What I can do for you

AI productcapabilities

I break down AI scenarios from a product manager's perspective, connecting capability design, prototype execution, validation, and review in one loop.

1. AI Product Definition

2. RAG Product Design

3. Agent Workflow

4. Vibe Coding

Yichen Huang

Product

Manager

AI Workflow / Product Loop

About me

I am a product manager transitioning from traditional internet products into AI product management. I am building hands-on experience through real projects around RAG, Agent workflows, Prompt design, AI product evaluation, and Vibe Coding. Rather than simply calling a model, I care about designing AI product loops that are testable, controllable, and iterative: starting from real scenarios, designing the right AI workflow, building runnable prototypes, and improving through feedback.

00

Featured AI Projects

00

Core Capability Areas

00+

Case Review Dimensions

I focus on turning ambiguous AI opportunities into real scenarios, product workflows, runnable prototypes, and case studies that can be reviewed.

RAGAgentPromptVibe Coding

Featured projects

AI product projects

These projects show how I understand RAG, Agents, Prompt design, AI product evaluation, and product loops through real practice.

RAG + Memory Workflow

Revive

A historical-saved-content reuse tool for knowledge workers, helping past collections become usable again when writing proposals, reviewing work, or preparing reports, while gradually adapting to task preferences.

Live

RAGTask Preference MemoryEvidence CitationStructured GenerationKnowledge Reuse

Problem

Many knowledge workers save articles, cases, methods, and experience-based content for years, but when writing proposals, doing retrospectives, or preparing reports, they still need to search, reread, filter, and summarize everything manually. The problem is not only that content is hard to find; saved materials rarely turn into directly usable work output at the moment a real task appears.

Target Users

Product managers, operators, marketers, consultants, and content workers who keep long-term collections and need to turn external information into proposals, reviews, reports, planning materials, or communication drafts.

Existing Alternatives

Users usually rely on platform bookmarks, read-it-later tools, knowledge bases, or manual research combined with general AI. These tools can store or generate content, but they require high upfront organization effort, provide unclear evidence for results, and rarely adapt to a user's own output habits over time.

Product Solution

Revive is not a general second brain or a heavy knowledge-management system. It focuses on reusing saved content when a task happens. Users import a batch of content, then start from a concrete task. The system generates structured output that can be copied, edited, and traced back to supporting evidence.

AI Workflow

The system is built on a RAG workflow: imported content is parsed, chunked, and indexed. When users submit a task, the system retrieves relevant passages from the current content set and generates structured results. Key conclusions are connected to citation evidence that shows original text snippets, not only source titles.

Memory System

On top of RAG, Revive adds task preference memory to remember how users apply knowledge to complete work. It captures preferred output structures, citation habits, writing style, and negative preferences, then uses them as soft constraints during generation so results gradually fit real work habits.

AI Trend Agent

REFINE

An AI trend-generation Agent for the AI industry, designed to filter high-value content from noisy information streams and turn it into structured outputs that can be reviewed over time.

MVP completed

Multi-AgentLLM-as-JudgeInformation FilteringEvaluation SystemPrompt Iteration

Problem

AI product managers need to continuously track model, product, technology, and business updates, but daily information sources mix promotional content, SEO articles, and repetitive summaries. The signal-to-noise ratio is low, and traditional reading often leads to information being consumed and forgotten instead of becoming reusable product judgment.

Target Users

AI product managers, operators, and knowledge workers who need to follow AI industry updates every day and turn external information into competitor observations, product judgment, and topic inputs.

Existing Alternatives

Newsletters, social platforms, X, tech media, and Perplexity can help users access information, but they still rely heavily on manual filtering and lack stable quality judgment, structured accumulation, and continuous tracking around personal focus areas.

Product Solution

Refine is not simply an AI tool that automatically writes trending topics. It is an information-filtering and trusted-accumulation system for AI PM workflows. The product aggregates technology media RSS content, then uses Agents to judge, distill, and integrate the information into a low-noise, reviewable AI industry daily brief.

AI Workflow

The workflow is designed as a four-step Agent chain: search, judge, distill, and write. Search handles content recall, judge handles quality scoring, distill creates structured summaries, and write integrates the daily brief. With modular decomposition, recall gaps, misjudgment, summary distortion, and writing quality issues can be diagnosed separately.

Project Highlights

By aggregating RSS sources, Refine improves recall quality. Around the judging Agent, it builds a small-sample cold-start evaluation system, defines a three-dimensional scoring rubric, and calibrates LLM-as-Judge with a human golden set, reaching a 90% human-AI agreement rate.

Frequently askedquestions

A quick overview of how I think about AI product work, what this portfolio demonstrates, and where interviewers can start.

Because AI products are not only about connecting models. They require redesigning information flows, decision flows, and human-AI collaboration. This direction needs product judgment, prototyping ability, and continuous validation.

It highlights RAG, Agent workflows, Prompt design, AI output evaluation, Tool Calling, and fast prototyping with Codex, Claude Code, and Cursor.

Each project explains the problem, target users, existing alternatives, AI workflow, MVP scope, validation method, failure boundaries, and key learnings.

Start with the project section, then review the capability section and About section. The projects show how I apply AI capabilities to real task scenarios.

Product thinking

A reserved area for future essays on AI product reviews, methods, and reflections.

2026.06

The value of RAG products is not just search, but task-based knowledge reuse

A reflection on how Revive turns retrieval into an actionable product loop.

2026.06

In Agent product design, boundary control matters more than automation

A product breakdown of Agent workflows through planning, tool calling, and failure recovery.

Let's work together

Contact

If you want to learn more about my AI product projects, prototype thinking, or case reviews, feel free to reach out.

Send email