A trust score out of 100 for every loan document
Banks receive salary certificates, bank statements and IDs as PDFs and photos, and a small share are edited. We are building a tool with a banking domain partner that scores each document for signs of tampering and explains every flag in plain language.
Background
A banking professional with years in UAE lending came to MornningStar with a clear problem: reviewers cannot check every document deeply, and the tools on the market give a yes-or-no answer nobody can defend to a regulator. MornningStar is the technical builder; the partner brings the banking knowledge.
Challenge
A useful score has to be explainable. Documents are in Arabic and English, often mixed. And the best checks, like matching against genuine bank templates, need real documents that only a bank partner can provide.
What we are building
- A scoring engine that returns a trust score out of 100 with a list of reasons, not a single verdict
- Metadata checks: creation and edit history, software used, inconsistencies
- Font and layout checks that catch pasted-in numbers and names
- Arithmetic checks: do the totals, balances and salaries actually add up
- Language checks using Claude: wording that does not fit a genuine issuer
- Bilingual Arabic and English OCR, with a plan to fine-tune on real bank documents in the pilot
Tools used
Claude and Claude Vision APIs, Python document forensics, PaddleOCR / EasyOCR, Vercel.
What changed
In development with a banking domain partner, ahead of a bank pilot on real documents.
A score with reasons, not a black-box yes or no.
Metadata, fonts and layout, arithmetic, language.
Arabic and English, in the same document.
Other work
Ad tracking rebuilt so a beauty brand could sell across India
Two clean signals, one national market
Walk through your apartment before it is built, with the real view from your floor
Each unit type set up in under 7 minutes
See why a line slowed down before the shift report is written
Downtime explained by cause, in minutes
Tell us what slows your business down
Thirty minutes on a video call. We will tell you honestly whether AI can help, what it would take, and roughly what it would cost. No slides. No hard sell.