Skip to content

AI underwriting engine for a lending platform

An LLM-assisted underwriting engine that lifted application throughput sixfold while lowering default rates.

Business impact

  • 6x

    6x application throughput

  • 31%

    31% lower default rate

  • 4-week

    4-week payback

Industry
FinTech
Timeline
14 weeks
Team
6 engineers, 1 designer, 1 architect

Business challenge

Manual credit review capped approvals at 300 applications a day and risk decisions were inconsistent across analysts.

Solution

An LLM-assisted underwriting engine with document extraction, a feature store, and human-in-the-loop review dashboards.

From the engagement

Workshops, working software, measured results

Team sketching a system on a whiteboard
Strategy before code
Shipping in two-week increments
Developer focused on a screen in a dim office
Deep work
Approach

How the work unfolded

Scroll to follow the programme from first workshop to measured result.

  1. 01

    Discovery

    Shadowed analysts for two weeks to map decision criteria and identify the documents driving most of the review time.

  2. 02

    Data foundation

    Built a feature store consolidating bureau data, bank statements and application history with lineage and access controls.

  3. 03

    AI extraction & scoring

    Deployed document extraction and an explainable scoring model, with LLM-generated summaries for analyst review.

  4. 04

    Human-in-the-loop rollout

    Launched to one analyst pod first, measured agreement rates, then expanded once accuracy thresholds were met.

Technology stack

Next.js
NestJS
PostgreSQL
LangChain
AWS

Have a similar challenge?

We will walk you through how this programme was shaped and what a version for your organisation would look like.