AI customer service agent with automated prompt optimization driven by rep feedback
The team needed an automated feedback loop that could deduplicate customer service representative feedback and use it to improve the agent's system prompts over time, so the agent gets better automatically. The client wanted features delivered fast, but the existing codebase was too fragmented for anyone to iterate quickly, so setting up Claude Code for the entire team was essential to match that pace.
Joined a cross-functional team as one of several engineers. My focus was on stabilizing the codebase and laying the groundwork for the automated prompt feedback loop.
Refactored both the frontend and backend from a fragmented state into a clean, modular structure built to be Claude Code friendly. This let me deliver client requirements faster using custom agents and skills, and made it just as easy for anyone else on the team to do the same.
Designed the end-to-end architecture for the feedback pipeline (prompt deduplication and system prompt optimization), then created detailed UI mockups and validated them with the client to lock in requirements. Once the foundation was solid, the project was handed off to another team to build on top of it.