Yiran Zhu Product Design @ Goldman Sachs
Work About Resume Email
1 of 10 exceptions
Exception only
AI extracted with potential issues
View reasoning
Manual input
Olivia Kiu Mar 3, 2025 14:22 EST
View reasoning
Citation 1
DocumentA.pdf
Page 29

The Client hereby acknowledges that the reallocation of allocated capital remains contingent upon the successful verification of all secondary identifiers. In the event of a discrepancy exceeding the established 0.05% threshold, the automated validation protocol will trigger a mandatory manual review.

AI Data Review Transformation

Designed an AI-driven data review that transformed a 7-hour manual process into a 30-minute exception-based audit, reducing vendor dependency and build user trust through explainable AI.

AI-HITL AI UX Patterns

Problem Business cannot scale

Every loan required two rounds of manual effort from 2 team members 7 hours total reviewing hundreds of data points against legal documents, plus 24 hours wait time on vendor.

It can never be done on time, and business can never scale. Limited capability of current legacy system and human error pose risk to data accuracy.

~500

Data fields to review

Per process

7+

Hours review time

Across two + rounds

24

Hours wait time

On vendor

Solution AI-driven exception based review

As the sole designer, I redesigned the experience to move from fully manual review to AI-driven exception-based auditing.

I built a concept prototype in a different context to demonstrate the problem I tried to solve without business information, using AI-native development process. It's still in progress.

View Vibe Coding Demo
  • Efficiency gain

    Mapping data to their original document coordinates addressed user's biggest time spent during review

  • Trust in AI

    Explainable AI and confidence-based filter allows users to bypass verified data and focus on uncertain answers

  • Overhead reduction

    Redefined operational workflow from 2-person double check to single-user oversight

  • Increased accuracy

    Limit manual touchpoint reduced human error, and robust real time data validation prevented unverified data getting into system

  • Cost saving

    In-house AI solution cut the millions of vendor cost

Impact

PHASED DELIVERY - MEASURED

98%

Decrease on wait time

PHASED DELIVERY - MEASURED

40%

Decrease on review time

FULL VISION - PROJECTED

10K+

Hours saved annually

FULL VISION - PROJECTED

92%

Decrease on review time

FULL VISION - PROJECTED

$600K+

Cost saving annually

FIRST

Across platform AI pattern

For AI human interaction on AI extraction output