Prior

The problem

Tenant screening is a slow, manual workflow with an uncomfortable property: two underwriters looking at the same application can reach different conclusions, and neither can fully articulate why. The inputs are scattered across structured application data and a pile of supporting documents, and the judgment that connects them lives in someone’s head.

Prior, built at LatentAxis, was an underwriting ML system for the landlord and tenant screening vertical. The goal was to turn that manual workflow into a consistent, explainable scoring pipeline, one whose outputs could support, not replace, human underwriting judgment.

How it works

Prior combines two kinds of evidence. Structured applicant data comes in through a FastAPI service backed by PostgreSQL, and a document processing pipeline extracts signals from the supporting paperwork. The combined features feed a scikit-learn model that produces a risk assessment for each application.

The output is never just a number. Every assessment carries the reasoning behind it, so an underwriter can see what drove the score and weigh it against their own judgment.

Decisions worth surfacing

Explainability as a requirement, not a feature. Decision transparency shaped the design from the start. A score an underwriter can’t interrogate is a score they either rubber-stamp or ignore, and both outcomes defeat the purpose. Prior’s outputs were built to be examined.

Assist the human, don’t replace them. Underwriting carries consequences for real applicants, and the system was deliberately scoped as decision support. The pipeline delivers consistency and speed; the accept-or-decline call stays with a person who can see exactly what the model saw.

Outcome

Prior shipped as a LatentAxis product: a working underwriting scoring pipeline with explainable outputs for the tenant screening vertical. It replaced a slow manual screening workflow with a consistent process whose every score could be traced back to the signals that produced it.