← This portfolio

this site · how I decide · 2026

How I decide: the options behind each call.

I work by laying out two to four options, each with what it costs, then picking one and writing down why. These are real choices from building this site and my projects, including the options I passed on.

Decisions 27 · Options weighed 80 · Went against the recommendation 5 times

  1. 1Lay out the options. Two to four, each with its trade-off in one line.
  2. 2Pick one, and say why. The recommended option is a starting point, not the answer.
  3. 3See it before committing. When it's visual, mock it up first, like the eight ideas I tried on a test page.
  4. 4Write it down. So the next decision starts from this one.

Design · This site · Sep 2026

How far should Apple's typeface go?

Went against the recommendation

  1. Apple's system font everywhereRecommended

    Real Apple type scale and weights. The site would read as native on a Mac.

  2. Poppins headings, system font for UI

    The brand stays mine; labels and buttons go native.

  3. Keep Poppins throughoutMy pick

    No typeface change. Apple's influence comes through color, spacing and motion.

Why

  • Friendly geometry: round letters feel warm, not corporate
  • One font, less noise
  • Continuity with earlier versions of the site

Design · This site · Oct 2026

My name mark: a red seal or a brush?

  1. A 도장 name seal

    Official and formal. The first version of Stamp used it.

  2. My name brushed in inkMy pick

    김채운 in black ink on paper.

Why

  • Shows a person: brush strokes carry motion and imperfection, so the mark reads as human

Design · This site · Oct 2026

Which design becomes the whole site?

  1. Desktop

    Folders and windows, a macOS metaphor.

  2. Ink

    Monochrome, brush wordmark, serif headings.

  3. Desktop Craft

    Apple-style polish, blue accent, glass menu bar.

  4. StampMy pick

    Quiet and gallery-like. Real work as the picture.

Why

  • Quieter: the desktop was fun but busy; Stamp gets out of the way
  • Faster to read: the work is one scroll away instead of behind windows

Design · This site · Oct 2026

What should "live previews" mean on the Builds cards?

  1. Click to try it liveMy pickRecommended

    Card keeps its screenshot; a button opens the running app. Nothing loads until asked.

  2. Always-live embeds

    The real sites running in every card. Most alive, but slow, and phones struggle.

  3. Animated previews

    Short looping recordings. Alive and fast, but not interactive.

  4. Link previews

    A preview image per project for shared links.

See it on the Builds tab →

Design · This site · Oct 2026

Which background?

  1. Cool lavender-blue washMy pickRecommended

    Lavender to near-white to sky. Light and calm, no warm tones.

  2. Aurora

    Slow-drifting blobs of lavender, sky, mint and peach. More personality.

  3. Subtle neutral hue

    White fading to a cool gray-blue. Closest to before.

Why

  • The work is the color
  • Calm, not loud
  • Light and airy, like morning light
  • Keeps black and white from feeling stark

Design · This site · Oct 2026

How should the homepage put builds first?

  1. Swap tabs: Builds firstMy pickRecommended

    Builds opens by default; Work moves to second. Smallest change, keeps the tabs.

  2. Featured builds above the tabs

    Keep Work as the default, but show three flagship builds before the tabs.

  3. One scrolling page, no tabs

    Builds, then Work, then About on one page. Best for skimmers, a bigger change.

  4. Keep Work first

    Outcomes at known companies first; builds one click away.

Why

  • I want people to see what I've built first

See the Builds tab →

Design · This site · Oct 2026

Inside Builds, what should lead?

  1. Live products firstMy pickRecommended

    The four apps you can try live, then more products, mods, coursework, side quests.

  2. Newest first

    One list by date. Shows momentum, but mixes big and small builds.

  3. Three flagships, then the rest

    Large cards for the strongest three, everything else compact.

Content · This site · Oct 2026

The site's numbers predate my resume. Which should it use?

  1. Match the resumeMy pickRecommended

    Update every page to the resume's figures.

  2. Keep the site's numbers

    Only if both sets were true and the resume just shortened them.

  3. Go line by line

    Decide each figure separately.

Content · This site · Oct 2026

When should the new resume and headlines go live?

  1. After the numbers are fixedMy pickRecommended

    Ship together, so the live resume never contradicts the homepage.

  2. Push now

    Live right away, with the resume and homepage disagreeing for a while.

  3. Not yet

    Review on the test site first.

Content · This site · Oct 2026

What should the homepage design notes cover?

  1. Stamp, plus the journeyMy pickRecommended

    The current tokens and type, with a short strip on how the site got here.

  2. Stamp only

    Just the current design system.

  3. Bring back Desktop Craft

    Restore the desktop design first, then document it.

Read the design notes →

Content · This site · Oct 2026

What happens to the draft case study on choosing a brand?

  1. Rebuild for Stamp, publish laterMy pickRecommended

    Restyle it, update the ending, and publish once it's in my own words.

  2. Fold it into the design notes

    Drop the separate page.

  3. Leave it as is

    Don't touch the draft for now.

Content · This site · Oct 2026

Is a hand-brushed 김채운, traced to vector, still a next step?

  1. Yes, still plannedMy pick

    Keep it on the backlog.

  2. Maybe later

    An idea, not a commitment.

  3. Drop it

    Remove it from the plan.

See the backlog →

Products · AI Visibility Checker · Oct 2026

Where should the checker live?

  1. Its own app and repoMy pickRecommended

    A new Vercel app with a serverless function that fetches the URL, plus its own case page.

  2. A page inside the portfolio

    Easiest to find, but mixes a product into the portfolio's code.

  3. A Cloudflare Worker

    Fast and free, but a different platform from my other web apps.

AI Visibility Checker →

Products · AI Visibility Checker · Oct 2026

How much should version 1 check?

  1. One page, every signalMy pickRecommended

    Crawler rules, llms.txt, sitemap, structured data, previews, and text without JavaScript. Free and fast.

  2. Crawl the whole site

    The same checks across many pages. Slower, and more requests to other people's sites.

  3. Also ask AI assistants

    Ask a model whether it cites the site. Closest to real GEO, but it costs money per check.

Products · AI Visibility Checker · Oct 2026

Should the code be public?

  1. PublicMy pickRecommended

    The case page can link to the source.

  2. Private

    Keep it private for now, possibly for Lucent.

Products · Palette Generator · Oct 2026

Where should the weekly trend scan read brand colors from?

Went against the recommendation

  1. Startup listsRecommended

    New Product Hunt launches plus a curated list. Needs an API token.

  2. A curated list onlyMy pick

    Hand-picked sites kept in the repo. No token; "latest" means the sites we add.

  3. Design galleries

    The most designed sites, but their terms prohibit scraping.

Palette Generator →

Products · Palette Generator · Oct 2026

How should Explore show each scanned site?

  1. Name and paletteMy pickRecommended

    Company name, its palette, and a link. No screenshots or logos.

  2. Palette only

    Anonymous palettes by mood and industry. Safer, less interesting.

Products · Palette Generator · Oct 2026

How often should the trends refresh?

  1. WeeklyMy pickRecommended

    Current enough; about 150 site visits a week.

  2. Daily

    Fresher, but brand colors rarely change daily, and it means 7x the deploys.

  3. Manual

    Only when I ask.

Coursework · Spotify Review Triage · Oct 2026

The 10,000 development reviews were already labeled. Reuse those labels in the final run?

  1. Reuse the development labels

    Keep the earlier labels for those 10,000 reviews.

  2. Classify them again in the final runMy pick

    Label them fresh with the final setup.

Why

  • Every graded label should come from one final configuration

Spotify Review Triage →

Coursework · Spotify Review Triage · Oct 2026

How close must the model's sentiment score be to my label to count as a match?

  1. Exact

    Any difference counts as wrong.

  2. Average error only

    Report the mean error, with no pass or fail.

  3. Within ±0.5, with ±0.25 also reportedMy pick

    The same or the next step on a five-step scale. The average error is reported too.

Why

  • Sentiment runs from -1 to 1, so an exact match would mark near-identical scores wrong

Coursework · Pac-Man DQN · Sep 2026

How much should the agent explore?

  1. 0.20My pickRecommended

    Balanced: random one move in five after warm-up.

  2. 0.10

    Follows its own estimates more; settles into a habit faster.

  3. 0.30

    More variety in memory, less practice of its own policy.

Pac-Man DQN →

Coursework · Pac-Man DQN · Sep 2026

How many training games?

  1. 100My pickRecommended

    Best shot at real learning, with four checkpoints along the way.

  2. 50

    A middle ground, two checkpoints.

  3. 25

    Meets the minimum; learning may be modest.

  4. 5

    A setup check; may learn nothing.

Coursework · Pac-Man DQN · Sep 2026

Which learning rate?

Went against the recommendation

  1. 0.0001Recommended

    The standard value. Stable, best with longer runs.

  2. 0.00025My pick

    Slightly faster learning, so change shows up in a short run.

  3. 0.001

    Aggressive; values can blow up.

Why

  • 2.5x the reference, so change would be visible in only 100 games, without going as far as 0.001

Coursework · Pac-Man DQN · Sep 2026

Where should training run?

Went against the recommendation

  1. Google ColabRecommended

    A cloud GPU, matching the assignment guide.

  2. Locally on my MacMy pick

    Apple GPU acceleration, no session expiring mid-run.

Coursework · Personal Wiki · Sep 2026

Which sources should a fully offline study wiki use?

  1. Three notes I write

    Fully shareable and easy to verify.

  2. My Custom LLM write-up

    My own project, my own words.

  3. Wikipedia articles

    Course topics, saved locally with attribution.

  4. A mixMy pick

    My Custom LLM write-up plus two Wikipedia articles: retrieval-augmented generation and transformers.

Personal Wiki →

Coursework · Personal Wiki · Sep 2026

One source was too long for the local model. Cut it off, or fix it?

  1. Fix it: split into a third subjectMy pickRecommended

    Every part reaches the model, and each call stays short.

  2. Fix it: send more per call

    Fewer notes, but longer prompts and slower calls.

  3. Keep the cut-off

    Document it as a known limitation.

Coursework · Personal Wiki · Sep 2026

The offline recording missed some required chat checks. Re-record, or document the gap?

Went against the recommendation

  1. Re-record them offlineRecommended

    About two minutes with Wi-Fi off.

  2. Submit, and document the gapMy pick

    Say plainly in the README which checks ran online.