Insurance quoting platform: a leaner data layer and tests before launch

The problem

The platform quotes and binds insurance. Pricing comes from a rating engine, policies are bound through the app, and applications arrive as PDFs. I joined while it was being readied for launch.

What I did

I replaced the ORM with plain SQL on a Postgres connection pool: more than 14 services, 16 routes and 39 models. Twelve nested queries became two database views, and 90 packages came out of the dependencies. Validation moved to a lighter library at the same time.

After a day of magic-link login failures I moved authentication to Better Auth, across 14 routes, in a day. Existing users kept their passwords through a SQL migration of the stored hashes.

Then I put tests where mistakes cost money: the rating engine, the PDF parser, binding and coverage. That came to 269 unit tests, and the end-to-end suite went to 31 passing with none failing.

Getting ready for production

I set up continuous deployment to staging and production on AWS, with the client’s own domain. While preparing production for user acceptance testing I found an email configuration mismatch that would have silently failed every magic-link login, and added a guard against seeding production with test data.

Other work included an email system with stored templates and delivery logs, a billing fix for tax that was being left off, and time-limited, audited admin impersonation that can’t make payments or bind policies.

What a client gets from this

Fewer moving parts, a data layer the team can read, and tests around the parts of the product where a mistake reaches a customer’s bill.