The "Health Check" of AI Products: How I Build Evaluation Systems
Fine-tuning and prompting are only 20% of the work. The real challenge lies in setting standards. A guide to building a data-driven evaluation system for RAG and Agents.
Zhexi Zhu
Senior Computer Science student at the University of Michigan, minoring in Mathematics.
In Summer 2025, I worked at Tencent Advertising, leading the 0-to-1 build of 'Miaoxiaosi,' an intelligent Agent. It is Tencent Ads' first personalized vertical Agent with creative analysis and generation capabilities, providing merchants with a one-stop solution for short video ad creation. Additionally, I designed an automated 'Image-to-Video' workflow, effectively solving the pain point of video asset scarcity for merchants.
In Summer 2024, I served as an AI Product Manager at Shanghai Starbit, where I built an Agentic RAG intelligent assistant for IT operations. By designing multi-turn hierarchical retrieval strategies, I significantly optimized the accuracy of the system's responses.
I am currently exploring independent development, with interests in multimodal AIGC applications in advertising, AI Agent applications, AI product design, and Web full-stack development.
Personal Work
2025Full-Stack Developer
LecNotes.ai
An AI-powered platform that transforms lecture slides into comprehensive study notes. Built with Next.js and integrated with LLM APIs for intelligent content analysis and professional formatting.
2026Developer
Product Ad Video Generator
A Coze Skill that generates 30s+ professional product ad videos with one click. Features AI visual understanding, script generation, natural TTS, AI digital human avatar, and intelligent BGM matching.