Kyuyeon
Park
Builds.

Shipped a production CS agent for Korea's largest food delivery platform and replaced 9-hour manual QA with a 5-minute automated evaluation pipeline.

Diagnosed tool routing failures in a live system, consolidated proxy architecture from 2 to 1 tool call, and iteratively refined prompts via batch testing — pushing CSAT to 3.4/5 at 5% regional rollout (200–250 conversations/day). Built an LLM-as-a-judge evaluator scaling to 1,000+ conversations. Shipped a Slack/Jira/GitHub issue tracker adopted team-wide with zero onboarding.

Found the real bug — RDB-to-VectorDB misrouting — that the team assumed was a prompt problem.

Built a ReAct-based multi-agent LLM system for enterprise sales data consolidation. Traced end-to-end data flow across agent planning, tool routing, and DB layers. Shipped a domain-specific typo correction module at 95%+ accuracy, resolving 9 business-critical retrieval failures.

Rebuilt a monolithic crawl pipeline into a distributed system and cut cloud storage costs by 70%.

Re-architected a 100M+ page Korean web crawling pipeline into a distributed K8s/Redis system with memory buffering and compression. Designed a 2-step vector search tool planning system, filtering a 120k QA dataset to 20k high-quality samples (83% reduction) to replace LLM fine-tuning.