.

We have monthly batch reports. Do we need real-time?

Usually hourly change data capture with exactly-once loads is the right step. It removes the multi-hour batch window without the operating cost of a streaming cluster.