Mindy's Portfolio
Live: mindy-portfolio.vercel.app
My personal site — a clean Portfolio and History up front, an About page, and a set of small AI/data products you can click into. Plain HTML/CSS/JS, no build step, with a few Vercel serverless functions in /api for the interactive pages.
What's on the site
- Home (
index.html) — a short intro, a Portfolio list of everything I'm building (name + one-line focus), and a History timeline (roles, internships, NYU).
- About (
about.html) — who I am, why this site exists, and what I believe about building with AI.
- Buddy — Wellness Buddy, a wearable-free wellness check-in prototype with its own repo: mindyjwu/wellness-buddy, live at wellness-buddy-chi.vercel.app. Linked from the Portfolio list like Global Explorer and GenAI.
/wellness on this site redirects there.
- The AI Ecosystem — an interactive map of who funds, builds and powers AI: 26 layers and ~320 players, from what people spend on down through labs, clouds, chips, fabs and energy to raw materials. Its own repo: mindyjwu/ai-eco-diagram, live at ai-eco-diagram.vercel.app. Linked from the Portfolio list like Buddy.
- Mood Wall (
mood.html) — tell it how you feel and get a wall of real art, books, music and film chosen for that feeling. Claude picks; the browser resolves every pick against The Met, Open Library, iTunes (30-second previews) and Wikipedia, so every card is a real thing with a real link. Save cards to a local wall, share a mood by URL. Fifteen walls are hand-authored in data/moods/source.json and snapshotted once into static JSON so they load instantly and cost nothing.
- SIGNAL (
signal.html) — an AI-news feed across 16 industries, built to refresh each morning, with an estimated read time on every story, a shared archive of past sweeps, and a downloadable podcast script.
- Turnout (
turnout.html) — a rotating-host community. A fair rotation wheel, themed party ideas ranked by an on-device logistic-regression model that learns from ratings, a hill-climbing guest matcher that mixes friends with new people by shared interests, a host-cost model with a circle pot, and a post-party "lock in new friends" network. Demo build: state lives in localStorage, no backend yet.
- Stock Advisor (
stock-advisor.html) — scores stocks on fundamentals, technicals and news sentiment, then grades its own calls. Sample data, not investment advice.
- People Like Me (
people-like-me.html) — vetted beauty and body providers, ranked by people who share your hair, skin type and budget.
- Festival Clash Resolver (
festival-clash-resolver.html) — an offline-first set-clash resolver and route planner for electronic-music festivals.
- Plus external builds linked from the Portfolio: Global Explorer (a 3D globe of cities worth visiting) and GenAI.
Stack
Plain HTML, CSS, and JavaScript — no framework, no build step, so it loads instantly for anyone. The interactive pages call the Claude API through Vercel serverless functions in /api. A small Python script filters MoMA's open dataset for the Mood Wall art.
Repo layout
*.html one page per project (plain HTML, no build step)
css/, js/, assets/ shared styles, scripts and images
api/ Vercel serverless functions (Claude calls, Turnout alert email)
data/ committed Mood Wall and SIGNAL snapshots
scripts/ snapshot tools and the MoMA dataset filter
tests/ Node tests for the Turnout alert function (npm test)
docs/turnout/ Turnout guides, plus sql/ (numbered migrations and database tests)
docs/ other product notes
Running it locally
No build tools needed for the front end — open index.html directly, or serve the folder:
python3 -m http.server 8080
The /api routes need a Node environment that supports Vercel serverless functions:
npm install -g vercel
vercel dev
Content snapshots (demos and credit control)
Both the Mood Wall and SIGNAL commit their last real result to the repo, so they load instantly, cost nothing during demos, and fall back gracefully if the Anthropic account is out of credits. Anything that isn't today's data is labelled with its real date, never passed off as "today".
# Mood Wall
node scripts/snapshot-moods.mjs capture # resolve the authored walls into static JSON (spends credits)
node scripts/snapshot-moods.mjs freeze # typed moods serve the nearest curated wall, never call the API
node scripts/snapshot-moods.mjs unfreeze
node scripts/snapshot-moods.mjs status
# SIGNAL
node scripts/snapshot-signal.mjs capture # pull a fresh sweep and save it (the only command that spends credits)
node scripts/snapshot-signal.mjs freeze # page serves the snapshot only, never calls the API
node scripts/snapshot-signal.mjs unfreeze
node scripts/snapshot-signal.mjs status # what's saved, and how many days old it is
node scripts/snapshot-signal.mjs archive # save the current snapshot into the archive (seeds a day)
node scripts/snapshot-signal.mjs index # rebuild data/signal/index.json from the day files
Commit and push after any of these. Freeze before a demo: zero API calls, instant load, no cold-start wait.
SIGNAL refreshes itself daily
.github/workflows/signal-snapshot.yml runs capture --scheduled every morning (10:17 UTC, about 6am ET), commits the result, and Vercel redeploys. Visitors only ever read the committed file, so traffic never costs credits. This job is the one thing that spends: three Claude calls a day.
- Refresh now: Actions tab → Refresh SIGNAL snapshot → Run workflow.
- No secrets to set up. The job calls the public endpoint; the Anthropic key stays in Vercel.
- It never replaces good data with worse. The run fails, and the previous snapshot stays live, if a sweep errors (including out-of-credits), if the edge returns a stale cached sweep, or if the result has fewer than 8 stories or 5 industries.
- The page tells you when it stops, however the snapshot was made. The badge reads
DAILY while an auto-refreshed snapshot is within two days old. At three days, any snapshot flips to an amber STALE and the status bar says how many days old the stories are. (An earlier version only warned about snapshots the daily job had written, so one that had never been refreshed looked intentionally frozen.) GitHub's default notifications also email the account that last edited the schedule when a scheduled run fails.
- If a run fails with a credit-balance error, it is billing, not code. Top up under Plans & Billing and re-run.
The archive
Each successful capture also saves its sweep as data/signal/<date>.json, and data/signal/index.json lists the days. The Archive tab reads those files, so it is one shared history rather than each visitor's browser memory. The index is fetched only when Archive is first opened, and day files are fetched seven at a time. The index is always derived from the day files (snapshot-signal.mjs index rebuilds it), so a bad merge can't leave it out of sync. A run that fails writes nothing to the archive. The workflow commits new day files with git status, not git diff, because a new file is untracked and git diff cannot see it.
The archive keeps every day. It adds roughly 10 to 15 KB a day, about 5 MB a year, which is fine to keep in the repo for a long time. If it ever gets large, drop old day files and run snapshot-signal.mjs index.
Environment variables
Copy .env.example to .env and set ANTHROPIC_API_KEY (get one at console.anthropic.com). In Vercel, set it under Project → Settings → Environment Variables.
Deploying
Push to GitHub, import the repo in Vercel, add ANTHROPIC_API_KEY, deploy. Every push after that auto-deploys.
