Musashino Sitter Navi (work in progress)

Demo video: Watch on YouTube

Inspiration

Tokyo has a wonderful subsidy program: parents can get up to ¥2,500/hour toward babysitter costs. But to use it, my city's website offers a list of 37 approved providers — as 37 separate links. Each company's site is different. Fees, sick-child support, cancellation rules: you dig for them one site at a time, usually at midnight after the kids are asleep. I know, because I'm a parent in Musashino City and I gave up halfway through that list.

The program is great. The last mile is broken. So I built the last mile.

What it does

Musashino Sitter Navi puts all 37 government-approved providers on one mobile page:

• Location-aware: one tap on "use my location" (or type any address) sorts providers by distance, geocoded via Japan's official GSI service
• Filters that match real life: no enrollment fee (16 of 37!), sick-child care, night care, same-day booking, support for children with disabilities
• Everything on one card: rates, minimum hours, cancellation policy, tap-to-call, official links — plus a subsidy summary (¥2,500/hr day, ¥3,500/hr night, 144 hrs/year)
• Privacy by design: no account, and entered addresses are never stored
• Radical honesty: the title says "(work in progress)". Date/time filters exist in the UI but are labeled "not implemented" — because provider availability data isn't public, and we refuse to fake it

How I built it

Built solo, by a non-engineer, entirely with Codex + GPT-5.6 (Sol) in VS Code, as my practice run for Global Build Week Tokyo. Zero hand-written code. The pipeline: Codex writes → pushes to GitHub → Cloudflare Pages auto-deploys to my own domain.

The data work was the real project:

1. Collected public government data: Tokyo's approved-provider list, Edogawa Ward's provider comparison survey, Setagaya Ward's on-site inspection records (official addresses), Musashino City's subsidy terms
2. Codex structured it all into JSON with sources and retrieval dates embedded

Challenges I ran into — and the one that matters

The AI hallucinated addresses. Codex embedded office addresses for all 37 providers from its own memory. They looked perfect. When I cross-checked every one against official ward inspection records and Japan's National Tax Agency corporate registry, 4 of 37 were wrong — including one placed in the wrong prefecture. The pattern was revealing: long-established companies were accurate; companies founded in 2024–2025 were wrong half the time. AI memory is strong on old facts and weak on new ones.

Every address in the app is now verified against official sources, and the footer documents exactly which source confirmed what, and when. My biggest lesson: AI output becomes information only after you've traced it to a source. Until then, it's plausible-looking text.

Accomplishments that I'm proud of

• A real civic tool, live at https://sitter.moshimur.dev/, that already saves this parent (me) the 37-link ritual
• 37/37 provider records verified against official government sources
• Honest UI: "(work in progress)" in the title, "not implemented" badges on unfinished inputs — trust over polish

What I learned

Codex makes building easy enough that the hard parts become choosing the right problem and verifying the output. A non-engineer's advantage is knowing the problem from the inside — I'm the user, at midnight, with that list of 37 links.

What's next

The true final form — "find a sitter available at this date and time" — can't be built by one person: availability data lives inside each provider's systems. If a provider, a government office, or another engineer picks up this concept, parents all over Tokyo win. Honestly, if this app becomes unnecessary because someone builds the real thing, that's the happiest possible ending. (The full story is on my blog, in Japanese.)

Built With

• cloudflare-pages
• codex
• css3
• github
• gpt-5.6
• gsi-geocoder
• html5
• javascript
• open-data