Work

Selected problems I have worked on, and what I built to solve them.

📄 Prefer one page? Download my résumé (PDF, in Chinese)

Wealth management, 0→1 · ByteDance Fintech · 2025–now

The problem. Plenty of people are curious about investing but have nowhere obvious to start. The business is referral-based: we help users discover, follow and understand securities and funds, then connect them with licensed brokerages and fund distributors.

What I built.

• A discovery homepage: a card-based content supply with its own ranking pipeline, plus vertical content such as short videos with AI summaries and interest-based stock ideas.
• Subscriptions, watchlist and in-app messages to bring users back at the right moments.
• A traffic-slot system with an operating playbook, measured against holdout groups.
• A finance Q&A agent that answers investing questions where users already search.

Agent tools as a platform · Baidu Qianfan · 2025

The problem. An agent is only as useful as the tools it can call. Capabilities were scattered, uneven in quality, and hard for agent builders to find.

What I built.

• Lifecycle management for the tool layer, turning in-house and third-party capabilities into standardized MCP servers and Skills.
• Quality standards for tool calls, so agents pick and use tools more reliably.
• The Baidu Search Skill, packaging Baidu search as a Skill any agent can install. It became the most-downloaded official search-engine skill on ClawHub.

Search that understands supply · Meituan · 2023–2025

The problem. Good search depends on knowing what each merchant and product actually offers. That knowledge was spread across tag services, filter services and recall pipelines, and every change needed engineering time.

What I built.

• One tag data platform that unifies supply understanding, the knowledge graph and business data behind search.
• A tag operations system that lets business teams manage tags for recall, filtering and display in real time, without waiting on engineers.
• Search result strategies: scenario-specific filters, and recommendation reasons that combine LLM generation with statistical signals.
• AI helpers for my own team: a conversational agent for reviewing search quality, and an operations assistant with a knowledge base and MCP tool calls.

Earnings nowcasting · Oxford DataPlan · 2020–2023

The problem. Investors are regularly surprised when companies report earnings.

What I built. As part of the founding team, models that forecast company KPIs from alternative data, delivered as daily predictions through a data API.

Experimentation tooling · eBay China

What I built. An automated A/B test analysis tool, and causal evaluations of marketing, such as whether coupon campaigns actually drive incremental sales. The tool earned an eBay Spot Award.

Reading job ads with GPT-3 · London Business School × Jacobs · 2022

Before ChatGPT, I used GPT-3 to summarize and extract key information from thousands of job descriptions for an attrition early-warning project, and designed a scorecard to compare models on conciseness, completeness and cost.