← CV
POIZON app icon

AI Product Manager Intern · POIZON, Shanghai · Jun–Aug 2026

Two pieces of work on POIZON’s Shopping Agent: building Gift Agent, a sub-agent for context-aware gift decisions, from 0 to 1; and improving the quality of the Shopping Agent’s web search through evaluation and iteration.

1 · Gift Agent

Results

15Kdaily UV reached by Gift Agent
+4.60%Shopping Agent conversations per user after launch (significant in A/B testing)
+74.42%Shopping Agent products browsed per user after launch (significant in A/B testing)
12.4%Gift Agent performance improvement from eval-driven iteration

Demo

Gift Agent entry page in the Dewu app
Entry page
Gift Agent conversation: analysis steps, product recommendations and gift-giving tips
A full conversation · scroll to read

What I did

1. 0-1 product design

Defined the intent taxonomy, decision factors, dialogue strategies, and retrieval logic that let Gift Agent help people decide what to give in a given context.

2. Launch

Launched Gift Agent to 15K daily UV, lifting Shopping Agent’s conversations per user by 4.60% and products browsed per user by 74.42%.

3. Evaluation

Defined 7 eval metrics based on unique product value & output quality, and built 200-case golden sets to measure Gift Agent’s quality.

4. Eval-driven improvement

Diagnosed issues from online evaluation results, LLM execution traces, and 50+ pre-release user feedback, then refined prompting and multi-turn context management, improving Gift Agent performance by 12.4%.

2 · Web Search Evaluation & Iteration

Results

+2.1%entry-to-order conversion from node-level quality work
+1.96%ordering UV from the same quality work

What I did

Node-level quality improvement

Led quality improvement for the query decomposition and web search nodes in Shopping Agent’s LangGraph pipeline: defined 12 node-level metrics and real-case eval sets via SQL, diagnosed failure patterns from execution traces, and partnered with engineering to prioritize fixes.