March 2026 · Finance · Working demo
An AI credit desk where officers still sign off
CreditRisk gathers the records on each borrower, runs the numbers and drafts an assessment. People make the decision, and the system keeps a record of who decided what.

- steps from borrower name to drafted assessment
- 14
- steps from borrower name to drafted assessment
- roles: officer, senior officer, administrator
- 3
- roles: officer, senior officer, administrator
- people for every grade change: one proposes, one approves
- 2
- people for every grade change: one proposes, one approves
The problem
A credit review pulls from borrower records, financial statements, news, legal research, spreadsheets and approval emails, and then has to be watched for months afterwards. The work is scattered, but the final decision still needs a clear trail of why.
An AI that simply produces a rating is no use here. It has to collect the sources, keep the facts apart from its own judgement, show what it is unsure of, and hand the decision to the officer responsible.
What we built
- 1
Collects the records
A 14-step pipeline finds the company, gathers financial, news, legal and registry records, computes the ratios and prepares a structured assessment.
- 2
Keeps facts and judgement apart
Calculations, source documents and the model's suggestions stay separate, so an officer can see which is which.
- 3
Routes decisions through people
Queues for officers and senior officers hold every submission, approval, rejection and return. A suggestion never overwrites the recorded grade.
- 4
Keeps watching afterwards
Warning signs, scheduled checks, reports, cost tracking and an activity log continue after the borrower is onboarded.
Where people decide
A grade change is proposed by one person and approved by another. The system will not let anyone approve their own change.
Destructive actions in the chat need a confirmation that expires if it isn't given.
Where it stands
A working demo that shows the whole path from a new borrower to a reviewed decision in one place: web app, Python API, PostgreSQL workflow state, analysis pipeline, review queues, monitoring and reports.
The same pattern fits compliance, underwriting and investigations: anywhere an agent should show its work before a person acts.
What it doesn't do yet
This is a demo with representative data. It is not running at a bank, it is not regulated advice, and it never makes a credit decision on its own.




See What an Agent Could Do in Your Business
Describe how the job is done today and who signs off. A short note is enough.