Halcyon Loans
An underwriting copilot that gets reviewed by humans, not bypassed.
- Client
- Halcyon Loans
- Sector
- Lending · Mumbai
- Engagement
- 4 weeks
- Type
- Client engagement
The problem
Halcyon's underwriters spent five hours per file stitching bureau pulls, bank statements, and policy rules into a decision. Throughput was capped by headcount and two senior reviewers disagreed on a third of edge cases.
What we built
- 01
A grounded underwriting agent
The agent reads the application bundle, pulls bureau and bank data through typed tools, applies Halcyon's written policy, and drafts a decision with an evidence trail for every claim.
- 02
An eval set that mirrors the floor
We labelled 600 historical files with senior-underwriter decisions and ran the agent against them on every change. Nothing shipped that regressed agreement with the senior bench.
- 03
A review surface, not a black box
Officers see the draft, the confidence band, and the evidence. They accept, edit, or override, and every override is captured as a new eval case, so the system gets sharper with use.
How we measured it
Decision time, agreement with the senior bench, and override rate, tracked per officer and per policy clause. We tuned against the agreement score, not raw speed.