Skip to content

Tell Liora

An offline iPhone companion that helps Filipino mothers decide when to go to the hospital.

Category
AI
Year
2026
Liora chat screen greeting "Kumusta? Tell Liora how you feel." with a note that the rules decide, three Taglish starter prompts, and a message bar with microphone and voice buttons above an "AI can make mistakes" notice.
Tell Liora Today screen on period day 1, with a two week date strip marking the period days, quick log buttons for period, symptoms, mood and more, a Tell Liora message bar, and a logged today card above the tab bar.
Get Liora ready for offline setup step with a capiz window illustration and three on-device models marked ready: Gemma 4 at 2.8 GB, a 557 MB voice add-on and a 278 MB source card search, above a Start using Liora button.

Overview

Tell Liora is built for the hard moment at home: late at night, a bad headache, no signal, and the health centre closed. A mother tells Liora how she feels in Tagalog, English or Taglish, by typing or by voice. Gemma 4 E2B runs on the iPhone through llama.rn and only reads her words. It answers fixed yes or no questions, each with a confidence. A decision model built from WHO and DOH warning-sign tables then makes the call: go to the hospital now, one follow-up question first, or a calm answer that quotes its source word for word. The AI never decides and never writes medical text. Caution only goes up, so any danger sign counts and a skipped question counts as serious. Liora also keeps a period, symptom and mood log, with a cycle calendar and an explained next-period estimate. After one model download on Wi-Fi, it works in airplane mode, with no account, no server and no analytics. Liora does not diagnose; it helps her decide when to go. A team of four built it at the AppBuildersPH Hackathon 2026, under the theme Local AI.

What it does

  • Gemma 4 E2B runs on the iPhone GPU through llama.rn, so it works in airplane mode with no account or server
  • Rules built from WHO and DOH tables make every decision; the AI only reads her words
  • A "go to the hospital now" card with WHO's fixed words and one tap to call or text her emergency contact
  • A chat agent that logs periods, flow, symptoms and moods from Taglish messages, with Undo
  • Liora Live: hands-free voice, transcribed on the phone by Gemma 4's audio encoder
  • A cycle calendar with logged period days and an explained next-period estimate

The challenge

The first build ran in Safari with Transformers.js, but one tab could not hold Gemma 4 E2B. iPhone WebGPU caps a buffer at 1,024 MB, and the model's token embedding table is 1,120 MB. The team moved to a native iPhone app that reads the model from storage instead of copying it into memory.

Architecture

  1. Screens

    • Expo Router
    • Native iPhone tabs
    • Liora chat
    • Liora Live voice
  2. App state

    • Zustand stores
    • Storage on the phone
  3. On-device AI

    • llama.rn
    • Gemma 4 E2B
    • Gemma 4 audio encoder
    • EmbeddingGemma 300M
  4. Content

    • Fixed copy
    • Reviewed WHO and DOH source cards
  5. Core rules

    • WHO and DOH decision model
    • Triage
    • Cycle maths
    • Agent policy

Built with, 12

  • TypeScript
  • React Native
  • Expo
  • Expo Router
  • NativeWind
  • Reanimated
  • Zustand
  • Zod
  • llama.rn
  • Gemma 4
  • EmbeddingGemma
  • Vitest

Screens, 3

  • Setup step 2 of 4 asking "Where are you now?" with Pregnant selected above Recently gave birth and Neither, and an optional weeks pregnant field.
  • Log your day sheet with a week of dates, period flow options from none to heavy, and symptom chips including cramps, headache, back pain, bloating, nausea and pelvic pain.
  • Calendar in dark mode with October period days as filled circles, the estimated November period as dashed circles, and the estimated fertile window and likely ovulation days in green.