YK Studio

Manifest Match is live and free to try - paste a repo URL, answer five questions, get a report.

I built it after watching my own submissions get held up. App store rejections almost never come from bad code. They come from a listing that says one thing while the code quietly does another: location you forgot to declare, an ad SDK nobody mentioned, no account deletion path.

So the tool reads the actual code and reports where the two disagree. Every finding cites the file and line it came from, so you can check it yourself in ten seconds.

The design decision I'd defend hardest: what it cannot verify is reported as "needs checking," never silently passed. A clean report you can't trust is worse than no report at all.

No signup. Just a URL.

Open-sourced the agent that runs my apps' ops overnight.

Three guards plus a weekly report: reviews every 6h, revenue daily, crashes every 6h. In a full run it classified five reviews - three drafted and filed without a word, two escalated to me with the draft attached. Silence is the default.

Two rules shaped it more than any feature:

1. The model judges, the code measures. Anomaly detection is plain unit-tested Python, never a prompt. The agent decides what a deviation means and how to say it, but it never does the arithmetic.

2. Nothing irreversible. It drafts replies but never posts them. It proposes a price change; it never applies one.

The most interesting failure wasn't technical. On the first real run it wrote a beautiful reply to an angry customer and invented a cause for the bug - pointing at a setting that doesn't exist on Android. It read well, which is exactly what made it dangerous. Now a customer-facing draft may only state what the tool data supports; hypotheses go to me, labelled.

Runs fully offline in CI with no cloud credentials, so you can clone it and see it work in one command. Apache-2.0.

Kiroku is live on Google Play, now available worldwide.

Most journaling apps lose me by day four. Writing into a void stops feeling like anything. So I built the smallest possible fix: you write one line about your day, and the app writes one line back.

Not advice. Not a coaching lecture. Just a reply, so it feels like someone read it.

It also scores the mood of each entry and draws the curve over time, which turned out to be the part people keep coming back for - seeing how a month actually went, not how you remember it.

One line a day is the entire commitment. Built solo with Flutter, Firebase and Claude.

Narabu is live on Google Play, in 177 countries.

It predicts how long you'll wait at a popular restaurant and paints the answer on a map: green = walk in, yellow = short wait, red = long line.

The part I care about most is the honesty. A restaurant with no reports still gets a prediction from day one (time, day of week, weather), but the app tells you the confidence is low instead of pretending. Every prediction carries its confidence level.

Built solo on weekends with Flutter and Firebase. The prediction engine is deterministic math, not an LLM - so it is testable and I can explain any number it shows.

Starting with Tokyo's famous ramen shops.