Junhwan Seok

Built and shipped a production ETL pipeline ingesting global market news every 15 minutes with embedding and event clustering. Reduced daily LLM spend by 76% through model routing, prompt caching, and batching. Replaced an LLM classifier with deterministic logic plus embedding fallback, improving accuracy from 19% to 46% at a quarter of the cost. Moved financial computations from LLM to rule-based code, eliminating fabricated numbers and cutting the core module from 1,458 to 365 lines.

Led auth, data modeling, and API backbone for an AI writing assistant with four engineers. Owned all Alembic migrations and contributed 63 backend commits, the most on the team. Designed two three-agent LangGraph pipelines for source extraction and draft generation.

Built a batch scheduler ingesting Seoul Open Data into MongoDB on a fixed cycle. Owned the meeting-domain API. Developed for the Seoul Open Data public-data competition.