Back Home

Lifted first-attempt verification pass rate by 8%

Lifted first-attempt verification pass rate by 8%

Lifted first-attempt verification pass rate by 8%

INDUSTRY:

TRADING

YEAR:

2026

EXPERIENCE:

PRODUCT DESIGN

Back Home

01. about

01. about

Finex is an Indonesian retail trading broker, regulated by Bappebti. In 2025, Finex ranked #1 in Indonesia by trading volume. I work as a product designer on the verification team, embedded in one of the product's most critical flows.

Finex is an Indonesian retail trading broker, regulated by Bappebti. In 2025, Finex ranked #1 in Indonesia by trading volume. I work as a product designer on the verification team, embedded in one of the product's most critical flows.

[ 01 ]

[ 00 ]

Over 800,000

registered users

[ 02 ]

[ 00 ]

$17M

average monthly trading volume

[ 03 ]

[ 00 ]

28,000+ verifications per month

a ~15-minute process, manually reviewed by licensed specialists

02. The challenge

02. The challenge

Audit the verification form and propose design solutions that move the metrics that matter most: first-attempt verification rate, completion rate, drop-off rate, successful re-verification after rejection, and median processing time.

Audit the verification form and propose design solutions that move the metrics that matter most: first-attempt verification rate, completion rate, drop-off rate, successful re-verification after rejection, and median processing time.

☹️

Only 59% of verifications succeed

on the first attempt

☹️

44% of rejections
are photo-related

44% of rejections are photo-related

users submit blurry or skewed photos and never retake them

☹️

30% never return after
a rejection

30% never return after a rejection

they don't come back to fix errors – unaware of review timelines or that they can appeal

03. Research

03. Research

The verification team had never run any research – so that's where I started. I set up moderated usability tests of the verification flow and in-depth interviews, deliberately sampling across the entire funnel: users who passed, users who were rejected, users still mid-flow, and users who registered but never began. The usability tests showed where the form broke; the interviews explained why.

The verification team had never run any research – so that's where I started. I set up moderated usability tests of the verification flow and in-depth interviews, deliberately sampling across the entire funnel: users who passed, users who were rejected, users still mid-flow, and users who registered but never began. The usability tests showed where the form broke; the interviews explained why.

[ 01 ]

[ 00 ]

20 respondents across 4 user segments

[02]

[ 00 ]

12 usability tests

[03]

[ 00 ]

8 in-depth interviews

[04]

[ 00 ]

15+ hypotheses to improve the form

04. Insights

04. Insights

Both the tests and interviews pointed to the same handful of problems again and again.

Both the tests and interviews pointed to the same handful of problems again and again.

[ 01 ]

[ 00 ]

More questions built more trust, not less. Long forms and heavy personal-data requests actually raised users' confidence – and Bappebti's oversight reinforced it.

[02]

[ 00 ]

Poor document photos are the single biggest failure point. Most users don't try for a clean shot, and unreadable ID numbers drive 44% of all rejections.

[03]

[ 00 ]

After submitting, users are left in the dark. They don't know how long review takes, when they can start trading, or whether they can still fix anything.

[04]

[ 00 ]

Users believe more is better. Many assume that completing more fields improves their odds of approval.

05. Hypotheses

05. Hypotheses

I turned each insight into a testable hypothesis – and tied every one to a specific metric from the challenge.

I turned each insight into a testable hypothesis – and tied every one to a specific metric from the challenge.

The primary ones:

The primary ones:

[ 01 ]

[ 00 ]

If we guide users through photographing their document, photo-related rejections will drop sharply.

Metric: photo rejection rate

If we guide users through photographing their document, photo-related rejections will drop sharply.

Metric: photo rejection rate

[02]

[ 00 ]

If we explain data security up front, fewer users will drop off on the steps that ask for sensitive data.

Metric: drop-off rate

If we explain data security up front, fewer users will drop off on the steps that ask for sensitive data.

Metric: drop-off rate

[03]

[ 00 ]

If the result screen clearly lays out next steps, review time, and how to re-verify, more rejected users will come back.

Metric: return after rejection

If the result screen clearly lays out next steps, review time, and how to re-verify, more rejected users will come back.

Metric: return after rejection

[04]

[ 00 ]

If the form shows where you are, what's next, and how long it takes, more users will finish it.

Metric: completion rate

If the form shows where you are, what's next, and how long it takes, more users will finish it.

Metric: completion rate

06. Prioritization

06. Prioritization

I plotted every hypothesis on a value/effort matrix to decide what to ship first – and started with the photo problem, since it had the biggest impact on the metrics.

I plotted every hypothesis on a value/effort matrix to decide what to ship first – and started with the photo problem, since it had the biggest impact on the metrics.

07. Fixing the photo problem

07. Fixing the photo
problem

Photo quality drove the most rejections, so by the numbers it was the obvious place to start. Here's the flow as it was – and where it broke down.

Photo quality drove the most rejections, so by the numbers it was the obvious place to start. Here's the flow as it was – and where it broke down.

08. The approach

08. The approach

Two goals shaped the solution: help users capture a usable photo of the document, and catch a bad one before it ever reaches review.

Two goals shaped the solution: help users capture a usable photo of the document, and catch a bad one before it ever reaches review.

👀

Document mask anchored to the ID photo

If a user correctly captures at least the face-photo area on the document, they'll most likely capture the rest of the document correctly too.

👀

AI quality check

An ML model scores each shot on ~10 signals
like blur, glare, crop, and readability, then blocks anything
unfit for verification.

An ML model scores each shot on ~10 signals like blur, glare, crop, and readability, then blocks anything unfit for verification.

👀

Preview before submit

Capture and preview currently look identical. A dedicated preview screen with quality feedback stops bad photos from going through.

09. Design

09. Design

In the first iteration, we added a mask, a detailed preview screen, and a model to detect photo quality issues.

In the first iteration, we added a mask,
a detailed preview screen, and a model
to detect photo quality issues.

10. Design v.2

11. Design v.2

In the second iteration, we replaced manual photo capture with automatic capture. The photo is taken automatically when our model determines that the document, or the person holding it, is properly positioned and can be clearly read.

Main

Projects

Contacts

LinkedIn

Resume

Main

Projects

Contacts

LinkedIn

Resume

[ Results ]

Key Results & Impact

First-attempt verification
success rate

59%

→

73%

First-attempt verification
success rate

59%

→

73%

Photo-quality rejection rate

44%

→

19%

Photo-quality rejection rate

44%

→

19%