Project · 01
Connie Quest
Family use · reward system being redesigned- Q01
Why I built it - Between 'a child knows the right values' and 'a child lives them' there's a long river. After reading 'How to Become a Person', I needed a tool that could carry the book's developmental model into my daughter's everyday actions. She became the very first user.
- Q02
In one line - A family-scale PWA that turns a child's character growth into a Read → Know → Act loop — gamified to reinforce action, played as a game to deliver meaning.
- Q03
What makes it different - Most kids' apps revolve around ads and in-app purchases. Connie is built so that only the reinforcement of good behaviour is gamified. Family data is fully isolated with Supabase RLS.
- Q04
Today - My daughter is still the only user. Her appetite for rewards grew faster than we expected, so I'm rebuilding the reward system. Other work has paused active development for now.
- Q05
What's next - Once the reward system is back, we resume — and explore a collaboration with the book's author to deepen content fidelity (still searching for a contact).
Project · 02
Marketing SH.AI
Friends-only beta · auto-publish not shipped- Q01
Why I built it - After AX, the natural next question is GEO (Generative Engine Optimization). But most owners aren't familiar with GEO, and most can't afford an agency. They still need a starting line for marketing.
- Q02
In one line - A one-click GEO content-automation platform for small business owners — from idea discovery through publishing, condensed into a single click.
- Q03
What makes it different - Vibe Coding has dropped the 'invest' side of ROI so far that the era of instant apps has arrived. Services we couldn't justify before become viable. Marketing is no exception: we compressed an agency-grade GEO workflow into a single-click interface.
- Q04
Today - Running a friends-only beta. The current focus is timely idea discovery built on Google Trends data. Auto-publishing is not yet implemented.
- Q05
What's next - Target channels: blog, LinkedIn, YouTube, Instagram, X. We'll ship auto-publish once usage and revenue signals show up — then expand into per-brand customisation.
Project · 03
GolfShin
Me + friends- Q01
Why I built it - Booking a weekend tee time meant cycling through a dozen club sites by hand. The tee times of popular non-partner courses were scattered across the web. That was the starting point of my very first toy project.
- Q02
In one line - A unified view of tee times at 34 popular Korean golf courses the major booking apps don't cover — on a single screen.
- Q03
What makes it different - We scrape only the non-partner popular clubs, in real time, and bring them into one comparison interface. A small case of information-asymmetry, dissolved.
- Q04
Today - My friends and I use it for actual tee-time decisions. No monetisation in mind. I'd be happy if it lived on as 'a small AX example for everyday inconvenience'.
- Q05
What's next - The next step is automating the booking itself. The broader vision is for everyone to use Vibe Coding to solve their own everyday inconveniences — the same belief behind the free AI classes I give as a contribution.
Project · 04
Ikeike Kitchen
Site built · uploading existing content- Q01
Why I built it - A friend — an Instagram chef whose father is fighting cancer — created natural recipes for him, but they were scattered across SNS and never compounded. AI and IT felt distant to her, and that beautiful content was at risk of going dark. I couldn't leave it that way.
- Q02
In one line - A subscription recipe site that brings an Instagram chef's natural, seasonal recipes back to life with AI content and Coupang Partners.
- Q03
What makes it different - The core asset is her natural, local, in-season cooking — recipes she originally made for her father. AI handles content generation, Coupang Partners turns it into a frictionless ingredient-buying flow. Content and revenue move on a single rail.
- Q04
Today - Site construction is complete. We're uploading her existing content in stages.
- Q05
What's next - The next goal is to open the platform to other chef-influencers and grow it into a multi-creator content engine.
Project · 05
AstroECCOUNT
In production · deployed at Astros (my brother's trading company)- Q01
Why I built it - Astros, the trading company my brother runs, processed transaction data, reconciliation, and closing entirely by hand — three accounting staff carrying the entire load. Closing season meant overtime as the norm; people were carrying what the system should be carrying. Being a family company, it became the case I wanted to use to really prove out what AX should mean.
- Q02
In one line - An in-house AX system that automates the accounting and reconciliation of my brother's trading company Astros — designed as a 'human + system, working together' model, with the team operating MCP themselves.
- Q03
What makes it different - It started as a plain ERP automation. Mid-engagement I proposed swapping it for MCP (Model Context Protocol). Rather than 'the system does the work for you', I rebuilt it so the team gives Claude natural-language instructions for accounting tasks and validates the results. Vibe Coding training came as part of the package, so the team can now add and modify MCP tools on their own. The system keeps evolving in their hands, long after the consultant left.
- Q04
Today - 60% of what three people used to spend a full day on now runs through the system. Closing-season overtime dropped, and the team has moved from data entry to analysis and decision-making. The team operates MCP themselves today — a true human-plus-system workflow.
- Q05
What's next - Next, automation expands from accounting into customs clearance, logistics, tax, and internal controls. New automation isn't surfaced by the consultant any longer — the team discovers it and implements it as MCP tools. A 'human + system' AX model, proven first inside the family business.
Project · 06
Dental Insurance Term Helper (DIT)
v1 in production · Telegram @DentalInsuBot · Slack · Hospital B2B API- Q01
Why I built it - Dental insurance policy terms are hard for ordinary consumers to parse. Dental-treatment plans are non-standard and differ from insurer to insurer, while indemnity (silson) coverage follows government standard terms that vary by generation depending on when you enrolled. To actually confirm 'is my treatment covered' on the basis of the terms, you had to dig through dense policy documents yourself. I wanted to close that information gap with a tool that answers only from the policy clauses, never from guesswork.
- Q02
In one line - A RAG chatbot that searches dental and indemnity (dental-treatment) policy clauses and answers with the supporting source. Ask in Telegram and it finds the relevant clauses and replies with citations.
- Q03
What makes it different - The core is a grounded-or-silent principle. With no retrieved policy basis it declines rather than guessing, and citations are pulled only from the policy-clause metadata, never generated by the LLM. For indemnity it asks when you enrolled and points you to the matching generation's government standard terms. It makes no product recommendations, comparisons, or enrollment pitches — it is an information-only tool.
- Q04
Today - v1 runs in production (Telegram @DentalInsuBot and Slack). It loads a corpus of per-insurer dental policy terms plus the generational standard terms for indemnity insurance. It now also pins the exact policy version in effect at your enrollment month and cites it with its effective date (positive-pin), compares revisions across policy versions, warns about insurance fraud, and offers a B2B chatbot API channel that dental clinics can embed on their websites — all live.
- Q05
What's next - Next, we are broadening coverage to every indemnity generation and evaluating a usage-based pricing model.
Project · 07
Hanbyul LLC (법무법인 한별)
Live · 5 locales · daily legal-news pipeline- Q01
Why I built it - Hanbyul is a full-service firm covering corporate law, litigation, finance, IP, and M&A — but the web presence made none of those practice areas, or the attorneys behind them, legible. It is also the first door foreign counterparties walk through, so multiple locales were a requirement, and above all it had to be an asset that updates daily rather than a brochure that ages.
- Q02
In one line - A law-firm website on Next.js 16 and Supabase, serving five locales (Korean, English, Japanese, Chinese, Spanish), with legal news collected and summarised automatically every day.
- Q03
What makes it different - It isn't a static profile site. A Vercel cron job runs at 06:00 KST, crawls legal news, summarises it with GPT, matches each item to the attorney whose practice area it touches, and pushes the result through DeepL into all five locales. Notarisation attachments moved from the public folder into Supabase Storage behind RLS, and the repository's top-level rule forbids schema changes or destructive SQL without explicit approval.
- Q04
Today - Live, with the news pipeline running daily. Analytics were consolidated into a single GTM container and conversion events rebuilt on dataLayer. Feature development has quieted down; the site is now in a content-and-operations phase.
- Q05
What's next - Broaden the source set and practice-area tagging to sharpen news-to-attorney matching, and continue the GEO/AEO work for search and AI-engine visibility.
Project · 08
TheChain Lawyer
Built and operated, then handed over — domain and database moved to the attorney in Aug 2026- Q01
Why I built it - The practice of 이영남, a former deputy chief prosecutor with 21 years in the prosecution service. In blockchain, crypto-fraud, and Web3 matters, clients typically call without knowing what to even ask. The goal was to let someone work out the procedure and the issues at stake before that first phone call.
- Q02
In one line - A blockchain- and AI-focused legal services site in three locales (Korean, English, Chinese), with AI-assisted intake and precedent search covering the pre-consultation stage.
- Q03
What makes it different - The design draws a hard line so the AI never impersonates legal advice. Precedents are shown verbatim — holdings and summaries straight from the national statute-and-case API — and the site does not draft documents itself; it links to six Claude skills the attorney published, for the visitor to download and run in their own Claude. That structurally avoids exposure under Attorney-at-Law Act §109. Consultation content is stored encrypted, and phone-click and AI-consultation-start conversions are tracked through GTM.
- Q04
Today - After the build and the initial operating period, the domain and Supabase project were transferred to the attorney's own accounts in August 2026; the practice runs it independently now. The six legal skills published there are listed — with attribution — in SH Consulting's own Skill Library.
- Q05
What's next - The firm operates it going forward. Support continues in an advisory capacity for skill distribution and content expansion when asked.
Project · 09
Mindskin Dermatology
Live · 7 locales · automated content sync and translation- Q01
Why I built it - A dermatology clinic at Hongdae with two board-certified specialists. Treatment, device, and pricing information was scattered across ads and blog posts, which made it hard for a patient to pick only the procedure they actually needed — and as enquiries from Japan, China, Taiwan, Vietnam, and Thailand grew, localisation became a revenue path rather than a nicety.
- Q02
In one line - A dermatology site that manages procedures, devices, pricing, and journal content in a database and serves it in seven locales (Korean, English, Japanese, Simplified Chinese, Traditional Chinese for Taiwan, Vietnamese, Thai).
- Q03
What makes it different - None of the localisation is machine-converted. Traditional Chinese for Taiwan was translated fresh from the Korean rather than converted from Simplified, because Taiwan uses different words — 雷射 not 激光 for laser, 資訊 not 信息 for information — and a character converter passes all of that silently. Device and medical terminology is explicitly barred from DeepL and handled by a rule-prompted LLM instead: DeepL renders 탈모 (hair loss) as Japanese 脱毛, which means hair removal — the opposite. A terminology linter sits in the deploy pipeline as a gate.
- Q04
Today - All seven locales are live, with nightly news and blog sync jobs. Switching Naver blog ingestion from RSS summaries to full source posts raised median body length from 562 to 2,642 characters and eliminated truncated articles entirely. Search registration is complete on Search Console, Bing, and Naver Search Advisor, canonicalised on the apex domain.
- Q05
What's next - Keep expanding the device and treatment catalogue and the per-locale channels, and hold llms.txt and structured-data consistency so AI-search citations stay accurate.
