09 — Models with reasons
Models that price risk, read documents and make decisions — each with an explanation and an audit trail attached.
In regulated finance, a model output without a reason is unusable. Adverse action notices, credit decisions, monitoring alerts and pricing all have to be explicable to a customer and defensible to a supervisor, sometimes years later.
We build the pipeline that makes that possible: versioned models and features, decisions recorded with the inputs and model version that produced them, human review where the stakes require it, and monitoring for drift and disparate impact rather than accuracy alone.
Common questions
Can large language models be used for credit decisions?
For structuring and extracting information, yes, and that is where the value usually is. For the decision itself, the constraint is explanation: a decision must be reproducible and attributable to specific factors. In most jurisdictions that pushes teams toward interpretable models for the decision, with language models handling the unstructured inputs feeding it.
What does model governance require in practice?
Versioned models and features, a record of every decision with the exact inputs and version that produced it, documented validation before deployment, ongoing monitoring for drift and disparate impact, and a named owner. If a decision from eighteen months ago cannot be reproduced, the governance is nominal.
How do you monitor for bias in a financial model?
By measuring outcome disparities across protected groups continuously rather than at validation only, testing proxies that correlate with protected attributes, and monitoring the decision distribution for drift. This is an ongoing control, not a pre-launch checkbox, because population shift changes model behaviour without any code changing.
Next capability
Compliance automation
Bring us the hard part.
Forty-five minutes with the people who would actually run the build.