YEOYOUNG NA

Designed agentic workflows and core production infrastructure for B tv’s AI platform, supporting 4 agents and 1 orchestrator via FastAPI backend services and a vector database layer. Built RAG-based Product and customer support agents using hybrid retrieval (semantic vector search) with Reciprocal Rank Fusion (RRF), reaching 97% end-to-end answer accuracy. Optimized RAG serving APIs to keep average retrieval latency under 200 ms while improving horizontal scalability. Implemented Langfuse-based LLMOps observability for multi-agent workflows with prompt-level tracing and response-quality review. Added PII detection and masking prior to storage, and built batch pipelines for metadata extraction, embedding generation, and vector index construction with automated refresh and reindexing.

Operated and re-engineered the B tv search metadata platform by simplifying pipelines and optimizing queries to reduce batch processing time while improving data reliability and maintainability. Provisioned and governed an enterprise Azure OpenAI environment for internal service deployment. Built a GPT-based content block naming generator integrated into a recommendation admin system. Created keyword-based content similarity serving and integrated LLM-driven contextual retrieval to improve recommendation/search relevance. Operated and tuned a large-scale Hadoop data platform supporting recommendation and analytics workloads.

Contributed to migrating an on-premises data lake to AWS, owning EMR architecture design and Spark optimization for large-scale batch processing (recognized with features at AWS Industry Week 2022 and Snowflake Data Cloud World Tour Seoul). Implemented fine-grained access control with AWS Lake Formation (LF-TBAC) to strengthen governance and security. Designed a multi-cloud RAG architecture across AWS SageMaker, Azure AI Search, and GCP Vertex AI for an internal chatbot, including preprocessing and vectorization pipelines. Built a real-time viewer engagement pipeline using Spark Structured Streaming processing 150 GB/hour, with Kibana dashboards for audience behavior and content performance. Delivered an internal ChatGPT-style web app for text/image generation using Azure OpenAI and Django, plus a Streamlit dashboard for API cost tracking. Supported Snowflake data mart development and built Boto3-based EMR/EC2 monitoring tooling visualized in Streamlit.