I build interactive web experiences with AI.
Fig. 01-A · Concept product landing page
Interactive marketing site for a concept headphone brand. Scroll-driven 3D product animation, drag-to-inspect feature cards, animated spec sheet. Rev 2 recuts it as a measurement instrument: draggable frequency plots, a soundstage isolation demo, live dB readouts.
Fig. 01-B · Concept product site
Museum-hall site for a concept AI that operates your computer. Five marble exhibits you can drag to turn, rendered live in 3D, a hero relief that assembles itself from scattered fragments, and scroll-driven gallery reveals throughout.
Fig. 01-C · Concept brand site
Five-page site for a concept natural-diamond house, where every page is a trade process you operate. Stir a live GPU fluid sim to polish the rough crystal, scroll to fold a parcel paper over it, lift stones with brass tweezers, and grind facets on a scaife at the real pavilion angle.
Fig. 01-D · Concept brand site
Five-page site for a concept luxury padel clothing house, drawn up as a Virtual Enterprise brand. Four garments hang on a steel rail as live cloth simulations — push through the rack and the hangers knock and swing. Click one and it lifts off the rail until the fabric fills the screen, then unravels, unzips or whips away into its own page.
Fig. 01-R · Original research
An original study of whether AI models inherit human money biases. Ten open models took seven classic behavioral-finance experiments, 31,500 decisions run overnight on one GPU. They beat people on sunk costs, lost to a random number, and took a losing bet the moment it was called a trade. The site lets you take the same tests, flip a physics-simulated coin, and find out which model decides most like you. The paper is included.
The pages above are the visible half. These are the systems underneath: a full-stack product with a green test suite, a local agent that runs my machine, and the skills that make a model do a job the same way twice. None of these are marketing pages, so there's nothing to demo. The counts do the work instead, and Stir and Jarvis are both on GitHub if you'd rather just read the source.
Fig. 01-S-A · Full-stack SaaS
Discontinued by choiceAI restaurant reputation platform. It synced Google reviews, learned an owner's voice from their own writing samples, and drafted replies they could approve and post in one click. Multi-tenant from the schema up: Postgres row-level security, encrypted OAuth tokens, sliding-window rate limits, and a full WCAG contrast and screen-reader pass.
All 188 tests across 20 test files pass.
Restaurants drown in reviews and answer almost none of them. Stir's whole differentiator was the AI: reply drafts in the owner's own voice, plus insights pulled out of months of feedback they'd never read.
The incumbent review platforms shipped the same feature. Once AI drafting was a checkbox inside the tools restaurants already paid for, my one advantage was gone, and they had the distribution and the integrations I didn't.
I stopped. Not because the build failed. The codebase was finished, hardened, and green: tenant isolation, encrypted tokens, 188 passing tests. Competing head-on with a feature I no longer owned would have been sunk cost wearing the costume of persistence.
Auth, multi-tenancy, rate limiting, migrations, and what it actually takes to make something safe for other people's data. Killing this was cheaper than learning that lesson later, on something bigger.
Verdict: knowing when to stop is a skill. It just doesn't come with a demo — only the code, and the reason.
Fig. 01-S-B · Local agent harness
Source openA Jarvis-style console for my own machine, built on the Claude Agent SDK. Push-to-talk speech in through local Whisper, a local neural voice out, webcam vision, long-term memory in local embeddings, reminders that fire native notifications after a restart, and a global hotkey that summons it from anywhere. Everything that can run offline does, so the voice, the memory, and reading the screen all cost zero tokens.
Allow by default. Stop only for what can't be undone.
Fig. 01-S-C · Authored Claude skills
In useA skill is how you get a model to do a job the same way twice. /cms is the big one: a playbook plus twelve production templates that retrofit a password-gated, self-serve CMS onto an already-deployed Next.js site, so a non-technical owner edits their own copy, photos, and hours without touching code or calling me. It detects the real stack instead of assuming one, and refuses to improvise if the project isn't shaped the way it expects. Three more ship inside Jarvis and shape how it briefs me, closes out the day, and works the web.
Five months, from static pages to systems. Nothing here is hidden because the growth is the point.
Hand-written HTML and CSS, no framework and no build step. Learning what a layout actually is, and how far taste gets you before tooling does.
I'm Daniel Yerushalmi, a student at Syosset High School on Long Island. I got into building sites the moment AI made it possible to move at the speed of taste: I decide how something should feel, then I direct Claude until the pixels agree.
What I care about is the craft layer. Type, motion, and the small details most people only notice when they're missing. The tools are new. The standards aren't.
I direct AI to design,
build, and ship.
Landing pages. Brand sites. Research agents.
The model writes fast. I make it right: the taste, the direction, and the call on when it's actually good enough to ship.
Every project here is deployed and real. AI-assisted building isn't the caveat. It's the skill.
Define the feel before the features.
Prompts, references, taste. The machine gets a creative director.
Claude writes fast. Versions are cheap.
The orange pen comes out. Cut, tighten, polish.
Deployed on Vercel. Real links only.
A website is a storefront, and a storefront is a business decision. Alongside the building, I've taken college-level marketing, accounting, corporate finance, and Virtual Enterprise through LIU, on top of the business track at Syosset High School. The pixels are half the craft. This is the other half.
Fig. 05-A · College credit · LIU
Marketing · in the syllabus
Positioning, segmentation, and the four Ps. Why a product earns attention before it earns a sale.
Applied: every build here starts with how it should feel.
Filed · LIUFig. 05-B · College credit · LIU
Accounting · in the syllabus
Debits, credits, and the financial statements under every business. Reading a balance sheet closely enough that the numbers actually balance.
Applied: a project is only real once it's accounted for.
Filed · LIUFig. 05-C · College credit · LIU
Finance · in the syllabus
Time value of money, valuation, and cash flow. Deciding what a thing is worth, and whether it's worth building at all.
Applied: taste says what to make; finance says why.
Filed · LIUFig. 05-D · College credit · LIU
Virtual Enterprise · in the syllabus
Running a simulated company end to end, from org chart to profit and loss. Business treated as a system you operate, not just one you study.
Applied: I ship whole things, not just screens.
Filed · LIUFig. 05-E · Syosset business track
Syosset · elective
Turning an idea into a business plan and a pitch worth backing.
Syosset · elective
Weighing risk against return before committing a dollar.
Syosset · elective
Budgeting, credit, and saving. The fundamentals, applied to a real life.
One slot on this desk is reserved. The next build is in concept: feel first, then the redlines. Watch it land here, or on LinkedIn.