← Work
0022024

AI that explains a 42-page mortgage, credibly

An AI-powered assistant that reads Spanish mortgage contracts and explains every clause in plain language. The hard problem wasn't the model — it was designing explanations people in a low-trust situation would actually rely on.

Role
Lead Design Engineer · UX, prompt design, and accessibility — with an ML engineer and two full-stack devs
Status
Client work · production · UI abstracted
AI/UXGPT-4OCR pipelineReact NativeWCAG 2.1fintech
Client UI — shown with permission

The problem

Spanish mortgage contracts average around 42 pages of dense legalese, and in our research almost nobody read them before signing. People didn't discover what they'd agreed to until a clause activated years later — which is how mortgages became one of the biggest sources of banking complaints in Spain. Customers wanted to understand; they were just afraid that asking questions at the signing table made them look slow.

Decision

The trust paradox

We needed AI to build trust with customers who didn't trust banks — and who didn't trust AI either. The research pointed at the answer: people said they'd believe an AI explanation if they could compare it against the original text, and if a human expert had verified it. So the design never replaces the contract: original clause and plain-language explanation sit side by side, every explanation cites its clause number, unusual terms get flagged for human review, and each contract carries a visible legal-verification stamp.

How it's built

How it works

I co-designed the AI architecture with our ML engineer: a two-stage prompt system — extraction first (contract parsed into structured JSON: amounts, dates, clauses, parties), then explanation (plain-language summaries with mandatory citations, "you" language, and a hard rule against advice beyond the contract). I owned the prompt design end to end, plus the OCR error-handling flow for low-quality scanned PDFs and the WCAG 2.1 AA accessibility pass Spanish banking regulation requires.

Honest outcome

In testing, trust in the AI explanations roughly doubled against the control design, and comprehension moved clearly in the right direction. The client's numbers aren't mine to publish, so I won't dress this up with precise figures — what I can defend is the design system of trust: side-by-side text, citations, human verification. Every one of those decisions is documented here.