I'm Young-Kyu Yoo. I spent 17 years at Intuit working on QuickBooks, ending as a Principal Software Engineer, on software used by millions of small businesses. I stepped away in 2014 for a family break. In 2026 I came back to full-time development and built four full-stack AI prototypes — the architecture, the data models, and the LLM interaction design, end to end.
All four were built with Claude Code. Three were for product owners I partnered with; DEFINeD is my own. Each has real authentication, a real database, and a live deployment — I don't consider a prototype done until it runs in the world.
Seven games — from Spelling-Bee-style honeycombs to syllable puzzles and typing games — built on reusable puzzle engines. Every one of ~49,000 lexicon words carries an AI-generated, human-reviewed dictionary clue, produced by a multi-model pipeline: Anthropic Batch API generation with GPT and Gemini cross-review.
Supabase backend with 21 versioned migrations covering auth, daily puzzles, leaderboards, head-to-head challenges, friends, and AI-powered custom word lists. Includes a SHARK2 gesture-typing swipe keyboard decoder and a daily puzzle system with server-authoritative dates, data versioning, and client cache invalidation.
Students reason through multi-turn Socratic dialogue with an LLM that probes and scaffolds rather than lectures, adapting to grade level. Fully bilingual (English / Korean) with voice input and output. Each session ends with an LLM-generated learning report — vocabulary, concepts, and logic — plus curated book recommendations drawn live from two Korean national library APIs.
Built as a pnpm monorepo with an adapter boundary that keeps shared packages region-neutral for future markets. Four LLM providers run behind a single interface with retry. Currently in alpha with a school in Korea.
One intake — topic, grade, standard, duration — cascades into a lesson plan, worksheet, exit ticket, slide outline, and parent email that all share the same lesson context. Every output exports as a real PDF (server-rendered with full Hangul and math font coverage) or a real PowerPoint deck with speaker notes.
Supabase-backed lesson library with passwordless sign-in, row-level security, and public share links; AI picture-card generation for early-grade worksheets. The Korean edition encodes the 2022 revised national curriculum, standard lesson structure, and formal register switching per audience.
Roughly 36,000 lines of TypeScript across a React 19 client and a separate content-authoring admin CMS. Kids progress through an eight-level apron journey, earn coins and stars, and collect themed activity cards — and reaching Master level plants a real tree through One Tree Planted.
Offline-first architecture: bundled content seeds IndexedDB on first launch, then delta-syncs from Supabase. Native iOS shell via Capacitor with camera-backed photo journaling, 16 versioned SQL migrations with row-level security, and five Deno Edge Functions including AI image generation.
Coming back after eleven years meant learning a new way to build. This is the method I've settled into.
Each repo carries a working library of design docs — backend architecture, data models, migration plans — kept current as the system evolves, so the AI works from a spec, not from vibes.
DEFINeD's 49,000 dictionary clues were generated with Anthropic's Batch API, then cross-checked by GPT and Gemini, then reviewed by a human — me. Different models are good at different things; I use them accordingly.
Versioned migrations, row-level security, typed schemas, automated tests, and a second AI reviewing the first one's code before merge. The tools write faster than I ever did; the discipline is what keeps them honest.
I'm open to full-time and contract roles where experienced product engineering and AI-assisted development are both useful. Email is the fastest way to reach me.