Henry Kang

Lead research and engineering for a 4-person character-chat startup, covering LLM selection and fine-tuning, conversational memory and persona, and real-time inference. Built short- and long-term memory benchmarks that now decide which model and memory changes ship. Raised KV and prompt cache hit rate from under 10% to 30% in the custom routing architecture, and secured about US$200K in signed contracts.

Lead applied research on on-device small language models, cutting mobile inference latency from 14s to 3s (a 79% reduction) through quantization, layer pruning and runtime tuning. This work led to a first-author EMNLP 2026 main-conference paper on RAG with frozen SLMs, plus 3 papers submitted as corresponding author. Teach LLMs and applied AI to international students and mentor student teams building AI products.

Publishing independent AI research paper on mobile edge AI under constrained compute environments. Secured Google IT project support; authored and published AI e-books for the Korean market. Build the AI based Game.

Designed the AI architecture for a real-time conversational avatar product, including the LLM reasoning loop, fine-tuning and evaluation. Guided 4 junior scientists through data curation, fine-tuning, eval design and error analysis. The team was accepted to Y Combinator W25 and raised a $700K seed.