KIT AI

KIT AI is an offline-first emergency health assistant for situations where internet access is unreliable or unavailable, like airplanes, hiking trails, or low-connectivity regions. It provides objective, widely accepted first-aid information using a locally running AI model, without requiring network access at the time of use.

Live demo: https://kit-ai-smoky.vercel.app

Disclaimer: KIT AI is not a doctor, does not diagnose, and does not replace professional medical care. It shares general first-aid guidance only. In emergencies, seek professional help whenever possible.

---

Why KIT AI Exists

Most AI health tools require constant internet access. In real emergencies, that's often impossible.

KIT AI is built to:

• Work fully offline
• Run directly on the user's device
• Share clear, factual first-aid steps
• Avoid hallucinations, diagnoses, or personalized medical advice
• Update medical content safely when internet is available

---

What KIT AI Does

• Uses a local LLM (WebLLM) running entirely on-device via WebGPU
• Stores vetted first-aid knowledge in IndexedDB (syncs from backend when online)
• Provides a chat-style interface for health questions
• Never prescribes medication or gives diagnoses
• Syncs updated medical guidelines only when online

---

What KIT AI Does NOT Do

• No user accounts or logins
• No real-time cloud AI calls during inference
• No diagnosis or treatment plans
• No medication recommendations or dosages
• No replacement for emergency services

---

Architecture

flowchart LR
    subgraph Browser
        UI[React PWA]
        Worker[Web Worker: WebLLM]
        IDB[(IndexedDB\nmodel cache + medical cache)]
        UI -->|chat messages| Worker
        Worker -->|model weights, cached content| IDB
        IDB --> UI
    end

    subgraph Backend[Backend, online only]
        API[Node/Express API]
        Gemini[Gemini content pipeline]
        Mongo[(MongoDB Atlas)]
        TTS[ElevenLabs TTS proxy]
        Gemini --> Mongo
        API --> Mongo
    end

    UI -.->|GET/POST /api/medical, when online| API
    UI -.->|POST /api/tts/convert, when online| TTS

Loading

The frontend is a PWA that runs the model and the medical knowledge cache entirely in the browser. The backend is optional infrastructure for content generation and TTS. It is not required for the core offline chat experience to work.

Frontend (PWA)

• React + Vite
• Progressive Web App with service worker
• IndexedDB for:
  • LLM model weights (WebLLM cache)
  • Medical knowledge (synced from backend when online)

Local AI

• @mlc-ai/web-llm in a Web Worker
• WebGPU / WASM inference
• Model: Llama-3.2-1B-Instruct by default (swapped from 3B for faster load time on Vercel). A larger model can be configured via VITE_WEBLLM_MODEL_URL.
• A separate fine-tuned medical model, Pablo305/llama3-medical-3b-4bit, exists but is not yet wired into the PWA (see Known Limitations).

Backend (Online Only)

• Node.js + Express
• Gemini-based pipeline that generates first-aid guideline content
• MongoDB Atlas stores that content
• GET /api/medical — frontend fetches and caches when online
• POST /api/medical — ingest script updates content, gated behind MEDICAL_API_KEY
• ElevenLabs TTS proxy (POST /api/tts/convert) for higher-quality voice, with the browser's built-in speech synthesis as a fallback

---

Tech Stack

Layer	Technologies
Frontend	React, Vite, Tailwind CSS, PWA
Local AI	@mlc-ai/web-llm, WebGPU
Backend	Node.js, Express, MongoDB, Gemini API
Voice	ElevenLabs (online), browser SpeechSynthesis (offline fallback)
Storage	IndexedDB (frontend), MongoDB Atlas (backend)

---

Getting Started

Prerequisites

• Node.js 18+
• MongoDB Atlas account (or local MongoDB), only needed if you're running the backend
• Browser with WebGPU (Chrome 113+, Edge 113+, Safari 26+, Firefox 141+)

Backend

cd backend
cp .env.example .env
# Edit .env with your MONGODB_URI, PORT, FRONTEND_ORIGIN, MEDICAL_API_KEY
npm install
npm run dev

Ingest initial medical content (with backend running):

npm run ingest
# Or: npm run ingest path/to/medical-knowledge.json

Frontend

cd frontend
cp .env.example .env
# For API mode: set VITE_MEDICAL_SOURCE=api and VITE_MEDICAL_API_URL
npm install
npm run dev

Visit http://localhost:5173. On first load (while online), the model downloads and medical content syncs. After that, the app works offline.

---

Configuration

Frontend (.env)

Variable	Description
VITE_MEDICAL_SOURCE	api or static
VITE_MEDICAL_API_URL	Backend URL, e.g. http://localhost:3001/api/medical
VITE_WEBLLM_MODEL_URL	Optional. Points at a larger or custom model instead of the default 1B model

Backend (.env)

Variable	Description
MONGODB_URI	MongoDB Atlas connection string
PORT	Server port (default 3001)
FRONTEND_ORIGIN	CORS origin (default http://localhost:5173)
MEDICAL_API_KEY	Required to call POST /api/medical. If unset, that endpoint is disabled rather than open.

---

Medical Content Format

{
  "version": 1,
  "entries": [
    { "id": "topic-id", "content": "Medical content..." }
  ]
}

Use npm run ingest in the backend to push this format to MongoDB.

---

Known Limitations

• WebGPU is a hard requirement, with no fallback yet. If the browser or device doesn't support WebGPU, the app currently shows an error instead of degrading to a slower CPU path or a hosted model.
• The backend and TTS pipeline aren't connected in the live deploy. The Vercel demo runs frontend-only, against the static bundled medical content, with browser speech synthesis for voice. The Express/MongoDB/Gemini/ElevenLabs backend exists and works locally, but isn't deployed alongside it.
• The fine-tuned medical model isn't integrated into the PWA yet. Pablo305/llama3-medical-3b-4bit is trained and published, but the app still ships with the general-purpose Llama-3.2-1B-Instruct model.

---

Example Use Cases

• In-flight medical situations
• Hiking or camping emergencies
• Regions with limited internet access
• Disaster or outage scenarios

---

Ethics & Safety

KIT AI is intentionally conservative:

• When uncertain, it escalates to "seek professional help"
• It avoids personalized or speculative advice
• It prioritizes clarity, calmness, and safety

---

Team

KIT AI was built by a 5-person team at a hackathon. My (Pablo's) parts:

• The IndexedDB retrieval and medical-cache layer that lets the app work fully offline after first load
• The online/offline TTS fallback (ElevenLabs when the backend is reachable, browser speech synthesis when it isn't)
• Vercel deploy fixes to get the PWA actually loading in production
• A fine-tuned medical model, Pablo305/llama3-medical-3b-4bit, fine-tuned during the hackathon and not yet wired into the app

---

Project Status

Active development. Current features:

• WebLLM-based offline chat
• Medical knowledge cache with backend sync
• PWA with IndexedDB caching