AI is transforming the traditional internet industry. In the past, product validation cycles were long; today, product managers can use large-model capabilities to quickly vibe code interactive versions and validate product value in real scenarios. I want to move into AI because it not only improves the efficiency of validating product ideas, but also places higher demands on scenario judgment, technical understanding, evaluation awareness, and system design. These are the directions I want to develop in over the long term.

These two projects mainly demonstrate my ability to narrow scenarios, design AI product workflows, handle model boundaries, and think in terms of evaluation. Refine helped me understand the issues Agents face in information filtering, quality judgment, writing integration, and LLM-as-Judge evaluation. Revive led me to design a complete loop from content import to structured output around RAG, trustworthy citations, task reuse, and preference memory. Compared with simply writing PRDs, I care more about placing AI capabilities into real workflows and validating value through runnable products.

Calling a large model is only the underlying capability. The real product challenge is how to place the model inside a controllable, trustworthy, and verifiable workflow. Refine is not focused on "letting a model write daily reports," but on using a judging Agent and evaluation system to filter low-quality information while handling decision conflicts across multiple Agent modules. Revive is also not simply about "uploading materials for AI to summarize." It uses RAG citations to keep evidence trustworthy, then applies task preference memory so outputs continue to fit the user's habits while avoiding memory contaminating content evidence.