Bilingual health & breakup-recovery phone app with a guarded LLM coach on Cloudflare Workers AI
Bilingual phone app (water, stress, training, breakup no-contact journal) whose LLM button, Llama 3.3 on Cloudflare Workers AI behind a guarded Worker, talks users out of texting their ex. Live: cloudcrane.ai/blueprints/bar-log-6r4rkunmse
Business Challenge
After a breakup, the riskiest moment is the impulse to text an ex: usually late at night, alone, with nobody to talk to. General chatbots need an account, take time to open and keep the conversation. The goal was a one-tap intervention inside a daily health app people already open, that reads the person's own journal (mood, urge to text 1-10, whether they checked the ex's feed), answers in seconds in their language, stores nothing, and is never unsafe for someone in crisis.
Solution Delivered
Front end: a single-file, offline-first PWA on CloudCrane with Chinese/English UI, versioned data migrations (v1 to v5) that keep existing records, encrypted multi-device sync and a passkey (Face ID / fingerprint) lock. AI path: the app sends only the last 14 journal days plus the chosen language to a Cloudflare Worker. The Worker enforces an origin allow-list, a per-IP rate limit (6 per minute) and a request-size cap, and screens entries for self-harm language in English and Chinese; those get crisis-hotline guidance instead of a model call. Otherwise it prompts Llama 3.3 70B (with a fallback model) under a tightly scoped system prompt: blunt but caring, quote the user's own numbers, one concrete action for the next 10 minutes, never insult the person, under 70 words. Language: the journal is rendered in the user's language before prompting, after English labels were found leaking into Chinese replies. Privacy: an in-app switch turns the AI off completely, and then nothing leaves the phone; nothing is stored server-side. Quality: 17 released versions, each checked in CloudCrane's browser test runner, plus local Playwright tests (virtual authenticator, mock sync server); one version was rolled back after a performance regression.
Outcomes Achieved
Live and public. AI replies return in about 2-4 seconds on Workers AI's free tier. The app scores 95/100 on CloudCrane's mid-range-phone quality benchmark, and 33 of its 35 journeys and promises are proven by automated tests. Self-harm language never reaches the model; it is routed to hotline guidance. Chinese replies no longer mix in English terms after the prompt was localised. This is a personal deployment, so no user-count or revenue figures are claimed.
Measurable Business Outcome
AI reply latency ~2-4 s; app quality score 95/100; 33 of 35 automated journey/promise checks proven; 100% of self-harm-language entries routed to hotline guidance instead of the model (by design and tested)