B2B SaaS | Internal Tool | Platform Redesign

B2B SaaS | Internal Tool | Platform Redesign

A/B Test Platform @ Kwai

A/B Test Platform @ Kwai

Role

Product Designer Intern

Worked with a senior designer mentor, 1 PM, 1 FE and 2 BE engineers

Timeline

Q3 2024 July - August

Tasks

Hi-fi Prototyping

User Research

Design Iterations

Dev & PM Collaboration

Tools

Figma

Enterprise Tools

/ OVERVIEW
/ OVERVIEW

Kwai’s internal A/B testing platform serves as a “decision lab,” powering thousands of experiments daily across UI, content, social, commerce, and advertising.

I led the redesign of the metric configuration flow and experiment conclusion module, streamlining the end-to-end experiment process and redefining how teams set goals, interpret results, and scale learnings. The new experience accelerated decision-making and set a foundation for scalable experimentation workflows. Launched in Q3 2024.

Kwai’s internal A/B testing platform serves as a “decision lab,” powering thousands of experiments daily across UI, content, social, commerce, and advertising.

I led the redesign of the metric configuration flow and experiment conclusion module, streamlining the end-to-end experiment process and redefining how teams set goals, interpret results, and scale learnings. The new experience accelerated decision-making and set a foundation for scalable experimentation workflows. Launched in Q3 2024.

This case study was conducted in Chinese; UI screenshots are annotated in English for review.

This case study was conducted in Chinese; UI screenshots are annotated in English for review.

/ IMPACT
/ IMPACT

We succeeded in reducing 75% of metric configuration time, from 25 minutes to 6 minutes per experiment on average.

We succeeded in reducing 75% of metric configuration time, from 25 minutes to 6 minutes per experiment on average.

The redesign has benefited across 100+ departments within the company, effectively reducing the cognitive loads of various roles in the experiment.

The redesign has benefited across 100+ departments within the company, effectively reducing the cognitive loads of various roles in the experiment.

$ 30k+

$ 30k+

monthly cost savings through improved metric configuration flow and guidance

monthly cost savings through improved metric configuration flow and guidance

8.5 -> 3.9

8.5 -> 3.9

click reduction per test due to streamlined user flow, cutting down repeated work of metric tracking

click reduction per test due to streamlined user flow, cutting down repeated work of metric tracking

3.2 -> 4.5

3.2 -> 4.5

user satisfaction score after 20+ hours of user interviews and several design iterations

user satisfaction score after 20+ hours of user interviews and several design iterations

HIGHLIGHTS
HIGHLIGHTS

A streamlined experimentation platform enabling diverse teams to run, analyze, and act on A/B tests with clarity and speed

A streamlined experimentation platform enabling diverse teams to run, analyze, and act on A/B tests with clarity and speed

After

After

Before

Before

/ CONTEXT

Untangling the Complexity

Understand Kwai's A/B Test Platform

An internal A/B testing platform used across Kwai’s product, content, and operations teams to evaluate and optimize features through controlled experiments.


It supports thousands of daily experiments, helping teams make data-informed decisions by comparing different versions of user experiences, content strategies, and recommendation algorithms.

Two distinct user flows across different experimentation roles

I was given only three days to ramp up, during which I quickly uncovered a hidden complexity in the experimentation workflow. What initially appeared to be a single, unified flow actually diverged into two distinct paths, each shaped by differing user roles and goals. As the product designer, I needed to create a solution that would be comprehensive yet intuitive for all user types.


Like in most companies, a typical A/B test involves close collaboration between experiment owners and data analysts, and each role has different focal points.

I was given only three days to ramp up, during which I quickly uncovered a hidden complexity in the experimentation workflow. What initially appeared to be a single, unified flow actually diverged into two distinct paths, each shaped by differing user roles and goals. As the product designer, I needed to create a solution that would be comprehensive yet intuitive for all user types.


Like in most companies, a typical A/B test involves close collaboration between experiment owners and data analysts, and each role has different focal points.

Experiment owners, typically product managers, engineers, or operations teams, are responsible for configuring all aspects of the experiment, including defining goals, selecting metrics, allocating traffic, and specifying variant logic. Their primary focus is on ensuring that tests are properly set up and that live data is aligned with the intended experimentation objectives.

Data analysts, on the other hand, are responsible for validating the experimental design, monitoring data quality, and interpreting test results. They provide statistical insights that help determine whether the outcomes are reliable, significant, and actionable.

/ PROBLEM SPACE

Uncovering Pain Points

What We Learned from Last Quarter

To better understand user friction within the experimentation lifecycle, product and design team launched a platform-wide survey last quarter, and we collected qualitative feedbacks that help us gain in-depth reviews. The goal was to identify which phases of the A/B testing process caused the most confusion, inefficiency, or frustration.

The responses pointed to main pain points around two experimentation stages: Experiment Metric Configuration & Experiment Data Analysis Phases.

Experiment Metric Configuration

"The metrics involved in experiment analysis are quite complex."

456

average subscribed metrics per test

20%

tests have 1800+ subscribed metrics

44%

survey participants unable to distinguish between goal vs. observation metrics

Experiment Data Analysis

"I had to click into each metric one by one, and I still couldn’t tell how far along the experiment was. It felt so frustrating."

20%

Missing SRM Checks, risk of invalid traffic allocation

40%

survey participants are unsure if metrics are statistically significant

70%

survey participants take 20+ minutes to analyze the data every time they enter the result interface

Where Users Struggle the Most

After conducting preliminary research through 20+ stakeholder interviews, platform usage logs, and survey data, I uncovered a series of interconnected pain points from the surface that impacted both experiment owners and data analysts. These issues didn’t exist in isolation: one problem often triggered another, compounding into workflow breakdowns that delayed decisions, compromised data quality, and reduced overall trust in the experimentation system.


To better understand where and why these breakdowns occurred, I mapped critical moments across the experiment lifecycle and divided them by two experimentation phases.

Metric Configuration (Setup) Phase

where owners struggled with fragmented responsibilities and high cognitive load. However, we have found out that the pain points are intertwined into a vicious cycle, which means we need to exterminate the root cause to fix the problem.

Data Analysis Phase

where analysts (mostly) and owners alike faced challenges interpreting overwhelming, unstructured data.

PROBLEM STATEMENT

How Might We help users subscribe to the right metrics with goal clarity and cost awareness, so they avoid over-complication and focus on meaningful results at the Experiment Metric Configuration Phase?

How Might We provide users with clear experiment progress and success signals, so they can confidently interpret results and take action faster at the Experiment Data Analysis Phase?

/ GOAL

A New Direction for A/B Test

Product Goals

Pain Points to Opportunities

Transforming user pain points into design opportunities based on existing platform functionality, and also explore possibilities within constraints after discussing with the product and dev team.

Changing the Vicious Cycle

Restructuring Site Map

After understanding the usage patterns of different user roles, I realized that our original product structure was too flat and linear.


All experiment-related features were piled together without clear prioritization, and due to the system’s inherent complexity, critical metrics weren’t properly distinguished. While the interface appeared comprehensive, it actually confused users across roles.


To address this, I restructured the information architecture of the redesigned phases centering it around the key task stages users go through during the experiment lifecycle. This led to a redesign of both the experiment creation and result analysis processes, building on the original foundation with a clearer, task-oriented logic.

After understanding the usage patterns of different user roles, I realized that our original product structure was too flat and linear.


All experiment-related features were piled together without clear prioritization, and due to the system’s inherent complexity, critical metrics weren’t properly distinguished. While the interface appeared comprehensive, it actually confused users across roles.


To address this, I restructured the information architecture of the redesigned phases centering it around the key task stages users go through during the experiment lifecycle. This led to a redesign of both the experiment creation and result analysis processes, building on the original foundation with a clearer, task-oriented logic.

Metric Configuration - Metric Dashboard

Goal Metrics

- primary indicators allow the users to measure the effectiveness and success of an experiment

Observation Metrics

- secondary indicators allow the users to monitor potential negative side effects during an experiment.

SOLUTION - Experiment Design Phase Design

Metric Configuration - Primary Setup

Hover State Checkboxes

Reached 50 maximum subscriptions, Reminders of subscription cost

Search for Metrics

Search for Metric Category

Metric Categories

- subscribe metrics based on client nodes, modes, and subscription history

Tooltip Hover

50 maximum subscriptions

2

Add / Edit Metrics

3

Edit Metric Template

2

1

Goal Metric Re-bucketing Toggle

Tooltip Hover

The system recommends setting no more than five re-randomization metrics.

Active State Checkboxes

Cost Reminder

It is estimated that the department budget will be ¥ xx per day.

3

Edit Metric Template

FINAL SOLUTION PREVIEW - Result Analysis Phase Design

Result Interpretation

Experiment Prompt Guidance Design

/ ITERATION

Metric Dashboard

My design iteration preserved the visual style and core interaction patterns of the original version to ensure continuity for existing users. While keeping the primary user flows intact, I focused on enhancing the overall user experience by streamlining the path to view goal metric details and estimated costs on the metric dashboard.


These changes made key information more immediately accessible and improved the clarity of data presentation.

Showing a thumbnail of the metrics subscribed

no intuitive indication of goal vs. observation metrics, and re-bucketing (goal) metrics
New subscriptions were designed in a new section, while taking up plenty of spaces

Before : A/B Test Platform 1.0 Design

In the first two iterations, I combined the metric thumbnails with subscription actions to create a more seamless user experience. To enhance clarity and focus, I also highlighted several key goal metrics for easier viewing and comparison.

Initial Explorations

Iterations based on user feedbacks


  1. 💡 I want to understand cost impact earlier, before committing to configuration.

  2. 💡 I need to choose which goal metric to re-bucket, not all of them at once.

Toggles on each of the metrics

Showing cost reminder and budgets while selecting, streamlined the user experience

Final Design Revised Based on User Feedbacks

REFLECTION

In this internship project, I learned:

Depth, Not Just Speed

What I Would Do Differently

Design Thinking in Motion

Depth, Not Just Speed

What I Would Do Differently

Design Thinking in Motion

My Design Reflections at work

Let create work that matters
Your vision + my design = magic

Available For Work

+1 (585)-754 5515

jtang6730@gmail.com

Made with love, passion & tons of caffeine :)

All rights reserved,

Jiaxuan Tang 唐嘉璇 ©2025

Let create work that matters
Your vision + my design = magic

Available For Work

+1 (585)-754 5515

jtang6730@gmail.com

Made with love, passion

& tons of caffeine :)

All rights reserved,

Jiaxuan Tang 唐嘉璇 ©2025