Yeonguk Lee — Software engineer
building real products.

Every scan triggered an AI call for ingredient analysis. At realistic scale, this breaks the cost model — and makes users wait on every single repeat scan of the same product.

Designed a 3-layer caching strategy: TanStack Query client-side, product-level cache in PostgreSQL, and Claude prompt caching server-side. Repeat scans are free. New products cost 60% less.

LLMs that generate code directly are fragile and break on every revision. A "make this button blue" request shouldn't require regenerating the whole app — but most AI codegen tools do exactly that.

Designed agents to manipulate a structured AppState JSON instead of raw code, with specialized roles (Clarify · Design · Structure · Merge · Generate). Code is the output, not the medium — making revisions stable, diff-able, and inspectable.

A fleet operator dashboard needs live position updates, historical playback, and full CRUD on vehicles — and the frontend had to be developed in parallel with a backend that didn't exist yet.

Built a mock-API-first frontend with MSW so development was fully decoupled from backend readiness. Zustand for global state, TanStack Query for server cache, Recharts + Kakao Map SDK for visualization. CI/CD with Docker + Jenkins.

Most "share location" apps either show stale dots updated every minute, or rely on managed real-time services that hide what's actually happening on the wire.

Building a custom WebSocket server with explicit message protocol, in-memory room state for sub-second broadcast, and async batch persistence to PostGIS. Designed to learn the systems, not abstract them away.

Build a full ML pipeline — train a model, serve inference, wire up a frontend, and deploy publicly — as a student team without ML infrastructure budget.

Trained EfficientNet-B0 in PyTorch, exported to ONNX for fast CPU inference. Wrapped with FastAPI, paired with a Flutter client and Supabase + Cloudflare R2 for storage. Led the team end-to-end.

Group trip planning today means jumping between maps, messengers, and recommendation sites — context gets lost and friends fall out of sync.

Built one app that integrates Gemini for AI route generation, Naver Maps for native location search, and WebSocket for real-time multi-user sync — wrapped in GetX MVVM with a resilient async API layer.

Generic caffeine trackers ignore that everyone metabolizes caffeine differently — by weight, age, sex, smoking status, medication.

Modeled caffeine half-life decay personalized to body profile, then projected the curve forward to estimate the optimal sleep window. Built end-to-end in 36-hour hackathon sprint.