Hotari

Background

Plenty of people put in the work every day and still can’t see themselves getting better. Hotari turns what you did into skills, attributes and a short AI-written note — a two-minute check-in each night. No streaks, no guilt.

What I did

Everything, solo: product, design, engineering, testing and release. iOS and web share one Expo codebase; the backend is Hono on PostgreSQL; subscriptions run through RevenueCat. The global and mainland China versions are listed separately.

Problems and solutions

• Subscription entitlements: webhooks and the reconciliation endpoint share one sync function, with RevenueCat as the single source of truth. Purchases that didn’t carry over after switching accounts were fixed by re-querying on transfer events.
• “The model judges, the code computes”: the AI only picks a tier; code computes the points. A typical day went from +60 to +12, so the numbers stay meaningful.
• One codebase, two regions: separate bundle IDs and servers; the China build carries ICP filing, PIPL-compliant legal text and AI-generated-content labels.
• Four languages: every copy key is covered by a parity test, and the AI personas are rewritten per language (Japanese keeps its politeness levels).
• Performance: found that DeepSeek’s thinking mode was silently on in the Node SDK (about 20× slower), and cut a 42.2 MB font download from the web first load.