The work here starts from classroom constraints: curriculum language, teacher time, privacy, and the difficulty of preserving context across AI interactions.
An AI lesson planning assistant for Korean K-12 teachers. Built on RAG over 4,000+ national curriculum achievement standards, with DOCX/PPTX generation and MCP integration for both Claude and ChatGPT. Built by a pre-service teacher using AI development tools - no formal engineering background.
A small fine-tuned model acting as a persistent memory layer - generating domain-structured semantic cues that prime any frontier model. The compass persists. The engine is replaceable. Submitted to Red Bull Basement 2026 Korea. Inference server runs on-demand. Static demonstration available at the link.
A personal project mapping airline and hotel loyalty programs into a structured, queryable knowledge base - built out of curiosity about how complex rule systems can be made navigable through AI.