Underwrite.ai

The company builds and maintains a custom model from each client's own loan history and returns decisions through an interface, with no workflow, origination system or data asset around it. It names gradient boosting and argues why it beats logistic regression over fifteen to twenty variables, a model company rather than a software company that added models.

Decisions return in two to three milliseconds, and the flagship case celebrates a lender adopting the model as its sole underwriting method. Staff are said to focus on exceptions, but nothing defines an exception, who sets the threshold or whether the model flags cases it should not decide.

A Korean market study reports area under the curve of 0.958 for the boosted model against 0.906 for a well tuned logistic regression, the measure a credit risk function uses, disclosed for both. No sample size, observation period or out of time validation accompanies it, and a separate claim of above 98 percent accuracy sits loosely beside those figures.

An installment lender's early default rate is reported falling from 32.8 percent to 8.5 percent, and a United Kingdom auto lender moved from days to milliseconds. Every case is anonymous, with no customer, executive quote, customer count or decision volume published.

Shapley additive explanations are named, with a worked example decomposing a prediction, and every decision is said to carry documentation of its driving factors. The company also publishes when its product is the wrong buy, for lenders without performance history or whose scoring already separates borrowers well.

Clients supply anonymized loan and application data for model building, so the training corpus need not identify borrowers. Live decision data, retention and subprocessors go uncovered beyond a public privacy policy, and the Gramm Leach Bliley Act is absent despite community bank and credit union buyers.

No certification, attestation report or security contact is published. The vendor sits in both the training pipeline and the live credit path, and community banks examined on third party risk will need controls documented before onboarding.

Published material explains adverse action notice duties, disparate impact and how a neutral variable becomes a proxy, and a use case targets regulatory compliance audits. That is literacy rather than standing, with no license claimed, no statute cited by name and no position on emerging AI rules in its markets.

Disparate impact analysis tools compare outcomes across demographic groups before deployment, and models are designed to avoid proxy variables, with postal code tracking historical housing segregation as the worked example. The capability is handed to lenders, and no results for the vendor's own models or independent fairness audit are published.

Public terms of service are reachable without contact, more than many larger vendors offer. Because the vendor builds and maintains each model, its quality is the vendor's work product, yet no warranty, liability cap or remedy for an underperforming or biased model is published.

Each model is built from the client's own anonymized records and calibrated to that portfolio, so a buyer knows the model learned from its book, not a pooled corpus. Inputs are typed as bureau, bank transaction and rent payment data, but no supplier, library or infrastructure provider is named, nor what happens to a client's model when the relationship ends.

A two to three millisecond interface call suits embedding in origination, and the company candidly lists the bureau, verification and origination connections a deployment needs. No interface documentation, sandbox or named supported platform lets a lender confirm coverage of its own stack.

Hosting, regions and residency are unpublished for a service whose work spans several countries' lending rules. Client loan records move to the vendor for training, and where they sit during and after development, and whether they are destroyed on exit, is undescribed.

$2 per decision is the published unit and rate, self disclosed on a vendor supplied directory profile, with a 30 day self serve trial and public terms. How custom model development and ongoing maintenance are treated under a per transaction price is not described.

Use case pages target small business lenders, community banks and auto lenders among others, a coherent focus on firms too small for in house modeling. Published work spans markets from Korea to rural Mexico, yet no institution is named in any segment or market.