Turing College Logo

wae. — the subtitles YouTube can't make

wae. — the subtitles YouTube can't make

About this project

Subtitles translate the words. They don't translate what a Korean fan hears in them.

wae. explains the Korean nuance that non-Korean speakers miss — honorifics, speech levels, wordplay, editor captions, running gags, and cultural references that never make it into a subtitle. When a member says a mountain "feels like Gapyeong," Korean fans picture clean air and weekend trips outside Seoul; everyone else sees a place name. That gap is where the reaction happens, and wae. makes it visible at the exact second it lands.

It works as a Chrome extension. It reads the video's Korean captions in the viewer's own browser (nothing is downloaded), and a side panel synced to the player shows a short reaction tag at the moment a Korean fan would react, with an expandable "why?" card explaining the nuance in plain English.

Under the hood, a LangGraph pipeline splits the transcript into scenes and annotates them with a model (via LangChain and OpenRouter), grounded by hand-authored Korean culture and fandom knowledge retrieved through RAG. Every candidate passes a confidence gate before it reaches the screen, and machine-generated captions are flagged as uncertain rather than presented as fact. Quality is tracked with offline tests and a human-graded gold set.

Built with LangGraph, LangChain, ChromaDB, FastAPI, and a Chrome MV3 extension, deployed on Render. Development was AI-assisted with Claude Code, with the author — a Korean speaker and K-pop fan — as planner, reviewer, and author of all knowledge content.

Created byGain Choe
Published atSeptember 17, 2026
CourseAI engineering
Looking for
Potential usersCo-founderFreelance projectsNew job
OpenOpen LinkedIn