Timothy Park

A Python ingestion worker pulls creator-captioned videos with yt-dlp, classifies them as immersion or lesson content with Gemini, and tokenizes the transcripts with kiwipiepy down to lemmas and POS tags.

The frontend tracks every lemma a user is exposed to across 10k+ tokens per video, recomputes familiarity in real time, and surfaces videos in the 80–90% Goldilocks zone.

A self-healing NLP layer flags slang and compound words the dictionary misses, and a small admin console retriggers ingestion when a channel drifts.

Each card is anchored to a unique dictionary entry and a definition embedding. New definitions are matched against existing sentences with pgvector cosine similarity, so good content gets reused across decks and cache hit rates hold above 30%.

On a miss, Gemini drafts a beginner-appropriate Korean sentence that uses the term in grammatical context, and Google Cloud TTS produces a 0.8× audio file for clarity.

Bulk TSV imports run through PGMQ in batches of 25 with live progress over Supabase Realtime; the same backend serves a web app and an Expo mobile app.

Zero database, zero network calls. Every Hangul string is generated algorithmically: powers of 10,000 for Sino-Korean, decade maps for Native Korean, irregular hours (한·두·세·네·스무) for time.

A logarithmic dual-thumb slider snaps to significant-figure jumps from 1 to 10 billion, and a guide overlay annotates each digit block with 만/억/천/백/십 brackets.

Time Trial scores by character length so a long counting phrase isn’t worth the same as a short one — and the top five scores live in localStorage with a screenshottable result card.