Provenance, lineage, and explainability as infrastructure, because in regulated finance you must prove not just what the model decided, but what data it stood on and why.
Every other AI course treats data as a given input. This treats the data substrate, lineage, provenance, consent, and the explainability that flows from them, as the load-bearing infrastructure it actually is in regulated lending.
We architect end-to-end lineage, generate adverse-action reasons that map to model outputs, and run the disparate-impact and proxy analysis that 2026 enforcement is built on. Free, because it is the on-ramp to everything else.
Fair-lending and model-risk reviews begin at the data layer, because an examiner cannot test a model whose inputs are not yet proven accurate. This lesson shows where that review actually starts and what it asks of you first.
Data lineage in finance has to be a running system that emits machine-readable records as pipelines execute, not a diagram someone redraws before an exam. This lesson covers what to instrument, what to store, and how to reconstruct any model decision from source to output.
Most finance teams audit whether a model is fair before they audit whether the data was ever permitted for the use at hand. Purpose limitation is a data property, not a model property, and you have to enforce it at the column level before training starts.
A facially neutral feature can encode a protected class without anyone intending it. This lesson shows how to find proxies in your feature set and test for disparate impact, even as the legal ground shifts under ECOA.
Less-discriminatory-alternative search is a documented hunt through equally accurate models for one that reduces disparity. Here is how to run it as a repeatable process, not a one-off defense.
ECOA and Reg B require specific, accurate principal reasons for a credit denial, and a complex model is no excuse. This lesson covers how to generate reason codes from real model attributions and prove they match the decision.
Retrieval-augmented generation only earns trust in finance when every answer carries a verifiable record of which documents it pulled and which claims those documents actually support. This lesson covers the provenance plumbing that makes a RAG answer auditable rather than merely plausible.
A feature store earns its place in a regulated stack when it makes scores reproducible. Point-in-time joins, versioned features, and one definition for training and serving turn reproducibility from a hope into an audit control.
Most of the data behind an AI lending decision sits inside a vendor's systems, and the right to test it for fair-lending compliance only exists if you negotiated it into the contract first.