Kyung-soo Kim

AI Agent for Advertising (PoC)

• Developed and optimized Python APIs for automating product and service promotional video creation for small businesses using LangGraph.
• Developed LLM-based APIs for video material recommendation and script generation through prompt engineering, and implemented a feedback-driven refinement loop to enhance script quality based on self-feedback.
• Enhanced function-calling model performance via tool information data augmentation, SFT and research into efficient reasoning with reward shaping and model merging.
• Improved internal content quality score from 3.1 to 4.4, and reduced reasoning path length by 31% while maintaining performance on BFCL and in-house evaluation sets.

Smart NPC Agent for Gaming

• Designed a persona-specific LLM data pipeline and optimized RLHF for character-driven NPCs that guide users and deliver immersive dialogues.
• Enhanced character persona fidelity through character boundary SFT, hierarchical PPO tuning, and reward modeling using text generation regularization to address reward hacking.
• Developed an automated LLM evaluation pipeline with G-eval.
• Mitigated reward hacking issues and increased G-eval persona score by 1.2 points (from 3.6 to 4.8), preparing the service with 16 distinct character personas.

RL-based Stock Trading Agent (PoC)

• Developed a distributed RL framework that aggregates experiences from multiple workers to optimize trading agent performance, incorporating model segmentation and pre-trained market encoder freezing strategies.
• Applied and validated the agent with real transactions at Korea Investment & Securities, achieving +16.43bp over TWAP and +15.45bp over VWAP in a one-month live trading pilot.

Transfer Learning Optimization

• Researched methods to efficiently adapt pre-trained deep learning models for defect detection in Hyundai Motor's smart factory domain.
• Evaluated advanced transfer learning with regularization techniques and studied dynamic gradient adjustment strategies based on transferability to the target domain.
• Applied the proposed approach to ImageNet-pretrained ResNet-50, achieving a 10.2% performance improvement over the baseline.

Optimization of Refinery Process Control Using Reinforcement Learning

• Optimized SK refinery operation by implementing reinforcement learning (RL) models for critical valve control, replacing conventional PID systems.
• Designed process-specific MDPs and applied a conservative offline RL training approach with uncertainty penalties to address real-world challenges such as action delays.
• Achieved a 32% performance improvement over PID controllers and successfully deployed RL-based autonomous control for one week in live refinery operations.