CHUN WANG05Contact
AI PRODUCT MANAGERPORTFOLIO / 2026

Between human needs and machine capability

I make AI feel
genuinely useful.

NAME
Chun WANG
PREVIOUSLY
NetEase · Baidu
STUDYING
University of Washington
Explore selected work Start a conversation
OPEN TO WORKAI PRODUCT MANAGER · SHANGHAI / REMOTE
A graphite feather whose lower barbs are printed in pale cyan-blue
PRODUCT SENSE●AI LITERACY●USER EMPATHY●EXPERIMENTATION●0 → 1
01ABOUT / A PRODUCT POINT OF VIEW
02SELECTED WORK / CASE STUDIES

Make uncertainty testable.

Two agent products I helped build during internships at NetEase and Baidu. The diagrams show what was broken, what I added, and what changed.

01 / INTERNSHIP · NETEASEOFFICIAL SITE 2026

AI Learning Companion

The best AI assistance knows when humans matter most.

An AI learning assistant for K12 students that answers when human planners are offline and hands complex cases back to them. I mined 5.365M chat records to set priorities, then designed the agent architecture, knowledge base, tool calls and evaluation, from zero to gray launch.

  • AGENT ARCHITECTURE
  • RAG & EVALUATION
  • 0 → 1
80%Core needs covered by the top-3 scenarios the AI was built for
Agent architecture, from incoming message to reply or human handoffMESSAGE INREPLY OUTCHANNELSAPP · WECOMMEMORYSLIDING SUMMARIESINTENT ROUTERDEEPSEEK V4 FLASH · JSON + CLARIFYTOOLS + KNOWLEDGE17 READ-ONLY MCP TOOLS · 3-LAYER RAGGENERATIONQWEN 3.7 PLUS · GROUNDED ANSWERHANDOFF4-RISK CHECK → HUMAN PLANNERMYDESIGN
FIG. 01-AAgent architectureEvery message passes six layers. The router and the tool layer are where most of the design work went.
Before and after tool calling: an order-status questionBEFORELLM WITHOUT TOOLSSTUDENT ASKSORDER STATUS · 22:40LLM ONLYNO ORDER ACCESSGUESS, OR WAITFOR A PLANNERREPLY AT 09:00NEXT MORNINGAFTERINTENT → MCP TOOL → ANSWERSTUDENT ASKSORDER STATUS · 22:40ROUTEINTENT: LOGISTICSMCP TOOL CALLget_order_logistics()GROUNDED REPLYSHIPPED OCT 24 · 3 SECTIMEOUT → RETRY → DEGRADE TO HUMAN
FIG. 01-BFixing a dead end with a toolOrder questions used to stall until morning. One read-only logistics tool turned them into a three-second answer.
02 / INTERNSHIP · BAIDUOFFICIAL SITE 2025

Baidu Merchant Agent

A better AI conversation should lead to a better outcome.

I turned sales practices into structured dialogue workflows for Baidu's merchant agent: chain-of-thought guidance, emotion-aware replies, and a three-step lead-qualification flow that judges lead quality inside the conversation instead of after it.

  • CONVERSATION DESIGN
  • LEAD QUALIFICATION
  • KNOWLEDGE BASE
+20%Qualified-lead conversion vs. the previous approach
Lead qualification as a network radiating from the chat: the old route judged leads two layers out, the new route judges them in the first layerIN THE CHATAFTER THE CHATAFTER THE CALLASK FOR PHONELEAD POOLSALES CALLGRADED LATER01 CLARIFY02 FILTER03 TAG INTENTNOT A FITHOT LEAD→ STRAIGHT TO SALESCHATBEFORE · JUDGED AFTER THE CALLAFTER · JUDGED INSIDE THE CHAT+20% QUALIFIED LEADS
FIG. 02-AJudge the lead earlierGray is the old route: collect a phone number, call, then judge. Blue is the route I designed: judge inside the chat, so only hot leads reach sales.
How one customer message becomes a reply and a lead cardONE MESSAGE, FOUR DECISIONSCUSTOMERPRICE OBJECTIONEMOTIONPRICE-ANXIOUSSOP STAGEHANDLE OBJECTIONCOT DECISIONREASSURE, THEN ASKREPLYCALM, REASSURING TONELEAD CARDAIGC · PERSONALIZEDTONE FOLLOWS THECUSTOMER'S MOOD+3.2PP LEAVE-INFO RATE+10% SATISFACTION+5% WITH LEAD CARDS
FIG. 02-BOne message, four decisionsEmotion, sales stage and a chain-of-thought plan shape each reply, and decide when a personalized lead card appears.
Merchant materials flow into one knowledge table; unanswered questions loop back to the merchantMERCHANT MATERIALSSINGLE PAGEWHOLE SITEPRODUCT CATALOGUPLOADED FILES· THREE PARSING MODESKNOWLEDGE TABLEQA PAIRSKNOWLEDGE CHUNKSSOURCE + UPDATED ATINDUSTRY TAGSMERCHANT AGENTANSWERS CUSTOMERSUNANSWEREDGAP DETECTORFLAG MISSING ANSWERSASK THE MERCHANT TO FILL THE GAP+20PP INDUSTRY KNOWLEDGE COVERAGE
FIG. 02-CA knowledge base that asks for what it lacksMerchant materials land in one table. Questions the agent cannot answer loop back as requests for the merchant to fill.
03EXPERIENCE / A SHORT JOURNEY

From industrial design to AI products.

  1. NetEase

    AI Product Manager Intern

    Built an AI learning companion for K12 students, from zero.

    Jul – Oct 2026INTERNSHIP

  2. Baidu

    AI Product Manager Intern

    Merchant AI agent for business conversations.

    Jun – Oct 2025INTERNSHIP

  3. Baidu UX Bootcamp

    UX Design Trainee

    Baidu Netdisk redesign; outstanding participant.

    Apr – Jun 2023PROGRAM

  4. University of Washington

    Master of Communication in Digital Media

    Graduate study in digital media, Seattle.

    Sep 2025 – Mar 2027GRADUATE

  5. Yangzhou University

    Bachelor of Industrial Design

    Where I learned to design around people, objects and behavior.

    Sep 2020 – Jun 2024UNDERGRADUATE
RECOGNITION
  • Gold Award, UK International Art and Design Competition, 2024
  • Silver Award, Huaxia Award for Cultural Arts Design Competition, 2024

WHAT I BRING

Foresight, perspective, and the patience to iterate.

(A)
A line drawing of a light cone, with past and future opening out from a single point

FORESIGHT

The present sets the horizon

Each choice today decides which futures stay within reach.

(B)
A line drawing of a Möbius strip, a band with only one side

PERSPECTIVE

Both sides are one

Follow the user far enough, and you reach the business.

(C)
A line drawing of a spiral staircase winding upward around a column

ITERATION

Circle, and climb

Each loop comes back to the same question, a step higher.

04HOW I WORK / AI IN THE LOOP

OBSERVE → FRAME → MAKE → LEARN

AI in every step
of the loop.

Not a subject I study, but the way I work. Models do the reading, drafting and grading at scale; I keep the judgment.

  1. A complete archival moon specimen showing a cyan-blue waxing crescent
    01
    OBSERVE

    Listen

    Let a model sort the chats; I read the outliers.

  2. A complete archival moon specimen showing a cyan-blue first quarter
    02
    FRAME

    Define

    Argue the problem with AI until one bet is left.

  3. A complete archival moon specimen showing a cyan-blue waxing gibbous phase
    03
    MAKE

    Build

    Vibe-code the demo instead of writing a long spec doc.

  4. A complete archival moon specimen showing a cyan-blue full moon
    04
    LEARN

    Measure

    Let a model grade at scale; I read every failure myself.

TOOLBOX /PRDFIGMASQLLLM EVALRAGAGENT WORKFLOWVIBE CODINGCLAUDE CODECODEXUSER RESEARCH

05 / CONTACT · SAY HELLO

If the signal feels right,
let's build something useful.

I’m looking for an AI Product Manager role—and always open to conversations about products, models, and the people between them.