Millennial Works · 2025.08 ~ Present
Frontend Developer · Kiosk Client
I build the kiosk client that runs unattended AI experiences such as AI photos, tarot, personality tests, photo cards, and AI music at pop-up stores and exhibitions. It has run 65-70 events without major issues, from game companies, theme parks, and department store pop-ups to an event in Singapore. I migrated the legacy C# WPF client to Electron + Next.js, and the renewed version is deployed on ~150 kiosks, handling ~30K card payments and ~8K coupon uses in a year.
WebSiteUrl ->
I designed the kiosk client with Electron and Next.js, owning IPC, logging and data management, and device authentication. Each booth runs its own program, and per-store settings are toggled through JSON configuration.
Coupon codes, cards, Samsung Pay, and Apple Pay are integrated through the KICC payment module (payments, refunds, cardless refunds). Results print on photo, card, and POS receipt printers, and can be downloaded with an on-screen QR code. Downloadable files are deleted automatically after a set time.
next-intl supports Korean, English, Japanese, and Simplified and Traditional Chinese, and dayjs formats dates and times for the event's country.
Windows TTS reads the screen aloud, and Electron dual screens show an assistive display alongside the main one.
AI music content such as "make your debut song" depended on an external SaaS. Switching to the open-source ACE-STEP 1.5 Turbo brought the API cost per request to $0 and removed the dependency on an outside service. A 120-second song takes about 8-14 seconds in the queue.
In the AI photo content, users pick one of eight styles such as virtual character, anime, pixel, retro, figure, and cubism, and the photo is converted and printed. Style guide assets are managed in 5 languages.
Given questions like "Is there a fire?" or "Are the person's eyes closed?", the vision language model Clef-Flash judges the camera feed as yes or no, used for on-site safety monitoring. It runs a Q8_0 quantized model with llama.cpp CUDA on a single Windows PC (RTX 4070 Ti SUPER 16GB, 16GB RAM) without external APIs, at about 110-130ms per judgment.
The Idol model turns a single image into a 3D Gaussian Splat, used for event content that needed 3D output. On a single Windows PC (RTX 4070 Ti SUPER 16GB, 16GB RAM), one generation with 262,144 Gaussians and 50 steps takes about 50 seconds at 94% GPU usage.
I like pulling new models, running them on a local GPU, and measuring speed and resource usage to see if they can become on-site content. Models I've run so far:
FrontEnd Developer
KoJaem Portfolio
Copyright 2026. all rights reserved.