Interned on the Product Infrastructure org of TikTok’s Customer Service Platform (CSP), working along two throughlines: upgrading the AI customer-service bot from a rigid fixed-flow engine to an agentic architecture, and rebuilding how the platform evaluates service quality — and whether that evaluation is itself accurate.

Authored the SkillAgent design proposal: an LLM-driven agent that replaces the rigid JSON-graph execution engine behind the bot (a LangGraph multi-agent system). SOP flows defined as JSON graphs are recast as natural-language Skills (a readable Markdown replacement for the graph) that the LLM reads end-to-end, then plans and executes step by step through a ReAct loop — backed by a shared function/tool registry and memory-based multi-turn recovery.

Designed the agent harness: a version-controlled Skill repository with Libra-bucketed A/B experiments and gradual rollout by SOP (measured on CSAT and human-transfer rate), OpenTelemetry / Logfire observability of token cost, latency, and ReAct loops, plus safety guardrails (loop caps, retry, human fallback) — all requiring zero downstream changes.

Led the New Rule migration on the Manual Annotation platform: rebuilt the scoring standard from the legacy single-score, CSAT-weakly-correlated GE Rate into a two-part User Experience Score (UES) — a platform-view Service Quality Score (SQS) across 6 dimensions plus a user-view User Expectation Fulfillment score — adding gating cascades, Knowledge-Source-based branches (Skill / FAQ / SOP), auto-generated Human Result records for bot-to-human handoffs, and an A/B/C double-blind QC workflow. Migrated the dashboard’s core metrics from GE Rate to SQS/UES averages and QC accuracy, lifted annotation throughput ~120→250–300 cases/hour, and underpinned the team’s CSAT (50→60%) and eval-accuracy (77→83%) targets; built the frontend via spec-driven AI coding in days.

Led the end-to-end product lifecycle of AI Skills Hub (skillhub.tencent.com) — an AI capability platform for Tencent’s global studios. Independently owned product planning, front-end interaction design and content-ops strategy, aggregating 33,000+ AI skills plus daily AI×gaming industry insights, substantially accelerating AI adoption across production pipelines.

Contributed to the product definition of PUBG Mobile’s “Bruce” AI combat companion (600M+ downloads): combining LLM and reinforcement learning, drafted the PRD supporting real-time voice command + tactical cooperation, and positioned it in GTM as the first collaborative AI character in mobile shooter games that players can command by voice.

Designed and shipped a tag-based user segmentation system and content recommendation engine: used NLP to auto-tag user reading behavior and knowledge preferences, matching each user with similar peers and relevant articles — enabling group-level content distribution and personalized reach that measurably improved knowledge circulation and content discovery on the platform.

Contributed to a strategic due-diligence project in consumer retail, conducting executive compensation benchmarking for China’s leading F&B chains (Starbucks China / Yum China / Luckin) — systematically breaking down compensation structures and incentive mechanisms across management levels.

Synthesized fragmented data from annual reports, KJRC filings, and expert interviews to build compensation models — validating critical investment hypotheses and identifying opportunities in the consumer retail sector.

Before “Vibe Coding” became a buzzword (July 2025), pioneered a “PM × Dev in parallel” workflow: engineering set up isolated branches governed by a Spec Coding convention, letting PMs use an agent to prototype directly in the codebase and merge alongside feature logic. This shipped 10+ feature modules efficiently and compressed the PRD-to-experienceable-prototype cycle from weeks to days — later formally adopted team-wide after an internal share-out.

Designed a gamification system (virtual pet + gacha): defined the economy model, sink paths and milestone pacing. Post-launch it lifted silent-to-active conversion by 191% (from 358 to 2,400+ users), homepage DAU by +35%, and 2-month retention by +73%.

Independently pitched and shipped a player personality-test H5 live-op for Tencent Game Institute’s anniversary — owned the full chain from PRD, front-end implementation to analytics instrumentation, closing the PV→Share funnel. 20,000+ participants in 20 days; later adopted by Tencent’s Minor Protection team as a psychological-profiling tool for younger users.

Contributed to VALORANT Mobile’s launch live-ops strategy: used player-behavior analysis to shape the pre-registration activity, helping reservations surpass 70M. Distilled into a reusable launch playbook.

Authored several in-depth game-industry research reports covering breakout strategies for small studios, miHoYo’s overseas publishing model, and the WeChat mini-games ecosystem — one report was selected as an internal headline story and externally reposted by Tencent Game Institute’s official account, earning recognition from multiple director-level leaders.

At an internal Game Jam, served as Producer & Designer in a 3-person team and shipped the cooperative rhythm game OH MY BOSS — placed 3rd out of 20 teams, opening-day play covered 1,300+ employees.

Conducted 15+ stakeholder conversations to uncover shifts in consumer behavior, directly informing content planning and engagement strategies.

Created multimedia coverage of key events (e.g., local elections, tech expos, cultural festivals), ensuring alignment with organizational messaging goals. Converted field-level observations into strategic recommendations, boosting local relevance and brand resonance.

Focused on Chinese enterprises’ overseas localization — from Anta’s debut in Beverly Hills to community issues like the LA wildfires and social movements — bridging Chinese and English-speaking readerships.

Deep-dove into 20+ leading GaaS competitor ecosystems, systematically tracking 3,000+ version updates (balance tweaks, DLC releases, live-ops events) to distill a “frequent small updates + quarterly major releases” GaaS lifecycle model.

Used SQL to clean business data and built a Tableau dashboard to monitor DAU, retention and conversion in real time — delivering structured insight reports and regular readouts that earned recognition from multiple team leads.

Independently led end-to-end development of the himentor WeChat Mini-Program — from requirement definition, interaction design, and front-end implementation to back-end integration. Tech stack: WeChat Mini-Program framework (WXML / WXSS / JS) + Node.js backend + cloud database — shipped core pages, the user account system, and the mentor-mentee matching flow.

Designed and produced the full brand content system: defined visual language, wrote copy, and delivered asset kits to bootstrap the mini-program’s cold start.

Built and operated a Xiaohongshu product matrix — multi-account segmentation (product / mentor / mentee-story), established content cadence and hashtag strategy, and funneled Xiaohongshu traffic into the mini-program user base, forming a repeatable growth playbook.

Reported on the 2022 Global Digital Economy Conference — conducted on-site interviews and produced multiple stories for Xinhuanet’s digital-economy coverage.

Moderated and coordinated breakout sessions while contributing to content creation — working with editorial on theme design, guest outreach, agenda control, and copy review to keep on-site content aligned with the broadcast cadence.

Drafted tender proposals for the Tibet Autonomous Region Tourism Bureau project — from needs analysis and competitive benchmarking to budget structuring — independently delivering submission-ready business documents.

Served as video editor for Xinhuanet’s original program “Health Decoding” — handled full post-production (rough cut, fine cut, subtitles, audio-video sync, packaging) across multiple episodes, and produced program visual assets with Adobe Illustrator alongside the art team.