UnderwriteAI

Keep the reasoning behind a credit draft visible.

An underwriting workbench that separates document interpretation, calculations, rule checks and narrative so a reviewer can inspect how a credit case was built.

WHO IT IS FOR

Credit professionals and small lending teams exploring structured preparation of credit memoranda.

MY ROLE

Product builder · credit-domain design · workflow and control design · testing

The problem

Credit appraisal draws on fragmented borrower information, calculations, policy and judgment. A fluent AI summary can conceal missing inputs or blur the difference between an assessed rule and a question that was never checked.

The approach

UnderwriteAI brings my banking experience into the structure of the product. Its central design principle separates the work: models extract and draft; deterministic code calculates and evaluates configured rules; the qualified reviewer makes the lending decision.

The workbench is designed to assemble borrower evidence, financial metrics, risk arguments and a reviewable memorandum. Its market-specific design covers US, UK and onshore UAE contexts, with jurisdiction and currency kept explicit.

The public experience is currently a synthetic interactive demo. A visitor can inspect fictional source figures, change assumptions, see ratios update and open an example memorandum. The broader live document workflow remains under validation; public live trials are not open.

HOW THE WORK FITS TOGETHER

01Inspect evidence
02Review extracted figures
03Calculate metrics
04Assess rule coverage
05Challenge and draft
06Human review

Inside the product

Visible calculations

Debt / EBITDA, debt-service coverage, current ratio and loan-to-value are calculated from stated inputs. Missing evidence should remain visible instead of becoming an invented value.

Four distinct rule states

Finding, no finding, not checked and not applicable communicate different answers. A quiet checker is never an overall compliance clearance.

A challenge to the case

The wider workflow includes a separate risk perspective to question the proposal before the narrative is assembled.

Evidence-conscious handling

The implementation is designed for browser-side document extraction and redaction before text transmission, with server-side provider secrets. Real-data handling still requires validation.

A CLEAR VIEW OF READINESS

What is ready.
What still needs work.

The public demo uses invented business cases. It does not analyse real borrower files or demonstrate a completed live trial.
The system is decision support, not a lender, regulatory certification or a replacement for institutional policy and qualified review.
Rule coverage is bounded and must be checked for applicability and currency. UAE scope is onshore, excluding DIFC and ADGM.
No production deployment, accuracy guarantee or customer outcome is implied by the sample.

What comes next

Complete live workflow validation, test representative edge cases and document the limits of each jurisdiction before expanding trial access.

WHAT THIS PROJECT SAYS ABOUT MY WORK

In credit, “not checked” can be the most valuable thing a system tells you. Confidence should follow the evidence.

Explore the synthetic demo ↗Back to portfolio

Status reviewed September 2026. Descriptions distinguish implemented work, public demonstrations and remaining validation.