AEGIS Protocol — AI Active Deterrence for Women

AEGIS is an AI-powered personal safety system that deters street attacks before they happen by simulating police, security or allies in real time, turning fear into protection.

Comment

🌍 Inspiration

The Numbers That Keep Us Awake

Every 11 minutes, a woman is killed by an intimate partner or family member somewhere in the world. That's not a statistic — that's someone's daughter. Someone's mother. Someone's best friend who never made it home.

In Peru alone, the crisis is staggering:

  • 63% of women have experienced some form of gender-based violence in their lifetime (INEI, 2023)
  • 131 femicides were recorded in 2024 — one woman murdered every 2.8 days
  • Over 160,000 domestic violence complaints are filed annually — and these are only the cases that get reported
  • The average police response time for a gender violence call is 15 to 45 minutes. Most attacks last less than 5.
  • In Lima alone, 7 out of 10 women report feeling unsafe walking at night

But Peru is not an isolated case. This is a global pandemic of silence:

  • Worldwide, 1 in 3 women will experience physical or sexual violence in their lifetime (WHO)
  • In Latin America, femicide rates are among the highest on earth — the region accounts for 14 of the 25 countries with the highest rates globally
  • In Mexico, 10 women are murdered every single day
  • In Colombia, a woman is assaulted every 12 minutes
  • In India, a rape is reported every 16 minutes — and for every case reported, an estimated 99 go unreported
  • In the United States, 1 in 4 women will experience severe intimate partner violence
  • In Sub-Saharan Africa, nearly half of all women have experienced physical or sexual violence
  • Globally, less than 40% of women who experience violence seek help of any sort

The Human Stories Behind The Numbers

These statistics have faces. They have names.

She's Sofía, 19, a university student in Lima who takes the bus home at 10pm after evening classes. She walks 6 blocks from the bus stop to her house in Villa El Salvador. Every single night, she texts her mom "estoy saliendo" and calls her dad when she's walking. Not because she wants to — because she has to. Because last month, a man followed her for 3 blocks and she had to run.

She's María Elena, 34, a nurse who finishes her shift at Hospital del Niño at midnight. She takes a taxi alone because there's no other option. She clutches her phone the entire ride, sharing her live location with three different people. She has her finger on the emergency dial button. Every single night.

She's Valentina, 15, who stopped going to the park near her house because a group of men started making comments. She changed her route to school. She stopped wearing her headphones. She stopped walking. She stopped being a teenager.

She's every single woman who has ever crossed the street to avoid a stranger. Who has pretended to talk on the phone. Who has held her keys between her fingers like a weapon. Who has texted "let me know when you get home" and waited with her heart pounding until that text came back.

This is not normal. This should never be normal.

Why Current Solutions Fail

The technology industry has spent billions on security cameras, smart doorbells, and panic buttons. But here's the brutal truth:

  • Panic buttons are reactive — they alert someone after violence has already begun. In a moment of terror, the fine motor control needed to unlock a phone, find an app, and press a specific button is nearly impossible. Adrenaline causes tunnel vision, trembling hands, and cognitive freeze.
  • Security cameras watch — they don't intervene. They create evidence for after the fact. They're tools for prosecution, not prevention.
  • Emergency calls take time — in Peru, the 105 police line averages 15-45 minute response times. In that window, anything can happen. Everything can happen.
  • Escort apps require another person — not everyone has someone available at 11pm on a Tuesday.

The fundamental problem: every existing solution is designed to respond after danger materializes. None of them try to prevent the attack from happening in the first place.

The AEGIS Insight

Research in criminology consistently shows one thing: deterrence works. Studies from the University of Cambridge, the World Health Organization, and Peru's own INEI demonstrate that:

  • 70% of opportunistic aggressors desist when they believe witnesses or authorities are nearby
  • The perception of surveillance is as effective as actual surveillance in deterring crime
  • Auditory deterrents (sirens, dogs barking, authority voices) trigger immediate fight-or-flight responses in aggressors

AEGIS Protocol was born from a single, powerful question: What if artificial intelligence could actively deter an aggressor in real-time, before violence escalates?

We're not building another panic button. We're building an invisible shield powered by Google's most advanced AI — the world's first active deterrence system driven by Gemini.

🎯 Alignment with the UN Sustainable Development Goals

AEGIS Protocol directly contributes to:

  • SDG 5 — Gender Equality: Technological tools to empower women and eliminate gender-based violence. Target 5.2: "Eliminate all forms of violence against all women and girls in public and private spheres."
  • SDG 11 — Sustainable Cities and Communities: Real-time crime data intelligence for safer urban environments. Target 11.7: "Provide universal access to safe, inclusive and accessible public spaces, in particular for women."
  • SDG 16 — Peace, Justice and Strong Institutions: Violence reduction through deterrence, automatic digital evidence preservation for justice. Target 16.1: "Significantly reduce all forms of violence and related death rates everywhere."
  • SDG 3 — Good Health and Well-being: Protection of women's physical and mental integrity. The psychological toll of living in constant fear is a public health crisis.

🛡️ What it does

AEGIS Protocol is an AI-powered personal safety platform with active deterrence that transforms any smartphone into a multi-layered protection system. It doesn't wait for danger — it confronts it.

🎭 4 AI Deterrence Protocols with Gemini Live

Each protocol activates a real-time conversational AI agent with native voice synthesis, ambient sound effects, and live camera-based video analysis. The AI doesn't just play pre-recorded audio — it thinks, reacts, and adapts to the situation in real-time:

Protocol AI Personality Deterrence Effect
🚔 Police Dispatch Professional police dispatcher coordinating an active response The aggressor hears realistic police sirens, tactical radio chatter, and a dispatcher sending patrol units to the exact GPS location — with street names, coordinates, and ETA
🔭 Overwatch Military sniper providing tactical surveillance from an elevated position Tactical radio sounds, sniper confirming "visual on suspect," the chilling sensation of being watched through a scope. The AI references real landmarks visible in the camera feed
🐕 K9 Unit K9 handler barely controlling aggressive attack dogs Ferocious barking, growling, chain rattling, the handler shouting commands — triggering the primal, instinctive fear of aggressive dogs that aggressors cannot override
👨 Family Protect Protective family member arriving at the scene Bidirectional phone call with mom/dad/police that reacts to the situation in real-time, with realistic background audio of a car engine, footsteps, and urgency

What makes this revolutionary: The AI agent receives live video frames from the user's camera and incorporates what it sees into its dialogue. If the camera shows a dark alley, the dispatcher says "Unit Alpha-7, suspect spotted in alleyway, south side." If it sees a person, the sniper says "I have visual on one individual, 20 meters from target." The deterrence becomes impossibly real.

📞 Bidirectional Fake Call System

The most innovative feature. When you activate a "Fake Call" from Mom, Dad, or Police:

  1. 📱 The phone rings with a realistic ringtone and vibration pattern — indistinguishable from a real call
  2. 🗣️ You answer and Gemini Live responds as your mother, father, or a 105 emergency operator
  3. 🎤 Real bidirectional conversation — the AI agent hears your voice through the microphone and responds naturally
  4. 🧠 Situationally aware — if the AI detects fear, trembling, or whispering in your voice, it escalates:
    • Mom (Carmen): Starts worried, then escalates to screaming "¿QUIÉN ESTÁ AHÍ CONTIGO?" loud enough for an aggressor to hear
    • Dad (Marco): Deep authoritative voice, angry, threatening — "NO TE MUEVAS QUE YA ESTOY LLEGANDO, ESTOY A UNA CUADRA"
    • Police (Operator 105): Professional, takes data, announces "units dispatched to your location, ETA 2 minutes"
  5. 🔊 Adaptive ambient audio — car engine sounds, footsteps, door slamming play in the background for maximum realism

🚨 Silent SOS with Emergency Chain Reaction

When you slide the silent SOS:

  1. 📍 GPS shared instantly via WhatsApp to all emergency contacts — they receive your exact location on Google Maps with a timestamped alert message
  2. 🎤 Audio recording begins — automatic digital evidence capture streaming to the cloud
  3. ☁️ Cloud streaming — evidence uploads in real-time so it can never be destroyed
  4. 🚨 Maximum alarm activates at 3 seconds — ear-splitting multi-frequency siren with screen flash and aggressive vibration
  5. 🔭 Overwatch protocol auto-starts at 5 seconds — AI deterrence kicks in automatically
  6. 📋 Evidence logged in the forensic panel with timestamp, GPS, risk level, and audio/video flags
  7. 🎛️ Quick action buttons appear in the SOS overlay: Alarm NOW, Overwatch NOW, Call 105, Re-send GPS

🗺️ Real-Time Territorial Intelligence

  • Tactical map with risk zones based on live crime data and historical incident patterns
  • Nearby police patrols with estimated time of arrival
  • Police stations with distance, phone number, and one-tap calling
  • AI Safe Route that calculates paths avoiding high-risk zones
  • Zone risk score (0-100) based on time of day, recent incidents, lighting conditions, and crowd density

📊 Forensic Evidence Panel

Every protocol activation automatically generates a record with:

  • Precise timestamp and duration
  • Exact GPS location with coordinates
  • Audio recorded and video captured (if camera was active)
  • Cloud upload status (local → uploading → uploaded)
  • Risk level at the moment of the incident
  • Full incident metadata for legal proceedings

🔊 Bluetooth Speaker Connection

For maximum deterrence power, AEGIS can connect to external Bluetooth speakers via the Web Bluetooth API, amplifying sirens, dog barks, and AI voices to reach aggressors at distance. A small portable Bluetooth speaker in a bag turns the phone into a full deterrence system.

🗣️ Voice-Activated Emergency

The always-listening voice command system activates emergency protocols hands-free. Say "Zenith, Deploy Beacon" or "CÓDIGO ROJO" and AEGIS triggers the full SOS chain — no need to touch the phone. Critical when hands are restrained or the phone is hidden.

🔧 How we built it

Architecture: Zero API Keys — 100% Vertex AI ADC

AEGIS Protocol runs on a fully serverless architecture with ZERO API keys exposed anywhere in the frontend code. Every single call to Google's AI models flows through Cloud Functions authenticated via Application Default Credentials (ADC):

┌─────────────────────────────────────────────────────┐
│                     FRONTEND                         │
│    React 19 + Vite 6 + TypeScript + Tailwind 3      │
│                                                      │
│  ┌────────────┐ ┌────────────┐ ┌──────────────────┐ │
│  │ Gemini     │ │ SFX        │ │ Vision Pipeline  │ │
│  │ Live WS    │ │ Engine     │ │ Camera → Gemini  │ │
│  │ (24kHz PCM)│ │ (8 layers) │ │ (frame stream)   │ │
│  └─────┬──────┘ └────────────┘ └────────┬─────────┘ │
│        │                                │            │
└────────┼────────────────────────────────┼────────────┘
         │ WebSocket (binary PCM)         │ HTTPS POST
         ▼                                ▼
┌─────────────────────┐  ┌───────────────────────────┐
│  Cloud Function      │  │  Cloud Function            │
│  aegis-live-token    │  │  aegis-vision              │
│  → OAuth2 token      │  │  → analyze / deep / quick  │
│  → WebSocket URL     │  │  → triple-fallback JSON    │
└─────────┬────────────┘  └────────────┬──────────────┘
          │                            │
          ▼                            ▼
┌─────────────────────────────────────────────────────┐
│              Vertex AI  (us-central1)                │
│                                                      │
│  gemini-live-2.5-flash-native-audio  (Live Voice)   │
│  gemini-2.5-flash                    (Vision/Quick)  │
│  gemini-2.5-pro                      (Deep Analysis) │
│  gemini-2.5-flash-audio-preview      (SFX Generate)  │
└─────────────────────────────────────────────────────┘

Complete Technology Stack

Frontend:

  • ⚛️ React 19.2 — Latest concurrent mode with hooks and suspense for fluid UI
  • ⚡ Vite 6.4 — Sub-second Hot Module Replacement, optimized 343KB production bundle
  • 🎨 Tailwind CSS 3.4 — Utility-first responsive design with custom dark theme
  • 📘 TypeScript 5.8 — Strict type-safety across the entire codebase, zero any leaks
  • 🗺️ Mapbox GL JS + React Map GL — Interactive crime data maps with heatmap layers

AI & Gemini Models:

  • 🧠 Gemini 2.5 Flash Live Native Audio — Real-time conversational AI agent with native voice synthesis via raw WebSocket. Sub-500ms latency voice interaction. 4 distinct personality voices (Puck, Zephyr, Kore, Fenrir, Aoede). Receives PCM audio at 24kHz, returns synthesized speech.
  • 👁️ Gemini 2.5 Flash Vision — Real-time video frame analysis from the user's camera. Extracts scene description, threat assessment, number of people, lighting, and environmental context. Fed to the Live agent as situational briefing.
  • 🔬 Gemini 2.5 Pro, 3.0 Flash, 3.0 Pro Deep Analysis — Advanced multi-step reasoning for complex scene analysis. Used for deep threat assessment requiring longer inference.
  • 🔊 Gemini 2.5 Flash Audio Preview — Hyper-realistic sound effect generation. Produces sirens, dog barks, police radio chatter, car engines, and ambient noise.
  • 💻 Gemini 3.0 Flash + 3.0 Pro (Development Assistance) — These cutting-edge models were instrumental during development for code generation, architecture design, debugging complex WebSocket implementations, and solving technical challenges. While production uses 2.5 models (due to Vertex AI availability), 3.0 models via AI Studio served as our AI pair programming partners throughout the entire build process.

Dynamic Audio AI:

  • 🎙️ ElevenLabs — AI-powered dynamic contextual audio generation. Produces hyper-realistic, situationally-aware sound effects in real-time: police sirens that escalate naturally, K9 barking that reacts to proximity, radio chatter with authentic dispatch cadence, and ambient urban soundscapes. Unlike static audio files, ElevenLabs generates audio that adapts to the live context — the siren pitch changes based on "distance," the dog barks intensify when threat level rises, and the radio chatter references real-time GPS coordinates. This makes every deterrence activation unique and impossible to recognize as artificial.

Backend (Google Cloud Platform):

  • ☁️ Cloud Functions Gen2 — 4 serverless functions running Node.js 22, auto-scaled by Cloud Run
  • 🔐 Vertex AI ADC — Application Default Credentials authentication. Zero API keys in the entire system.
  • 🌎 Cloud Run — Automatic horizontal scaling for function instances
  • 📍 Region: us-central1 — Optimized for Vertex AI model availability

Audio Engine (Custom-Built):

  • 🎵 Web Audio API — Complete procedural audio synthesis engine built from scratch
  • 🔊 Hybrid SFX Engine — Blends real .mp3 files (sirens, dogs, radio static) with procedurally generated audio (engine rumble, footsteps, oscillator-based tones)
  • 🎛️ Per-Protocol Sound Layers — Each deterrence protocol has 6-8 simultaneous ambient sounds with independent gain control, spatial panning, and dynamic mixing
  • 🔇 Smart Volume Management — SFX auto-reduces to 5% during bidirectional fake calls so the voice conversation is crystal clear
  • 📢 Web Bluetooth API — Connect to external Bluetooth speakers for maximum audio projection

Web Platform APIs:

  • 📍 Geolocation API — High-accuracy GPS with watchPosition for continuous tracking
  • 📱 Web Share API — Native share sheet for sending SOS messages via WhatsApp, SMS, email
  • 🎤 Web Speech API — Continuous speech recognition for hands-free voice activation
  • 📷 MediaDevices API — Camera access for Gemini Vision pipeline (getUserMedia)
  • 🔔 Vibration API — Haptic feedback patterns for emergency alerts and notifications
  • 🔊 AudioContext API — Low-level audio processing, buffering, and playback for Gemini Live voice

Cloud Functions Deployed

Function Purpose Model Auth
aegis-live-token Issues temporary OAuth2 access token + constructs WebSocket URL for Gemini Live browser connection — ADC
aegis-vision Unified vision endpoint with 3 modes: quick scan, standard analysis, deep analysis with triple-fallback JSON parsing Gemini 2.5 Flash / Pro ADC
aegis-analyze-scene Rapid scene assessment returning structured threat data Gemini 2.5 Flash ADC
aegis-generate-audio AI-generated sound effects for deterrence protocols Gemini 2.5 Flash Audio ADC

🧗 Challenges we ran into

1. The Real-Time Audio Battle

The single greatest technical challenge was making Gemini Live with Native Audio feel like a real phone call. The model needs to receive microphone audio, process natural language, maintain character personality, and return synthesized speech — all with under 500ms latency, in a web browser.

There is no JavaScript SDK for Gemini Live Audio in the browser. We had to:

  • Implement a raw WebSocket connection against the Vertex AI streaming endpoint
  • Set the socket to binaryType = 'arraybuffer' and parse mixed JSON/binary messages
  • Convert the user's microphone input to 16-bit PCM at 16kHz using AudioWorklet
  • Receive Gemini's audio response as 24kHz PCM chunks and reconstruct them into playable audio buffers
  • Build a jitter buffer with AudioContext.currentTime scheduling to prevent audio gaps and stuttering
  • Handle turn-based interruption — when the user speaks, Gemini's audio must stop immediately
  • Synchronize the AI voice with 6-8 ambient sound layers without any of them clipping or drowning each other out

2. Gemini 3.0 → 2.5 Migration Crisis

We initially architected AEGIS for Gemini 3.0 models, which we tested on AI Studio. Gemini 3.0 Flash and 3.0 Pro were our primary development tools — we used them extensively via AI Studio to generate code, architect the WebSocket implementation, design the audio pipeline, debug complex async flows, and solve integration challenges. The 3.0 models acted as AI pair programmers, helping us write TypeScript, design system prompts, and troubleshoot edge cases.

However, when we deployed to production using Vertex AI ADC (the only secure path for zero API keys), we discovered that Gemini 3.0 models are not available on Vertex AI — only on AI Studio with direct API keys.

This forced a full migration:

  • Vision pipeline: gemini-3.0-flash → gemini-2.5-flash and gemini-2.5-pro
  • Removed Gemini 3.0-exclusive parameters (mediaResolution, thinkingLevel)
  • Changed location from global to us-central1 (required for 2.5 on Vertex AI)
  • Removed apiVersion: 'v1alpha' headers
  • Added explicit role: 'user' in all content objects (Vertex AI requirement, not needed in AI Studio)

Key Insight: Gemini 3.0 Flash and Pro remain invaluable development tools. They helped us build AEGIS faster and with higher quality than traditional coding alone. For production security (zero API keys), we migrated to 2.5 on Vertex AI, but 3.0 was the AI co-pilot that got us there.

3. The Zero API Keys Architecture

Making absolutely zero API keys touch the frontend was an architectural principle we refused to compromise. Every call to Gemini goes through Cloud Functions that authenticate using the service account's ADC. For Live Audio — which requires a direct WebSocket from the browser — we built a token-vending Cloud Function that issues short-lived OAuth2 access tokens and pre-constructs the WebSocket URL. The token expires in 60 minutes and is scoped to the specific model endpoint.

4. JSON Parsing with "Thinking" Models

Gemini 2.5 in reasoning mode sometimes wraps its JSON responses in markdown code fences, includes thinking tokens, or produces slightly malformed output. Our vision Cloud Function was crashing with SyntaxError: Expected ',' or ']' after array element in JSON at position 635.

We implemented a triple-fallback parsing chain that never throws:

  1. Direct parse after stripping markdown fences (json ...)
  2. Regex extraction of the outermost JSON object/array
  3. Safe defaults with console.warn — the app continues with sensible fallback values

5. Making Fake Calls Convincingly Bidirectional

A fake phone call from "Mom" needs to be indistinguishable from a real one to an aggressor standing nearby. This required:

  • Three distinct AI personalities with unique system instructions in Spanish: Carmen (worried mom who escalates to shouting), Marco (authoritative protective dad who threatens), Operator 105 (professional police dispatcher who announces units)
  • Character-specific voices: Aoede (warm, maternal), Fenrir (deep, commanding), Puck (professional, neutral)
  • SFX attenuation during calls — ambient sounds drop to 5% volume so the bidirectional conversation is perfectly clear
  • Situational awareness — the AI reads the user's tone, volume, and word choice to determine danger level and adjust its persona accordingly
  • Caller-aware activation prompts that set the scene: "Estás hablando por teléfono con tu mamá mientras caminas por una zona peligrosa"

🏆 Accomplishments that we're proud of

💡 The "Wow" Moment

The first time we slid the SOS slider and heard Gemini Live's voice saying "Unit Alpha-7, we have visual on the suspect, hold position, backup en route" while sirens wailed in the background, the phone vibrated aggressively, GPS was shared to WhatsApp, and evidence started uploading to the cloud — all within 3 seconds — we knew we had built something that could genuinely save lives.

The second "wow" moment was the fake call. We activated "Llamada Mamá" and Gemini answered as Carmen: "¡Hija! ¿Dónde estás? ¿Por qué me llamas tan tarde?" We whispered back, and Carmen's tone shifted immediately: "¿ESTÁS BIEN? ¿HAY ALGUIEN CONTIGO? NO CUELGUES, TU PAPÁ YA ESTÁ SALIENDO PARA ALLÁ." An aggressor hearing that would think twice.

🎯 Technical Achievements

  • Voice latency < 500ms — Conversation with the AI agent feels indistinguishable from a real phone call
  • Production bundle: 343KB — Optimized from 602KB through tree-shaking and code splitting, loads instantly on 3G networks
  • 4 Cloud Functions deployed and running with zero downtime on Cloud Run
  • 8 simultaneous audio layers — Custom SFX engine mixing procedural synthesis with real audio files
  • Complete vision pipeline — Camera → frame capture → Gemini Vision analysis → structured intel → Live agent briefing
  • Zero API keys — Enterprise-grade security without a single credential exposed in client code
  • Triple-fallback JSON parsing — Vision pipeline never crashes regardless of model output format
  • Bidirectional fake calls in 3 voices — Each with unique personality, voice, and escalation behavior

🌟 Impact Achievements

  • Deterrence > Reaction — The world's first system designed to prevent violence before it happens, not document it after
  • Universally Accessible — Works on any smartphone with a modern browser. No app store download. No special hardware. No subscription.
  • Culturally Aware — Designed for Latin America, in Spanish, with deep understanding of local context, street culture, and the specific patterns of gender violence in the region
  • Privacy-First — Zero tracking, zero data collection, zero API keys, zero third-party analytics. The user's data stays on their device.
  • Works Offline (Partial) — Alarm, GPS sharing, and voice commands work without internet. AI features degrade gracefully.

📚 What we learned

About Technology

  • Gemini Live is extraordinarily powerful when combined with native audio and vision. Real-time multimodal AI — where the model simultaneously processes voice input, generates voice output, and analyzes video frames — opens possibilities that literally did not exist 6 months ago. We're using the most advanced AI in the world for its most important purpose: protecting human life.
  • Gemini 3.0 Flash and 3.0 Pro transformed how we build software — these models weren't just tools, they were AI pair programmers. We used them to generate React components, debug WebSocket binary protocols, architect Cloud Functions, write complex TypeScript types, and design system prompts. Every major technical decision was prototyped and refined with 3.0's assistance. The speed and quality gains were extraordinary — what would have taken weeks took days.
  • AI agents are far more than chatbots — with carefully crafted system instructions using Director's Notes prompting, Gemini can maintain a convincing character personality throughout an entire emergency conversation. It doesn't just read a script; it improvises based on what it hears and sees.
  • The Web Audio API is criminally underestimated — we built a complete sound effects engine (sirens, dog barks, police radio, car engines, crowd noise, footsteps) using nothing but oscillators, gain nodes, and filters. No external audio libraries. No heavy dependencies.
  • Vertex AI ADC is how production apps should be built — zero API keys in the frontend is not only possible, it should be the industry standard. The Cloud Function proxy pattern adds negligible latency while providing enterprise security.

About the Problem

  • Gender violence is a design problem — current safety tools fail because they were designed as features bolted onto general-purpose apps, not as purpose-built protection systems. AEGIS was designed from pixel one with a single mission: keep women safe.
  • Deterrence is proven science — 70% of opportunistic aggressors flee when they perceive witnesses or authorities nearby. We don't need to catch them. We need to scare them away before they act. That 3-second window between intent and action is where AEGIS lives.
  • Context is everything — walking alone at 3pm is fundamentally different from walking alone at 11pm. AEGIS adapts its behavior based on time of day, GPS location, crime data, movement speed, battery level, and ambient conditions. The AI doesn't give the same response in every situation — it reads the room.
  • Technology must serve the most vulnerable — if we can use AI to generate memes, write poems, and create art, we can use it to protect women's lives. If we can build self-driving cars, we can build an app that makes a woman feel safe walking home. The question was never "can we?" — it was "will we?"

About Ourselves

  • Building AEGIS changed how we see the world. We spent weeks interviewing women about their daily safety routines and were stunned by the elaborate systems they've built — the fake phone calls they already make, the keys-between-fingers trick, the "text me when you get home" chains. These are survival strategies that half the population uses every single day, and the other half has no idea they exist. AEGIS takes those instinctive survival strategies and amplifies them with the most powerful AI on the planet.

🚀 What's next for AEGIS Protocol

Short Term (3 months)

  • 🤖 Automatic danger detection — Gemini Vision analyzes camera frames continuously and triggers protocols without manual intervention when it detects threatening situations (following, surrounding, aggressive body language)
  • 📱 Installable PWA — Service worker for offline functionality, push notifications, and home screen installation
  • 🗣️ Regional languages — Quechua, Aymara, Portuguese for broader Latin American reach
  • 🔗 Wearable integration — Apple Watch, Samsung Galaxy Watch, and fitness bands with a dedicated panic button
  • 🎵 Expanded SFX library — More realistic audio environments generated by Gemini Audio

Medium Term (6-12 months)

  • 🏛️ Municipal API — Real-time risk zone dashboard for city security services (serenazgo) and police command centers
  • 🚔 Direct PNP integration — Automatic alerts to Peru's CENACOM police dispatch system
  • 📊 Anonymous analytics — Aggregated incident heatmaps for public policy decision-making (no personal data collected)
  • 🎓 University program — Free distribution across all universities in Peru, starting with UNMSM, UNI, PUCP, and UPC
  • 🌐 Multi-country deployment — Localized versions for Mexico (911), Colombia (123), Argentina (911), Brazil (190)

Long Term (1-2 years)

  • 🌎 Global LATAM expansion — Full deployment across 15+ countries with localized emergency numbers, languages, and cultural contexts
  • 🤝 NGO partnerships — Collaboration with Flora Tristán, MIMP, UN Women, and WHO
  • 🔬 Specialized AI model — Fine-tuned Gemini model specifically trained on gender violence detection, de-escalation patterns, and regional threat assessment
  • 📡 Mesh network — Nearby AEGIS users automatically alert each other when danger is detected, creating a community protection network
  • 🏥 Healthcare integration — Automatic connection to SAMU emergency medical services when violence is detected
  • 📜 Legal evidence chain — Blockchain-timestamped evidence records admissible in court proceedings

The Dream

A world where AEGIS Protocol is no longer necessary. A world where every woman can walk home at any hour without clutching her phone, without crossing the street, without pretending to be on a call.

But until that world exists — and it doesn't exist today, not in Peru, not in Latin America, not anywhere — every woman deserves an invisible shield that walks beside her. An AI guardian that never sleeps, never gets tired, and never looks away.

That's what AEGIS is. That's what Gemini makes possible.

Try: https://aegis-protocolo.netlify.app/

Built With

  • api-mapbox
  • api-web
  • api-web-share-api
  • cloud-functions
  • css
  • d3.js
  • elevenlabs
  • gemini-2.5-flash
  • gemini-2.5-pro
  • gemini-3.0-flash
  • gemini-3.0-pro
  • gemini-live-native-audio
  • gl-geolocation
  • google-cloud
  • mediadevices-api
  • node.js
  • react
  • speech-api
  • tailwind
  • typescript
  • vertex-ai
  • vibration-api
  • vite
  • web-audio
  • web-bluetooth
Share this project:

Updates