Home

About

Work

Gold loan agent

Problem statement

Research

Loan deep dive

Lovable prototype

Middleware OS

Website redesign

Re-imagining email

Playground

Download CV

Pilot

Automating Gold Loan Risk Reviews with AI

Designing Explainable AI Supervision for High-Risk Financial Workflows

Designing Explainable AI Supervision for High-Risk Financial Workflows

Rootflo reimagined gold loan audit in the AI era with image analysis and I led the end-to-end UX design of an AI-powered Gold Loan Evaluation Agent used by banks to detect fraud, reduce appraisal errors, and prioritize audit risk in gold loan operations.

80% of high-risk applications correctly flagged for review

80% of high-risk applications correctly flagged for review

Standardized valuation with 65% reduced variance for manual appraisal

Standardized valuation with 65% reduced variance for manual appraisal

Skills

UX, User research, UI, Prototyping, Vibe coding

Team
Parvathy KuroorU1
Vishnu SatisU2
Vishnu RKU3
Nitin ShajiU4
Parvathy KuroorU1
Vishnu SatisU2
Vishnu RKU3
Nitin ShajiU4
PROBLEM STATEMENT

Manual gold loan evaluation is slow & not efficient

Manual gold loan evaluation is slow & not efficient

Gold loans are high-volume financial products in India. Branch staff appraise jewellery manually using traditional methods like touchstone testing. The cost of a single misjudgment could cost the banks.

Operational Challenges

Operational Challenges

  • 5–7% appraisal error rates

  • Stone-weight miscalculations leading to excessive disbursement

  • Delayed detection of spurious gold

  • High attrition at branches

  • No structured audit prioritization

  • Weak audit trail defensibility

USERS
Branch Agents

  • Semi-trained

  • Under pressure to disburse quickly

  • Resistant to heavy supervision

Auditors

  • Review thousands of packets

  • No risk prioritization system

  • Reactive fraud detection

Compliance Team

  • Require traceable decisions

  • Need RBI LTV adherence

  • Require audit defensibility

DESIGNING CORE EXPERIENCES

Loan Deep Dive

Redesigning a risk-heavy audit workflow for speed, clarity & signal density

CONTEXT

The Loan Deep Dive experience is where auditors evaluate individual gold loan cases flagged by the AI system.

Each loan contains:

  • Multiple jewellery items

  • Stone-weight estimates

  • Spurious design signals

  • Fraud similarity matches

  • Suspicious activity alerts

Goal: Enable faster, more accurate risk decisions without overwhelming users.

Goal: Enable faster, more accurate risk decisions without overwhelming users.

Loan list dashboard & individual loan details page

The Loan Deep Dive starts from a tabular loan overview with date, Loan ID, Branch & Zone, Loan amount and status.

Users can:

  • Filter by date range

  • Apply filters and sort through loans (risk, clarity, suspicious activity etc)

  • View risk status badges

Inside each loan details page, we get to view the AI image analysis results with metrics like:

  • Clarity Score

  • Overlap Score

  • Detected Items list

  • Stone-weight detection

  • Risks Identified

  • Suspicious Activities

Problem #1
Auditors were missing critical risk signals

While user testing our pilot with a few stakeholders in a bank, we made a few observations:

  • Users rarely scrolled beyond the ornament list table.

  • Critical risk information below the fold was being missed.

Solution 01: We changed the page hierarchy and introduced the final verdict right in the beginning

Solution 01: We changed the page hierarchy and introduced the final verdict right in the beginning

Solution 02: Stone weight data and the alerts all shown in the same table so the user can view all necessary information in one place

Solution 02: Stone weight data and the alerts all shown in the same table so the user can view all necessary information in one place

Problem #2
Stone-weight deviation detection

  • Approximately 50% of inspector flags were due to stone-weight deviations

  • Branches sometimes under-deduct stone weight, increasing LTV beyond RBI’s 75% ratio.

Design challenge: Blocking every deviation would slow approvals, while ignoring them increases risk.

Design challenge: Blocking every deviation would slow approvals, while ignoring them increases risk.

Solution: We introduced side-by-side comparison of branch entered stone % and AI estimated stone %

  • Highlight deviations >10%

  • Allow override with mandatory justification

Solution: We introduced side-by-side comparison of branch entered stone % and AI estimated stone %

  • Highlight deviations >10%

  • Allow override with mandatory justification

Solution: We introduced side-by-side comparison of branch entered stone % and AI estimated stone %

  • Highlight deviations >10%

  • Allow override with mandatory justification

Problem #3
Auditors were struggling to go through all loans after filter application

  • Users had to go through the list and click on each loan to view the details, which took more time

  • After applying a filter, the user needed to view all loans quickly

Solution: Introduced a multi loan viewer mode in the dashboard with the option to switch between list and slider views. The multi-viewer mode is applied by default when filters are applied.

Solution: Introduced a multi loan viewer mode in the dashboard with the option to switch between list and slider views. The multi-viewer mode is applied by default when filters are applied.

Gold audit prototype

Rapidly Prototyping Internal Image Analysis for Bank Branches

CONTEXT

Built a quick functional prototype in Lovable to demonstrate how banks could run internal gold image analysis directly at the branch level within their own infrastructure.

Goal: To validate feasibility, reduce stakeholder skepticism around internal deployment, and translate AI capabilities into a realistic, branch-friendly workflow that could be experienced interactively rather than presented conceptually.

Goal: To validate feasibility, reduce stakeholder skepticism around internal deployment, and translate AI capabilities into a realistic, branch-friendly workflow that could be experienced interactively rather than presented conceptually.

1️⃣

Branch-Level Image Upload

1️⃣

Branch-Level Image Upload

2️⃣

Instant AI Evaluation View

2️⃣

Instant AI Evaluation View

3️⃣

Structured AI Feedback

3️⃣

Structured AI Feedback

Initital prototype to match old UI style with Rootflo's embeds

Let's work together

I'm eager to know more about your work, let's talk.

About

Work

© 2025 Parvathy Kuroor

Back to top

Let's work together

I'm eager to know more about your work, let's talk.

About

Work

© 2025 Parvathy Kuroor

Back to top

Let's work together

I'm eager to know more about your work, let's talk.

About

Work

© 2025 Parvathy Kuroor

Back to top