KuCoin internal platforms

Design lead · 3 designers · 5 platforms · design system

Product
In-house back-office platforms replacing third-party tools (Zendesk, Ada)
Target users
200+ support agents, team leads, compliance officers, data teams
Scope
Programme lead; deep dive: agent workspace redesign and AI assistance
Team
3 designers (I led) · 2024–25 · Hong Kong
  1. 5Platforms, 0→1
  2. 6 monthsProgramme
  3. 2 weeksRedesign
  4. 20Pilot agents
The redesigned agent workspace beside the four other platforms: customer-service console, chatbot configuration, data platform and compliance system, all on one design system
KuCoin internal platforms
02 · Overview

Leading five platforms, and stepping in where it mattered

KuCoin set out to replace Zendesk, Ada and other third-party tools with in-house platforms. I led a team of 3 designers across all five platforms and a shared design system. When the most critical one, the agent workspace, left agents stuck, I stepped in to redesign it.

03 · Leading the programme

Five platforms, one design team

Third-party tools were costly and couldn't keep up with crypto-specific work. With a small team serving five user groups, my job was to decide who owned which platform, how much each MVP covered, and which shared rules kept quality even.

Lower costReplace Zendesk and Ada
Built for cryptoWorkflows shaped around crypto support
ConnectedLinked to the core system's risk, KYC, trading and account modules
My role on each platform
1 / 7

Everything else I led and reviewed. Screens captured in Chinese, as built for the internal teams.

Who owned what, and why
Agent workspace

Designer A, the strongest on visuals. Agents chat with customers here, so it shares many consumer-facing patterns.

Compliance system

Designer B, quick to grasp complex product logic and professional context, and to turn compliance rules into sound layouts and workflows.

Tickets and design system

Designer C, who joined for the later months after I asked the director for help when we were overloaded. C was keen on design systems, and that work needed no domain ramp-up.

Data platform

Me, alongside reviewing the other four.

Scope: five platforms, one launch

I scoped every MVP so all five platforms could go live at about the same time. Anything a platform didn't need on day one moved to a later iteration, including the shared empty, loading and error states, which arrived in the second.

A disagreement with developers

Developers were overloaded and wanted simpler interactions; my test was whether the experience still held up for the people using it. We agreed to build the design system first, to settle patterns once and speed up development. The trade-off: UI details such as spacing and type were rough for a while, and caught up once components were synced in design and code.

04 · Design system

Rules, not just components

Built on Ant Design, the system set more than button styles. It set the rules every platform followed, so an agent, a team lead and a compliance officer met the same behaviour wherever they worked. I reviewed every platform against these rules.

1Claim and lockWhoever claims a ticket first is assigned it. One rule, so two agents don't answer the same customer.
2Empty, loading and errorShared from the second iteration, so every table and panel behaves the same when it's empty, waiting or failing.
3PermissionsRoles see different actions, such as edit or view only, and in compliance, different modules.
4One interaction modelOne sidebar, one visual language, and the components B2B tools rely on: buttons, tables and decision trees.
05 · Stepping in · Agent workspace

The busiest platform needed a rethink

When a customer opens KuCoin Support, three platforms work behind the scenes. The agent workspace is where 200+ agents handle chats and tickets all day.

KuCoin Support chat on the customer side, with bot replies and related topics

Customers

KuCoin Support

Where customers ask for help

Agent workspace: chat list, conversation with an AI suggested reply, and user info

Support agents

Agent workspaceCase study focus

Where agents handle chats and tickets every day

Customer-service console: trigger conditions and an assign-to-team action

Team leads

Customer-service console

Sets routing rules, SLAs and metrics

Chatbot configuration: scenario list and the AI reply scenario with conditions and a behaviour

Chatbot operators

Chatbot configuration

Sets auto-replies, knowledge base and AI logic

Programme brief

Agent workplace

Where support agents take chats and work tickets — the busiest of the five platforms in the programme.

Design 2024.10 – 12 MVP 2024.12.25
Replaced · chatbot
Ada knowledge configuration: article list for the AI agent
Ada — chatbot config
Replaced · agent workspace
Zendesk agent workspace: ticket conversation and side panels
Zendesk — agent workspace
Built
In-house agent workspace: chat list, conversation and user info
In-house agent workplace Linked to the core system's risk, KYC, trading and account modules.
Before

Third-party tools were costly and inflexible for crypto support work.

Goal

Raise agent efficiency and experience, improve the customer-facing support experience, and replace the third-party stack.

First MVP

Designer A owned the first UI; when agents couldn't use it before launch, I led the redesign with A and B in January 2025.

Why build in-house
  1. 1Built for crypto — workflows shaped for crypto support, not generic ticket templates.
  2. 2Efficiency and experience — agents work faster; customers get better support.
  3. 3Deep integration — risk, KYC, trading and accounts in one place, so agents don't hop systems.
Why I stepped in

The first MVP was built under time pressure, before any agent had tried it. The screens were polished, but the flow was fragmented: in review, agents kept getting stuck, key actions were hidden and the screens didn't follow their workflow. That wasn't a way of working to change overnight. Compliance had just wrapped up, so I brought in designer B, the strongest on flow logic, and with two weeks to the development deadline took over the redesign with A and B.

06 · Research

Watch agents first, then redesign

With only two weeks, I started with observation rather than screens: 5 agents worked on the live MVP, 2–3 per session. We pinned every point of confusion to the screen, then grouped them into three structural blockers.

The live MVP during an observation session
  1. Unclear titleWhich field does it come from? Needs a label.
  2. No new-chat alertA live chat arrived without a clear signal.
  3. Bot and agent look alikeBot and agent messages look alike; consider collapsing bot replies.
  4. Narrow bubblesMessage bubbles are too narrow to scan.
Notes from one observation session, pinned to the live MVP.
07 · Three fixes

Three blockers, three fixes

Three blockers explained almost all the friction. I scoped each to one focused change that fit the development deadline.

Where agents got stuck
  1. Sign in & go online

    Set status to online and join the queue

    No friction observed

  2. Pick up the ticket

    Take an incoming chat or an unassigned ticket

    Blocker 1

    Agents didn't know to take a ticket before replying

  3. Read the context

    Check who the user is, their KYC level, any holds, and past tickets

    Blocker 2

    User, ticket and history split across tabs

  4. Write the reply

    Translate, pick a macro, add an internal note

    Blocker 3

    Most-used tools were the hardest to find

  5. Submit, escalate or close

    Submit as open, raise an admin ticket, or close and follow up

    No friction observed

Each blocker sits at the step where agents hit it.
1
Hidden claim step
4 of 5 agents

BeforeClaim button: tiny, three screens in

Before: three screens to claim a ticket, ending at a small claim button in the side panel
Scroll to see all 3 screens
After: the reply box reads Take the ticket to start reply, with Take it

AfterTake it, right in the reply box

2
Context split across tabs

BeforeTwo tabs: user info, ticket info

Before: user info and ticket info as two tabs in the right panel
After: user info, ticket history and ticket info stacked in one right-hand column

AfterOne column: user, history, ticket

3
Everyday tools hidden in icons
None found them unaided

BeforeInternal notes: a tiny side-rail icon

Before: internal notes open from a tiny icon in the right-hand rail
After: Quick Reply, Add Note, Ask Help and survey as labelled buttons above the reply box

AfterLabelled tools, right above the reply box

Fix 2: three side-panel layouts we compared
  • Two zones, collapsible: user info and ticket info in separate areas.
  • Anchor navigation: one long column, with links at the top that jump to each part.
  • Collapsible sections (chosen): one column in the order agents check context: user, history, ticket, each able to expand or fold.

We chose collapsible sections: everything stays in one column, in the order agents read it, and they can fold away what they don't need.

Try the redesigned workspace

Try it here: start at the sign-in screen and follow the redesign as one flow, from going online to replying with an AI suggestion, raising an admin ticket and closing the chat.

Interactive prototype built from the Figma file. Use the steps along the top to switch pages, or tap an empty spot to highlight the active areas. Open full screen ↗

Agent workspace after the interface was aligned with the upgraded design system

Final workspace, after the interface was aligned with the upgraded design system.

08 · AI assistance

AI that supports the agent's judgment

I set one rule for both AI features: tie them to the ticket in front of the agent, and leave what reaches the customer to the agent. The design followed the Ant Design AI guidelines and a review of competing support tools.

AI reply suggestions

A draft reply, ready to edit

  • Detects the issue and suggests a reply to edit
  • Designed to cut typing and handling time
AI summary

The ticket at a glance

  • Sentiment, issue and a suggested next step
  • Pick up a case without rereading the thread
A suggested reply under the detected issue in the chat, with Paste to reply box and a close button

AI reply suggestionIn the chat, under the detected issue. The agent pastes it into the reply box to edit, or closes it.

Copilot panel: summary, emotion, consulting sensitivity, category, recommended reply and related issues

AI summarySummary, emotion, sensitivity flag, category and a recommended reply, in the side panel.

Where the agent stays in charge
1When it appearsAutomatically, on a customer message, whenever there's something worth suggesting.
2When it's wrongThe agent closes or edits it. Nothing goes out until the agent sends it.
3Sensitive ticketsThe summary flags sensitivity, and the agent handles the case under the team's guidelines.
4The agent sendsEvery reply is reviewed and sent by the agent. The AI never answers a customer on its own.

Still open: agents could close or edit a suggestion, but not regenerate or flag it. I'd add both, so a poor suggestion becomes feedback rather than a dead end.

09 · Result

Ready to pilot, not yet measured

Over six months, the team delivered five platforms on one shared design system, and the redesigned agent workspace was ready for a 20-agent pilot. Before the pilot ran, the team was restructured and I moved to a customer-facing project, so there is no usage data.

Programme
  • Five platforms built from scratch on one design system
  • Zendesk and Ada replaced by in-house tools
  • One interaction model across the three support tools
Agent workspace
  • In the new-design review, agents said the main pain points were resolved
  • Key actions now sit where agents work
  • Ready for a 20-agent pilot
How I'd measure it
1Time to first replyDoes taking the ticket in the reply box get agents answering sooner?
2Claim errorsHow often agents try to reply before claiming, the MVP's biggest blocker.
3Labelled tool useHow often agents use Quick Reply, Add Note and Ask Help, compared with the old icons.

The first pilot week, compared with the MVP, would set the baseline. The redesign stayed as the starting point for the workspace.

10 · Reflection

What I'd keep, and what I'd change, as a lead

1
Match designers to the problem, not only their strength

Matching by strength worked for compliance and the design system. On the agent workspace, a flow-heavy tool, strong visuals alone weren't enough. Next time I'd pair visual and flow skills on the busiest platform from the start.

2
Watch users before the MVP, not after

Skipping early validation left the MVP out of step with how agents work, so it needed rework and cost development time. Watching real users is the fastest path to clarity.

3
Still open: measure from day one

Without a baseline, the redesign's value rests on review feedback. Next time I'd set time to first reply and claim errors before the MVP ships, so any redesign can be judged against them.

Mini case study · Diting

Diting: self-serve data for analysts

Turning repeat data requests into templates, so analysts could pull data without writing SQL, and operations could later do it themselves.

Role
Product designer, hands-on on most modules
Team
1 other designer, PM, data engineers
Users
Data engineers and analysts; operations later
Year
2024 · KuCoin

1Context

Analysts and data engineers pull user data for operations, such as the list of users who qualify for a reward campaign. Every pull started as a request and ended as a spreadsheet. The first release served these two roles; later releases would let operations pull data themselves.

2Research: how a request moved

I had calls with users to walk through real requests, then mapped the flow for both data engineers and analysts. A reward request, for example:

  1. Jira request
  2. Confirm on Teams
  3. SQL in HUE
  4. Double-check
  5. Confirm with business
  6. Upload or email

What slowed it down

  • Repeat requests, such as reward lists, rewritten in SQL each time
  • Manual checks against reward amounts, or cross-checks with data engineering
  • Exports capped at 50,000 rows, so large pulls were split into parts
  • Results passed around by email

3Scope

The PM planned four modules, and I worked to the delivery order: data tools first, then the data service (APIs), with data assets and a data map opening data to operations later. Before designing, I also reviewed comparable data platforms, such as DataLeap.

4Design

Templates for repeat requests

A common request becomes a template. Running a task shows only the parameters it needs, then sends the result to the reward platform or by email automatically.

Three steps: create a template, create a task from it with only the needed parameters, then track task progress

Create a template, run a task from it, track progress. Red notes mark the auto-filled parameters and the automatic hand-off.

Self-serve query, no SQL

Pick a data model and columns, then combine and/or filters. Built for analysts first, and ready to open to operations later.

Self-serve query: data model fields on the left, selected columns and nested and/or filters on the right

Data map to find data

One search across tables, tasks, topics and APIs, grouped by business line. Admin pages cover accounts, roles and menus, with logs for auditors.

Data map: one search box, asset types and business lines, with top searched and top viewed tables

5What I learned

Walking through real requests mattered more than starting from a feature list. The template idea came from seeing the same reward request rewritten again and again.

Diting wasn't a company priority, and no usage data was collected.