0
NuraSafe

Project Details

Owner

Humza Sami Chughtai

Release Date

2025 – Present

Services

On-Device AI, iOS, RAG

Duration

Ongoing

Overview

NuraSafe is a privacy-first iOS app I built — private AI that runs entirely on your device, with zero cloud dependency for core use. Free forever. No internet required. Switch into one of 12 emergency modes and get expert-grounded answers when the network is gone.

  • 12 emergency modes: first aid, fire, flood, earthquake, radiation, conflict, chemical hazard, tsunami, wildfire, blizzard, shelter, power outage
  • 100% free — no subscription, no freemium wall
  • 0 bytes uploaded — conversations never leave your phone
  • App Store: apps.apple.com/app/id6761926910 · Website: nurasafe.io

Objective

Safety and sovereignty without a subscription — assistive AI that works offline, stays understandable in emergencies, and costs nothing.

  • Resilient when networks fail first in disasters
  • Modes that match the moment — not generic chat in a crisis
  • Frontier-grade grounding without sending data to the cloud

Process

Emergency-Mode Retrieval

When an emergency mode is active, NuraSafe switches from general chat memory to a dedicated pipeline — the bundled knowledge base is filtered to your scenario (fire, flood, radiation, etc.) plus general safety content, then hybrid search finds the best passages to inject into the prompt. Responses stay as grounded as top cloud assistants using curated safety retrieval — except nothing leaves your phone.

Intelligent Retrieval (RAG)

A lightweight on-device pass reads your message and recent conversation and decides whether NuraSafe's safety knowledge is needed or it's general chit-chat. If knowledge is needed, it produces a tuned search phrase; if that phrase drifts from your question, the system falls back to your wording.

Dual Search — Lexical + Semantic

Retrieval runs BM25 lexical scoring and E5 semantic similarity (384-d embeddings) in parallel. Vectors load into memory from the bundled KB with on-disk caching after first setup. Weak semantic matches are dropped; when embeddings are unreliable, NuraSafe leans lexical-only; when healthy, results rank 60% semantic / 40% lexical before the main model writes the answer — fully offline.

Tools & Stack

On-device LLM for chat and reasoning. Core ML runs multilingual E5 embeddings. Structured knowledge ships with the app and indexes locally. In-memory vector store for low-latency retrieval. Everyday chat often skips retrieval; in a crisis, retrieval turns on when your question needs domain-sharp answers.

Impact

Offline-first AI built for real life and real emergencies — by axon86.

  • 100% free forever — full capabilities, no paywall
  • 0 bytes of user data uploaded
  • 12 specialized emergency modes on one app
  • Hybrid BM25 + semantic retrieval — frontier-grade grounding, on-device
  • Works on everyday iPhone hardware — instant answers, no network latency
  • Live on the App Store: https://apps.apple.com/app/id6761926910