Hyunjun Kim MethodENTHKOJABangkok --:--Last build shipped 18 Oct
Hyunjun Kim

Hyunjun Kim I plan products and build them by directing AI. Before this I ran the web agency Sky C&I as its CEO and planner.

Nine products so far, from online stores and games to native iOS and Android apps, web platforms and business tools. I write the spec and set the bar, AI writes the code, and I check every result. Five are live in stores or on the web.

  • 안녕하세요
  • こんにちは
  • Hello
  • สวัสดี
Find the right buildContact
Scroll · drag a phone to turn it
MomentoiOS · AndroidTestFlight

A couple's map of memories.

Two native apps, SwiftUI and Jetpack Compose, on one backend. Places you visit together pin themselves to a shared journal.

Dev start to TestFlight
~7–8wks
Native platforms, one backend
2
iOS feature modules
24
  1. Journal timeline
  2. Shared calendar
  3. Home-screen widgets
Momento journal screen
Case study · MomentoPlanning · setting the bar · directing AI
First instruction to AI · 21 Mar 2026 When a couple is in roughly the same place and one of them takes a photo, collect it with the date and location. Can this be done?
Momento onboarding
Onboarding
Momento map of shared memories
Map · same time, same place
Momento home-screen widgets
Home-screen widgets
Momento journal timeline
Journal timeline
Same screen, two native apps

Drag to compare iOS and Android.

The calendar built twice, in SwiftUI and in Jetpack Compose, from one feature list and one backend. Not ported code: each follows its own platform.

Momento calendar on iOSMomento calendar on Android
iOSAndroid

  1. The bar I set

    “Don't port iOS code to Android. Follow each platform's conventions.” Two native apps on one backend.

  2. Where it got stuck

    Server dates arrived in mixed formats and broke iOS decoding. Comparing locations on the phone drained the battery.

  3. How I fixed it

    Directed a decoder that accepts several formats, and moved location matching to the server as a SQL function.

  1. First prompt
  2. First working build
  3. iOS on TestFlight
Read the full case
ProblemPlan & barDirecting AIFixesResults
MECHOOSEWeb · iOS · AndroidLive

Thai fashion, tried on by AI.

A live commerce app with real payments at mechoose.store. Shoppers send one photo and a stylist flow picks the outfit.

Idea to live web service
~1wk
Idea to App Store approval
~6wks
Backend API routes
137
  1. Category browse
  2. AI try-on flow
  3. Product detail and cart
MECHOOSE category screen
Case study · MECHOOSEPlanning · directing AI · deployment
First instruction to AI · 3 Feb 2026 I have no coding background, so I'm building the commerce app with Figma Make and GPT. My Page keeps erroring and I'm stuck. Can you help?
mechoose.store
mechoose.store home page
Live web store
MECHOOSE checkout
Checkout · real payments
mechoose.store/admin
MECHOOSE admin dashboard
Admin dashboard

  1. The bar I set

    Missing delivery info shouldn't block checkout. Collect it at app entry instead.

  2. Where it got stuck

    Interrupted payments left orders stranded, and hand-typed addresses caused failed deliveries.

  3. How I fixed it

    A resume-payment flow from a banner and My Page, plus address autocomplete that fills the postal code.

  1. First prompt · first working build
  2. Web live at mechoose.store
  3. App Store approval
Read the full case
ProblemPlan & barDirecting AIFixesResults
AttenNFCiOS · Android · WebIn development

Attendance you can't fake.

NFC and GPS attendance with automatic payroll for small Thai businesses. Designed in depth first; the core logic is running.

Idea to core logic working
~9days
Check-in verification layers
4
Payroll variables, configured
9
  1. NFC check-in
  2. Shift view, dark
AttenNFC check-in screen
Case study · AttenNFCPlanning · architecture & security · directing AI
First instruction to AI · 8 Sep 2026 Let's finish the full screen and feature design first for an NFC-tag and server attendance app. Flutter does both iOS and Android, right?
AttenNFC check-in on a phone
Check-in
stitch · onboarding
AttenNFC onboarding design
Onboarding, designed in Stitch
AttenNFC in dark mode
Dark mode
stitch · formula engine
Payroll formula engine design
Payroll formulas as settings
stitch · anomaly queue
Anomaly review queue design
Anomalies go to admin review
stitch · audit log
Audit log design
Audit log

  1. The bar I set

    Payroll formulas are settings, not code: nine variables an admin edits on screen.

  2. Where it got stuck

    Generated code fell short of the bar: five of nine variables, an NFC API mismatch and hardcoded GPS coordinates.

  3. How I fixed it

    Kept the schema and the design, had the code rewritten from scratch, and confirmed static analysis reported no issues.

  1. First prompt
  2. Core logic working
Read the full case
ProblemPlan & barDirecting AIFixesResults
Money NavigationiOS · AndroidReleased

Built in 48 hours.

A travel money helper from a two-day contest. It reads ATM fees and helps you haggle in Thai. Released on GitHub, as an APK and on TestFlight.

Contest start to submission
48h
Parallel AI sessions
8+
Public release channels
3
  1. Type the price
  2. Fair-price verdict
  3. Show the seller, in Thai
Money Navigation price verdict screen
Case study · Money NavigationU-Hackathon 2026 · AI track · Bangkok
Rule for the AI team · contest day If there are several sessions, make one the PM to manage order, and report status to me right away. After that I talked only to the PM session, which relayed every session's report in Korean.
01 · The brief

I brief one PM, not eight sessions

The plan, the bar and the house rules went into one board file. Every session read it first, and from then on I only talked to the PM.

02 · Fan-out

The PM opens the team

Eight working sessions, each in its own lane, plus 10+ Codex runs on the side.

Low-risk lanes switched to Sonnet
03 · Relay

Every report comes straight to me

Sessions report to the PM, and the PM tells me right away. That left me free to pull screen designs instead of watching terminals.

04 · Ship

48 hours to a public release

A GitHub release, an Android APK and a TestFlight build, all submitted inside the contest window.

0 false positives · 51 receipts
BOARD.mdHouse rules
  1. Touch only your own files
  2. One build at a time
  3. Only the PM opens sessions
  4. Stage only your own paths
PM → me · paraphrased

Auth reported done. The UI lead gets the next build.

Release is waiting for the current build to finish.

Pitch posted a first draft of the demo order.

Contest48 h
Enter a price
Type the price
Fair price verdict, ask for 110 baht
Verdict · ask for 110
Show the seller in Thai
Show the seller
github.com/g326555-lab
Public GitHub release
Public release v1.0.0-7

  1. The bar I set

    “Give it depth like a big-company app. It must not look AI-made.”

  2. Where it got stuck

    The ATM scan looked right on the phone, but could it be faking the demo numbers?

  3. How I fixed it

    Re-ran it on photos with different numbers and fixed column alignment: 0 false positives across 51 real receipts.

Read the build log
Topic choicePM modelFixesRelease
THE HOUSEUnity 6 · Android · iOSIn testing

An escape room, from zero engine experience.

A horror escape-room game taken from idea to store build, localized into four languages.

Idea to store build
~4wks
Store builds iterated
10+
Lines of game C#
17.7k
  1. “It's getting closer.”
  2. The ward
  3. The reaper door
THE HOUSE opening scene
Case study · THE HOUSEPlanning · level design · build & release
First instruction to AI · 20 Sep 2026 I'm thinking of making an escape-room game. Can I do this alone?
696instructions in a single conversation, the most in any of my AI work logs.
THE HOUSE scene: hold your breath
Real-time 3D with free look, not fixed panoramas01
THE HOUSE pawn shop scene
Every object is a clue, dummies included02
Store images in four languages
Store listing in English, Japanese, Thai and Traditional Chinese03

  1. The bar I set

    Joystick plus drag-to-look in real-time 3D, not fixed panoramas. English first, $1 per scenario.

  2. Where it got stuck

    Bundling every asset made the install too large, and hand-built versions kept missing settings.

  3. How I fixed it

    Checked the official docs, then directed on-demand asset delivery and a scripted build so every version ships the same way.

  1. First prompt
  2. First playable
  3. Store build
  4. Build 1.0.10_27
Read the full case
ProblemPlan & barDirecting AIFixesResults
AI Video StreamingWeb, serverlessLive

Streaming with no servers.

A video service on the Cloudflare edge: ten languages with AI-translated titles, and crypto subscriptions verified on-chain.

Languages
10
Versions in about three weeks
~70
Database migrations
39
  1. Home
  2. Player
  3. Subscribe
Streaming home screen
Case study · AI video streamingPlanning · bar · directing AI · deploy
The bar I set Payments must be verified, not trusted. A subscription starts only after the payment is confirmed independently. The live address and real content stay private on purpose. These screens are the real UI, run locally with sample data.
localhost · home
Streaming home on desktop
Desktop home
Mobile home
Mobile home
Mobile player
Player
Subscribe screen
Crypto subscription
Subscribe screen in Japanese
Same screen, Japanese
localhost · player
Desktop player
Desktop player

  1. The bar I set

    No servers at all: compute, database, media, rate limits and AI translation all on the edge.

  2. Where it got stuck

    About seventy versions in three weeks made it easy to lose track of why something changed.

  3. How I fixed it

    A running README explains every version, and one idempotent script makes each release a single command.

  1. First build on Workers
  2. Live, ~70 versions in
Read the full case
ProblemPlan & barDirecting AIResults
How I work

Spec-first solo build

The seven steps I repeat on every project, from the first rules document to a launch. Steps 3 to 6 run as a loop: instruct, check, correct.

01

Write the rules first

Quality starts in a rules document: how we work, the don'ts and the bar to verify against. Every session reads it before it starts.

02

Lock screens and data first

Designing and building at once gets tangled, so every screen and the data come first. Then they get wired in code.

03
Repeat steps 3–6

Small steps, checked by me

Not “it's done”. I run it on a phone or simulator and confirm it with my own eyes.

04
Repeat steps 3–6

A bar in words, not scores

I set quality by reference, big-company apps and tone, and let AI choose how to build it.

05
Repeat steps 3–6

Catch it and turn

I distrust plausible success (“this isn't hardcoded, right?”) and pivot as soon as something stalls.

06
Repeat steps 3–6

Evidence over self-report

AI's “I checked” is not proof. I re-confirm with the real code, the build, or different inputs.

07

Run AI like a team

Context lives in handover docs, and on multi-session work one session is the PM.

The full method

The rules behind every build

The seven steps above are the outline. The full page has the principles I hold AI to, the prompt skeletons I reuse, how I catch errors, and the mistakes that became rules.

  1. 01Framework: 7 steps from zero to launch
  2. 02Principles
  3. 03Prompt templates
  4. 04How I catch errors
  5. 05Mistakes → lessons
  6. 06The work in numbers
Read the full method
Smaller builds & experiments

Also built

Not every build is a flagship. These are the experiments, tools and content work around the main products.

iOS · Android · Experiment

Adonis AI

A camera-based AI trainer: on-device pose detection, no upload. I pivoted to Money Navigation at the hackathon.

Open case
Web · Private

MemoHub

Scattered notes and Telegram messages gathered, summarized by an LLM, with same-day reminders. Deployed on Vercel.

Open case
Server · Telegram

Trading bots

Crypto price and kimchi-premium alerts to Telegram. Used for information only, never auto-trading.

Open case
Content

H2 — AI short drama

Scenes generated with Google Flow (Veo 3.1) and edited in CapCut. Video content, not an app.

Ops for THE HOUSE

X / Threads automation

Posting automation with rules, drafts and logs as documents. AI drafts, I approve and publish.

Ideas I'm exploringMobile wedding invitationsKorean-style invites for the Thai marketX-ray AssistAge estimates from dental X-raysAI + OSAn assistant woven into AndroidAI videoShort-drama production with AI tools
Honest review

Solo with AI vs a dev team

Five projects' code checked against one rubric and compared with the same thing built by a two-to-three person team. Every gap comes with what I'm doing about it.

15Ahead of a small team
18On par
21Next steps I'm filling in now
ProjectAheadOn parFilling in now
THE HOUSE
Speed & costRelease & opsLocalization
Feature scopePerformanceReadabilityDocumentation
Testing & automationArchitectureVersion controlSecurityAccessibility
MECHOOSE
Feature scopePerformanceSpeed & cost
Security basicsRelease & opsLocalizationDocumentation
Testing & automationArchitectureCode consistencyHistory & monitoring
Momento
Feature scopeSpeed & costLocalization
ArchitectureCode hygienePerformanceSecurity basics
Testing & automationRelease & opsSecurityAndroid completenessAccessibility
AttenNFC
SpeedDocumentation
ArchitectureReadabilityPerformance
Feature completenessSecurityTesting & CILocalization & accessibility
AI video streaming
PaymentsSecurity basicsPerformanceSpeed & documentation
ArchitectureRelease & opsLocalization
Testing & version controlCode consistencyMonitoring

Speed, scope, localization and shipping are ahead. Automated tests, code structure and server-side hardening are behind, and I'm fixing them in order, money-critical parts first.

Before AI · 2022–2025

Two companies, founded

Before building with AI I ran a web agency and a medical-tourism platform. The job was the same as now: turn a client's problem into a spec, then direct the people who build it.

Sky C&IFounder · Service planning / PM2022 – 2025

A web agency, run from the spec

I met clients, defined requirements and drew every screen and feature into specs, then directed the dev team on logic and reviewed what came back. I did not write the code.

20+
Client projects
18
Member team
K-ETA screen
K-ETAVisa Korea · travel authorization
Zero call screen
Zero callZero Cool Tour · dispatch & settlement
FOHHN AUDIO screen
FOHHN AUDIOSite renewal + CMS
DEEPFARM screen
DEEPFARMSmart-farm hybrid app
VLYFT screen
VLYFTMJet Family · creator platform
DL E&C screen
DL E&COpen-innovation campaign site
AUMSSA screen
AUMSSANFT & metaverse brand site
THE e-roonFounder2022 – 2025

Wooah, Korea's clinics for the world

A platform matching international patients with Korean cosmetic clinics, set up to handle medical visas. It solved visas and trust for patients and no-shows for clinics.

  • Planning, brand and operations, 150+ partner clinics signed
  • Revenue beyond commission: ads, stays and transfers, visa services
  • Sold to a China-based company
150+
Partner clinics
3.5×
Revenue growth
Acquired
Wooah home
home
Wooah clinic search
clinic search
Wooah coordinator chat
coordinator chat
Wooah quote
quote
Wooah reviews
reviews
Where to start

What can I build for you?

Pick a kind of product. Each one opens the build that shows I can do it.

MECHOOSEWeb · iOS · Android · Live

A full store with real card payments, refunds, an admin panel and a stylist app, live on the web and both app stores.

~1 wk
Idea to live web
137
Backend API routes
Read the MECHOOSE case
Open case
Contact

Want to know more about me?

jun326555@gmail.com
© 2026 Hyunjun KimKorean · Japanese · EnglishBangkok, 3+ years