Mobile · Offline AI 2026

SOS Medfield

An offline emergency app that selects prepared first response medical instructions that works with hands-free use.

Measured highlights

100% offline guidance: including local speech and classification
Runs on an iPhone 12: without a network connection

Context

The idea came from encountering a fentanyl overdose on a hike, where poor cell service and not knowing what to do made the first few minutes harder.

SOS Medfield listens to someone describe an emergency, selects a prepared local first-aid protocol, and reads the steps aloud without needing an internet connection.

How a request becomes guidance

Listen
Local speech-to-text captures what the person says without sending it to a remote service.
Choose
Gemma classifies the situation and routes the request to a matching protocol. It does not write the medical instructions.
Guide
The app reads short steps from structured JSON and uses local text-to-speech so the phone can stay out of the user's hands.

From request to guidance

1 / 3

SOS Medfield landing screen asking what happened — The starting screen accepts an emergency description.

A follow-up question asking whether the wound is on an arm or leg — Short follow-up choices narrow the request.

Prepared guidance screen with next-step and speech controls — The guide presents the next prepared step with speech controls nearby.

My role

Responsibility: Frontend and speech mode
Team: Team of 3
Tools: React Native · Gemma 4 · Speech · Offline-first

I worked on the React Native frontend and built the hands-free speech mode. That included shaping the spoken flow so someone could hear the next instruction without keeping a phone in their hands.

Process & decisions

We stopped building a general chatbot.

A general chatbot could forget important context, produce long answers, or invent instructions. A text-only flow also expected someone in an emergency to stop and type.

Keep the instructions outside the model: We put the medical guidance in prepared JSON protocols. The model only interprets the request and chooses which protocol to open.
Carry the important context forward: New conversation details are folded into a running summary so the small local context window keeps the information that matters.
Make the flow work without typing: Hands-free speech lets the user talk to the app and hear the next step while attending to the person who needs help.

Speech and retained context

1 / 2

Hands-free mode with a prominent microphone control — Hands-free mode keeps speech at the center of the interaction.

Patient session summary screen — A session summary carries the conversation context forward.

Outcome

The finished path runs offline and gives one prepared step at a time. It covers situations including CPR, overdose response, choking, severe bleeding, allergic reactions, and hypothermia.

Before: Open-ended chatbot
After: Local protocols with spoken steps

01Context