Context
Pricing ran on a production workbook that had grown over years of changes. It worked, but every new rule needed a specialist, nothing could be called from other systems, and testing a change meant re-checking the whole file.
Approach
I started by separating what the workbook decides from how it calculates. Deterministic rating rules moved into a versioned rules engine. Machine learning was kept to one job: scoring the exceptions that underwriters should look at.
What was built
- A pricing API that channels and partners can call directly
- A rules engine with versioning, testing and an audit trail
- A rule-management console designed with underwriters
- An ML exception score that routes unusual quotes to people
- A deployment that can run on-premises or in any cloud
Outcome
Rule changes move from specialist edits to a reviewed, testable workflow. The same platform is designed to extend to other product lines without a rebuild.
What I'd do differently
I'd bring underwriters into the console design even earlier. The first screens were technically right but used our language, not theirs.