Statistical rigor, your credit policy, and an explanation you can audit.
Underwrite.ai hybrid models combine three purpose-built components: a gradient boosting machine for predictive power, a deterministic rule layer that encodes your policy, and a large language model that explains every outcome — without ever making the decision.
One application flows through three layers.
Each layer does one job well. The score comes from statistics, the decision comes from your policy, and the explanation comes from language — kept strictly separate so responsibility is never blurred.
Gradient Boosting Machine
Produces the risk score. Nonlinear, high-accuracy, and idempotent — identical inputs always yield an identical, reproducible result.
Credit Policy Rules
Makes the decision. A transparent, editable rule layer applies your cutoffs, overlays, and hard constraints on top of the score.
LLM Explainer
Explains the decision. Translates feature attributions and rule triggers into plain-language rationale and adverse-action reasons — read-only.
The strengths of each, none of the weaknesses.
Reproducible by design
The GBM is idempotent: rerun the same application and you get the same score, every time. No drift, no randomness in the number your decision rests on.
Your policy, encoded
The rule layer keeps credit policy where it belongs — with you. Adjust cutoffs and overlays without retraining, and know exactly why each rule fired.
Explanations, not black boxes
Every decision arrives with a plain-language rationale grounded in real feature attributions — ready for applicants, reviewers, and regulators alike.
The LLM explains. It never decides.
We deliberately keep the language model out of the credit decision. It receives the score and the rules that fired — then writes the explanation. It cannot change an outcome, override a rule, or introduce a factor the model didn't use.
That separation is what makes the explanation trustworthy: it describes a decision that was already made by deterministic, auditable components — not a story generated after the fact.
Auditable at every layer.
Because scoring, policy, and explanation are separate and deterministic, every decision can be reconstructed end to end — the exact score, the rules that applied, and the reasons given. Fully aligned with FCRA adverse-action requirements and fair-lending review.
