Founder portfolio
I read what people want first. Then I build.
As the founder of a two-person studio I opened and closed ten products in a year. Every case below follows the same order: the desire or anxiety I saw, what I decided and why, how the AI and product were designed, and what the numbers said back. The numbers that disappointed and the things I killed stay in.
What I do as a founder, in six stages
Find the problem, check the money, ship fast, learn from the numbers, run many products through automation — or kill it. Every figure under each stage comes from the cases below.
01
Finding the problem
Before building the app, I measure the desire with a single ad line.
- Fake-door ads, three creatives: CTR 13.43% · 7.39% · 5.33% — the winning line became the product’s core promise. Email conversion, though, was 1–5.7% against a 20% bar set in advance, and going ahead with the app anyway is kept on record as a lesson
- A 16% click-through judged separately from a 0-second median session — interest is not use
02
Market & revenue model
I design the revenue model with unit cost and regulation as constraints, not afterthoughts.
- Moved an AI companion from a daily-cap subscription to an in-app currency, using photos as the margin lever
- Proposed collapsing per-app server billing into one shared backend → first deploy two days later, 446 server tests, 0 failures
- Fixed the entry fee of a prediction product at zero so it stays clear of gambling regulation
03
Shipping fast
Two people and a set of AI agents opened and closed ten products in a year, and took several of them to the stores.
- App Store rejection → both causes fixed the day an agent investigation confirmed them → approved and live 9 days after the rejection
- AI keyboard: from idea to a Play closed test in two days
04
Learning & pivoting from metrics
When a number is wrong I retract my own metric first; when the result isn’t there I change direction.
- Retracted six of my own metrics as measurement artefacts (a 38-second dwell was really 1.5 seconds)
- Trend app: a 10-second engagement threshold found in logs from 2,013 devices → six redesigns → B2B pivot
- AI companion: share of users who signed up and never spoke, before vs after the in-app currency launch, 39.6% → 19.1% — logged together with the caveat that other changes in the same window could not be separated out
05
Running many products through automation
I designed my own attention and decisions with AI. Judgements arrive as cards; I approve.
- About 30 recurring automations on a personal server; AI calls cost $1–2 a day (list-price equivalent, two-day measurement); a dead process is back within 11 seconds
- AI agents build; a separate verification agent has to pass the work before it ships — otherwise it goes round again
- 98 of 102 “ship it?” approval requests were approved as-is → verified changes are now delegated
06
What I killed
One rule: if early revenue doesn’t show up, we stop.
- Shut down a low-viability project three days after its first commit
- Removed AI from a keycap keyboard on purpose — the product is play and decoration, not intelligence
- Moved a trend service’s centre of gravity from the app to a site subscription and B2B
Desire → design → result: nine cases
01Intimacy without intrusion
Beff
An AI friend who lives in your neighbourhood — and doesn’t only wait, but sometimes reaches out first.
Role · Co-founder · product · conversation design · growth
- Decision & grounds
- I started from a desire people already pay for — fandom. Three ad creatives (CTR 13.43% · 7.39% · 5.33%) picked the core promise, and billing moved from a daily-cap subscription to an in-app currency. I designed the currency and its name; my co-founder built the price table.
- AI & product design
- The first hypothesis was “someone worth waiting for”, with human-like reply gaps. Reading real conversations, I named “the AI isn’t interested in the user” as the churn cause, and the data overturned that hypothesis. My co-founder removed the “walk-away” mechanic and rewrote the rule prompt around intent; skipping a first-turn preparation step cut the first reply from 25 to about 5 seconds. I found the research behind it: delays of 15 seconds or more hurt how the relationship is perceived.
- Result
- 4.5 stars on the Korean App Store. The before and after periods differ in length and acquisition channel, so I record that the currency change alone could not be isolated.
02Attention — when waiting is hard to bear
ADHD & AI silence
An experiment on how an AI’s silence while it prepares an answer is experienced by people with ADHD traits.
Role · Study design · experiment app
- Decision & grounds
- Instead of assuming faster is better, I asked whether the way the same wait is shown changes the experience.
- AI & product design
- A three-condition study with an identical 78-second total wait, run in an app I built for it. Switching the manipulation check to a permutation test cut its false-alarm rate from 10.4% to a nominal 0.1%.
- Result
- Paper under submission.
- Finding the problem
- Learning & pivoting from metrics
03Social anxiety — finding it hard to say no
Excuse me
One convincingly real booking call, so you can leave an uncomfortable situation without being rude.
Role · Founder · product · revenue model
- Decision & grounds
- Chose a full renewal before launch: a robotic voice breaks the whole premise of a believable call. Payments went from “coming soon” to a real structure — two free voices, paid voices, and a Pro tier.
- AI & product design
- Natural human voices and longer background noise remove the doubt that the call is real; copy was stripped of AI-sounding phrasing and the legal-risk notes were written first.
- Result
- Preparing for store release. No revenue yet.
2
Free voices
4
Store products set up
- Market & revenue model
- What I killed
04Social anxiety — closing the chat because you don’t know what to say
K’ai
An AI keyboard that drafts the one-line reply and polishes Korean in group chats and DMs.
Role · Founder · model evaluation · pricing
- Decision & grounds
- Took a flow already proven daily on my own keyboard and made it a product. The differentiating feature was chosen by model evaluation, not taste.
- AI & product design
- A low-cost model keeps unit cost down, with an evaluation bar that forbids touching sentences that didn’t need fixing (zero over-correction).
- Result
- Demand is not yet validated.
99/100
Polish eval
0/36
Over-corrections
~₩0.03
Cost per call
2d
To Play closed test
- Shipping fast
- Market & revenue model
05Uncertainty — you can’t know if a job fits until you live it
LiveLike
Live a job you can’t easily try, with AI, and feel its judgement calls and pressure first-hand.
Role · Product · launch · store
- Decision & grounds
- Four days after an App Store 2.1(a) rejection, an agent investigation pinned both causes (a temporary server address, and running out of currency right after login); both were fixed that same day.
- AI & product design
- In its predecessor, a job simulator, grading was checked against working professionals so that plausible-but-wrong answers stopped passing. The product then pivoted to a chat format and became LiveLike.
- Result
- The resubmission was approved first time. Fast to ship; demand not yet proven.
- Shipping fast
- Learning & pivoting from metrics
06Habit — breaking the three-day cycle
PUSHERS
One push-up a day. The front camera counts; streaks and friend battles keep you going.
Role · Founder · product · growth
- Decision & grounds
- Launched with zero server cost. 198 cheat-motion combinations, generated in 3D simulation, were run against the camera engine and passes went from 39 to 0; the remaining tap-to-count path is being closed. A habit app you can fool is worthless.
- AI & product design
- “One a day” lowers the bar; the countdown voice and extra explanation were cut. Next on the design list is a streak / friend-battle / reminder loop.
- Result
- Live on the App Store. Day-7 retention is still 0%, so the habit loop itself is being redesigned.
68/71
Camera counting, 6 public videos
39→0/198
Camera cheat combos passing
59→9위
Brand search rank, 3 days
₩0
Server cost
- Shipping fast
- Learning & pivoting from metrics
07Reward & play — fun at your fingertips, a group with nothing to do
Key’s · Crobby
A keyboard where you collect keycaps and their sounds, and an outdoor cops-and-robbers game where the server referees on a map.
Role · Founder · product direction
- Decision & grounds
- AI was deliberately left out of the keyboard: what it sells is sound, motion and decoration, so the effort went into making the sound real. The outdoor game dropped childish characters for adults, and lets people start without logging in until the core action.
- AI & product design
- A reward loop (coins → keycap draws) plus audio-visual moments worth sharing; the outdoor game is judged on one question — does the same group come back every week?
- Result
- Both are MVPs in pre-release testing.
- Market & revenue model
- What I killed
08Curiosity — 90 minutes of short video a day, nothing left over
Tremer
Turning time spent on short video into time spent predicting trends.
Role · Co-founder · research · metrics
- Decision & grounds
- ₩15,000 of ads bought 1,137 clicks (₩15 a click), but behaviour was judged separately: a 0-second median session and 6.0% prediction participation. Verdict: interest, not use. The prediction entry fee was fixed at zero to stay clear of gambling regulation.
- AI & product design
- Found a 10-second engagement threshold in behavioural logs from 2,013 devices, and retracted six of my own metrics as measurement artefacts.
- Result
- After six redesigns the centre moved to a site subscription and B2B trend reports. The lesson is about order: willingness to pay should have been tested first.
16%
Ad CTR
0s
Median session
2,013
Devices analysed
6
Own metrics retracted
09My own attention — not enough hours to run many products alone
Dan OS
A personal operating system: AI turns the day’s data into decision cards; I approve or correct them.
Role · Design · operation (internal tool)
- Decision & grounds
- The scarcest resource is my own time. Notifications collect in one place, every automation follows “AI proposes, I approve”, and decisions where I am the bottleneck get delegated once the evidence piles up.
- AI & product design
- AI agents build, and a separate verification agent has to pass the work before it ships. Every approval or correction I make becomes input for the next judgement.
- Result
- This site update was made the same way and went live only after passing independent verification.
- Running many products through automation
If you want to start from what people actually want
Co-founding, product or research collaboration, investment — email is the way in.