Tested at scale.
Bugs found before users did.
A test harness that runs Even Up’s real code found weighted splits wrong in 39,948 of 56,000 cases. A $5.10 tax split three ways paid out $0.03.1
39,948→0
Avi Kaul
Builds apps. Ships them.
High school senior in Fremont, California.
Chrome extension
HumDocs
Pick a finish time. It handles the rest.
Types your text into Google Docs and Slides, on schedule.
iPhone app
Even Up
Scan the receipt. Split it fairly.
Settle up with one person at a time. Released October 9, 2026.2
Rejected. Then released.
App Review, round two.
Apple turned down Even Up 1.0 for app quality. I appealed, fixed what a clean install exposed, and shipped build 39.
Rejected›Appealed›Released
Private by design.
Your data stays yours.
HumDocs keeps your document in the browser. Even Up keeps your money records on your phone. Scans ask before a photo leaves it.
Shipped without sync.
On purpose.
I built shared balances over iCloud, then left them switched off in 1.0. I hadn’t tested them on two real phones, so they didn’t ship.
Checked, not trusted.
How I use AI tools.
I build with Claude Code and review its work like a teammate’s pull request. When an AI-written audit listed seven Even Up bugs, I checked each against the real code. Almost all came from stale copies of old files, so the audit was thrown out.
One price. Paid once.
The business side.
HumDocs Pro is $20, once. People use it in bursts, so a subscription would mostly be cancelled after a month. Payments run through Stripe, and the fixes that make runs correct always stay free.
$20
About.
I like the part after the prototype works. The edge cases, the review rules, the bug reports, and the release that has to go out again.
School
Mission San Jose High SchoolClass of 2027 · Fremont, California
National Merit SemifinalistAP Scholar with Distinction. 5 on AP Computer Science A and Principles.
Also
UC San Diego COSMOS, 2026Cluster 7, Machine Learning: “Can We Teach a Computer to Think?” Team final project in Python: a Naive Bayes classifier that tells human-written from AI-generated text.
Independent research“Likes, Comments, and Shares: How Interactive Features Guide Children’s Moral Development and Understandings of Justice.” A literature review, mentored by a Vanderbilt University PhD student in developmental psychology.
How I build
Solo, with AI coding tools.I use Claude Code to write much of the code. I review every change, test it on real devices and in real Google Docs, and decide what ships.