Smart Industrial Alert (SIA)

Enterprise industrial alert & operations platform

Smart Industrial Alert (SIA) is an enterprise industrial alert and operations platform for factory environments. It delivers real-time mobile alerting with AI-assisted supervisor assignment, predictive-maintenance forecasting, and shift and collaboration coordination - administered from a SuperAdmin console and deployed as a dedicated, white-label instance per customer on the customer's own Firebase and Cloudflare accounts. SIA is integration-ready for existing automation estates (microcontroller/IoT, PLC, SCADA, historian, MQTT, OPC UA, Modbus, and REST/webhook) and ships with operational monitoring, release documentation, and a trust center. It runs on top of the plant - coordinating alerts and people - without taking over control or safety functions.

What SIA does

• Real-time mobile alerting - targeted push with full-screen lock-screen alerts and a durable one-minute fallback so the right supervisors are reached.
• AI-assisted assignment - multi-factor supervisor matching with reinforcement learning; every decision carries an explicit reason and confidence.
• Predictive maintenance - an on-device forecaster that grades and adapts itself, surfacing next-24h machine risk on the Production Manager dashboard.
• SuperAdmin configuration - manage AI agents, branding/theme, AI models, industrial connectors, identity, and monitoring from one console.
• Dedicated customer instances - white-label deployment per customer with automated provisioning; the customer owns their project, keys, and data.
• Shifts, presence, collaboration & escalation - coordinate who responds, when, and how alerts escalate.
• Bilingual - full English/French with instant runtime switching.

Integration modes

SIA sits alongside your existing automation estate and ingests events through cloud-pull and edge-push connectors:

Connectors are configured self-service in the SuperAdmin console with a live link-verification handshake.

Safety & scope

SIA is a supervisory alert and operations layer. It is not a PLC, a safety-instrumented system (SIS), or a real-time control loop, and it does not actuate equipment or perform emergency shutdown. Existing SCADA/PLC control and safety functions remain authoritative; SIA augments them with alerting, coordination, and operational intelligence.

Enterprise pilot readiness

SIA is packaged for controlled enterprise pilots. Start here:

• Release notes - what's in this release and its scope
• Installation & deployment - stand up a dedicated instance
• Changelog - version history
• Pilot readiness checklist - go-live verification
• Trust center - security & operational posture
• Security policy · Privacy · Terms · License

Trust posture

SIA is built with security-by-design practices: enforced multi-factor authentication, SCIM 2.0 user provisioning, hardened Realtime Database rules, an immutable audit trail, an edge security agent (rate limiting, prompt-injection detection, anomaly scanning), automated CI / CodeQL / security scanning, a dependency audit, a published penetration-test scope, and a documented secret-rotation procedure - see the Trust Center.

SIA has not undergone external certification. We do not claim ISO 27001, SOC 2, IEC 62443, or any other third-party certification or attestation. SIA is at pilot maturity - intended for controlled deployments alongside existing systems - and in the dedicated-instance model the customer owns their environment and data.

Current release - v1.2.1 "Enterprise Pilot Readiness"

This release consolidates SIA for controlled enterprise pilots: dedicated white-label instances, self-service SuperAdmin configuration, SCADA/PLC/historian integration readiness, configurable and drift-monitored AI models, operational monitoring with alerting, and a full release-readiness and trust documentation set. See the release notes and changelog.

Overview

Smart Industrial Alert - SIA is a full-stack industrial supervision platform designed from the ground up for high-noise, high-stakes manufacturing environments. It replaces fragmented radio systems and paper-based workflows with a unified, intelligent layer that routes alerts, verifies identities through voice biometrics, predicts failures before they happen, and provides hands-free control to supervisors - all running offline on the factory floor.

Built for factories. Engineered for scale. Designed for resilient factory operations.

🏭  Factory Operations Layer
       ↕  Real-time sync
🔥  Firebase + Cloudflare Edge
       ↕  AI pipeline
🤖  TFLite + Claude + ONNX
       ↕  Voice interface
🎤  Offline ASR + Speaker ID
       ↕  Mobile/Desktop
📱  Flutter Multi-Platform App

Core Features

Usine (Factory)
 └── Conveyor Line
       └── Workstation
             └── Machine Asset (with full history)

• Alerts are always bound to a valid node in the hierarchy
• Asset management with machine IDs and asset history
• Location-aware AI assignment - closer supervisors score higher
• QR codes at each station let supervisors identify a workstation and load its alert history instantly

Voice Intelligence

┌─────────────────────────────────────────────────────────┐
│                   VOICE PIPELINE                        │
│                                                         │
│  🎤 Native Android AudioRecord (16kHz PCM mono)         │
│         ↓ RMS silence detection (900ms threshold)       │
│  📦 Sherpa ONNX Streaming Zipformer (~17MB int8)        │
│         ↓ 30% lower WER than alternatives in noise      │
│  🧠 TFLite Speaker Embedding (conformer_tisid_small)    │
│         ↓ Cosine similarity ≥ 0.80 to authenticate      │
│  🔍 VoiceCommandParser (num-word → digit + intent)      │
│         ↓ claim / resolve / escalate / navigate         │
│  ⚡ VoiceCommandDispatcher → AlertProvider action        │
│         ↓                                               │
│  🔊 flutter_tts confirmation (max vol + focus steal)    │
└─────────────────────────────────────────────────────────┘

Voice Commands

Speaker Verification

• TFLite conformer model generates speaker embeddings from raw audio
• Cosine similarity threshold: 0.80 - tuned for factory-grade accuracy
• Quality gates: minimum 600ms of speech, SNR ≥ 6 dB
• VAD trimming: silence stripped from both ends before embedding
• Enrollment UI: guided multi-sample enrollment flow for new supervisors

Lock Screen Voice Actions

• Full-screen VoiceLockRecorderActivity launches above Android keyguard
• FCM notification carries voice_claim action button - tap opens voice claim screen on the lock screen
• Post-notification flow: record → transcribe → verify identity → execute → speak confirmation
• fullScreenIntent with boostMediaVolume() ensures TTS cuts through factory noise

Offline-First STT

• Primary: Sherpa ONNX streaming Zipformer - fully offline after one-time model download
• Fallback: speech_to_text plugin for non-Android platforms
• Zero dependency on cloud connectivity on the factory floor
• Model stored in app data directory, compressed archive on first launch

AI & Machine Learning

┌────────────────────────────────────────────────────────────────────┐
│                      AI INTELLIGENCE LAYER                         │
│                                                                    │
│  ┌─────────────────────┐   ┌──────────────────────────────────┐   │
│  │  AI ASSIGNMENT      │   │  PREDICTIVE INTELLIGENCE         │   │
│  │                     │   │                                  │   │
│  │  Scoring Model:     │   │  • Morning briefing (daily)      │   │
│  │  • Workload weight  │   │  • Failure risk curves           │   │
│  │  • Expertise match  │   │  • Hourly probability dist.      │   │
│  │  • Location prox.   │   │  • Per-factory risk scores       │   │
│  │  • Success rate     │   │  • Assignee predictions          │   │
│  │                     │   │                                  │   │
│  │  Confidence: 0-100  │   │  Powered by historical alert     │   │
│  │  Cooldown: 5 min    │   │  patterns + factory metadata     │   │
│  │  Debounce: 1.5s     │   │                                  │   │
│  └─────────────────────┘   └──────────────────────────────────┘   │
│                                                                    │
│  ┌─────────────────────┐   ┌──────────────────────────────────┐   │
│  │  AI RESOLUTION      │   │  VOICE BIOMETRICS                │   │
│  │  SUGGESTIONS        │   │                                  │   │
│  │                     │   │  TFLite Speaker Embedding        │   │
│  │  Claude API via     │   │  • 16kHz mono PCM input          │   │
│  │  Cloudflare Worker  │   │  • conformer_tisid_small.tflite  │   │
│  │                     │   │  • Cosine sim >= 0.80            │   │
│  │  Context: type,     │   │  • VAD + SNR quality gates       │   │
│  │  location, history  │   │  • Multi-enrollment averaging    │   │
│  └─────────────────────┘   └──────────────────────────────────┘   │
└────────────────────────────────────────────────────────────────────┘

AI Assignment Engine

The rule-based scoring model evaluates every available supervisor against a weighted multi-criteria function:

score = w1·(1/workload) + w2·expertise_match + w3·location_proximity + w4·recent_success_rate

• Decision transparency: every assignment includes full per-supervisor score breakdown
• Rejection feedback: supervisors can reject assignments; feedback feeds back into the scoring model
• Deduplication: 5-minute factory lock with 1.5s debounce prevents duplicate assignments
• AI Logs Panel: admin view of every AI decision with full reasoning trace

Predictive Intelligence

• Morning Briefing: AI-generated daily summary of overnight alerts, resolution stats, and risk hotspots
• Failure Prediction: probability curves per factory line for upcoming 8-hour shift
• Risk Heatmaps: visual overlay on factory layout with ML-derived risk coloring
• Predictive Cards: dashboard widgets surfacing imminent-failure signals to supervisors before alerts fire

Claude-Powered Resolution Suggestions

• Cloudflare Worker proxies requests to the Claude API
• Prompt includes: alert type, workstation location, machine asset history, past resolutions for similar alerts
• Response rendered inline in the alert detail screen
• Graceful offline fallback when edge worker is unreachable

Architecture

┌──────────────────────────────────────────────────────────────────────────────┐
│                            Smart Industrial Alert - SIA ARCHITECTURE                             │
│                                                                              │
│   ┌──────────────┐   ┌──────────────┐   ┌──────────────┐   ┌────────────┐  │
│   │   FLUTTER    │   │   FIREBASE   │   │  CLOUDFLARE  │   │  SHOREBIRD │  │
│   │   APP        │   │              │   │  WORKERS     │   │  CODE PUSH │  │
│   │              │<──┤ Realtime DB  │   │              │   │            │  │
│   │ Provider     │   │ Firestore    │<──┤ Cron AI      │   │ OTA deploy │  │
│   │ Screens      │<──┤ Cloud Funcs  │   │ Escalation   │   │ Zero-down  │  │
│   │ Services     │   │ Auth         │   │ Push fan-out │   │ time patch │  │
│   │ Models       │<──┤ FCM          │   │ JWT signing  │   └────────────┘  │
│   └──────┬───────┘   └──────────────┘   └──────────────┘                   │
│          │                                                                   │
│   ┌──────▼──────────────────────────────────────────────────────────────┐   │
│   │                     PLATFORM CHANNELS (Android)                     │   │
│   │  VoiceLockRecorderActivity  ·  boostMediaVolume()  ·  Keyguard API  │   │
│   └─────────────────────────────────────────────────────────────────────┘   │
│                                                                              │
│   ┌──────────────────────────────────────────────────────────────────────┐  │
│   │                        ON-DEVICE ML STACK                            │  │
│   │   Sherpa ONNX (offline ASR)  ·  TFLite (speaker embedding)           │  │
│   └──────────────────────────────────────────────────────────────────────┘  │
└──────────────────────────────────────────────────────────────────────────────┘

State Management

• Provider pattern with AlertProvider as the central state atom
• Streams from Firebase Realtime Database, composed with RxDart operators
• Platform-specific service resolution via conditional Dart imports (_stub.dart / _io.dart)
• Background FCM handler registered with @pragma('vm:entry-point') for foreground & killed-app delivery

Data Flow

Firebase RTDB stream
  → AlertProvider (Provider)
    → UI rebuilds (Consumer widgets)
      → User action (tap / voice command)
        → AlertProvider method
          → Firebase write
            → Cloudflare Worker triggered (if needed)
              → FCM push to all supervisors

Technology Stack

Screens & UI

Authentication
  ├── Login Screen
  └── Role Gate (Admin / Supervisor / Production Manager)

Dashboard (Overview)
  ├── Health Score Gauge
  ├── Critical Alerts Card
  ├── Predictive Failure Card
  ├── AI Morning Briefing Hero
  └── Insights Strip

Alert Management
  ├── Alerts Tree (hierarchical, heatmap-colored)
  ├── Alert Detail (comments, resolution, AI suggestion)
  ├── Alert Scan Screen (mobile / web variants)
  └── Escalated Alerts (admin view)

Voice Features
  ├── Voice Claim Screen (lock-screen capable)
  ├── Voice Enrollment Screen (guided multi-sample)
  └── Push-to-Talk overlay

Supervision
  ├── Supervisor Tab (availability, workload)
  ├── Collaboration Screen (help requests, accept/refuse)
  └── Admin Escalation Dashboard

Factory Management
  ├── Factory Mapping Tab (custom layout)
  ├── Alert Locator (custom painter map)
  └── Asset Management

Intelligence
  ├── AI Logs Panel (assignment decision trace)
  ├── Predictive Heatmaps
  └── Morning Briefing View

Theme System

Smart Industrial Alert - SIA ships with a polished dual-mode design system:

• Material Design 3 ColorScheme integration
• context.appTheme.navy, context.isDark extension helpers
• Custom card theming, navigation bar theming, and input decoration
• Inter font family throughout

Platform Support

Android is the primary deployment target. Native VoiceLockRecorderActivity and full offline Sherpa ONNX pipeline are Android-only. Other platforms use speech_to_text plugin fallback.

Backend Infrastructure

Firebase Architecture

firebase/
  ├── Realtime Database   → Live alert streams, supervisor state
  ├── Firestore           → AI decision logs, structured metadata
  ├── Cloud Functions     → Assignment cooldowns, factory locks (15s TTL)
  ├── Cloud Messaging     → Push to supervisors, voice_claim actions
  └── Authentication      → UID-based role gates

Cloudflare Workers

Primary Worker (cloudflare_worker.js - 49 KB):

• Custom Firebase JWT generation (service account signing, no Admin SDK required at edge)
• Rate-limited: max 1 alert/push, 5 escalation checks, 20 factories per cron run
• Supervisor availability scoring at the edge layer

Secondary Worker (WORKER_UPDATE_FILTER_CLAIMED.js):

• Claimed-filter real-time synchronization for UX snappiness

Firebase Cloud Functions

• AI assignment cooldown enforcement
• Factory-level lock acquisition with 15-second TTL (prevents race conditions)
• Supervisor availability tracking
• Alert deduplication and processor matching

Getting Started

Prerequisites

# Flutter SDK (3.x recommended)
flutter --version

# Firebase CLI
npm install -g firebase-tools
firebase login

# Android SDK (API 21+)
# NDK required for Sherpa ONNX native libs

Installation

# 1. Clone the repository
git clone https://github.com/your-org/Smart Industrial Alert - SIAapp.git
cd Smart Industrial Alert - SIAapp

# 2. Install Flutter dependencies
flutter pub get

# 3. Configure environment
cp .env.example .env
# Fill in your Firebase project credentials and Cloudflare Worker URL

# 4. Configure Firebase
flutterfire configure --project=YOUR_FIREBASE_PROJECT_ID

# 5. Run on Android (recommended)
flutter run --release

First Launch

1. Voice Enrollment: Supervisors record 3–5 samples to build their voice profile
2. Model Download: Sherpa ONNX ASR model (~17 MB) downloads once, then fully offline
3. FCM Token: Device registers for push notifications automatically
4. Role Assignment: Admin assigns supervisor/manager roles via Firebase console

Configuration

Environment Variables (.env)

CLOUDFLARE_WORKER_URL=https://your-worker.your-subdomain.workers.dev
FIREBASE_API_KEY=...
FIREBASE_APP_ID=...
FIREBASE_MESSAGING_SENDER_ID=...
FIREBASE_PROJECT_ID=...
FIREBASE_STORAGE_BUCKET=...
FIREBASE_DATABASE_URL=...

Escalation Thresholds (Cloudflare Worker)

// cloudflare_worker.js
const ESCALATION_THRESHOLDS = {
  qualite:           10 * 60 * 1000,  // 10 minutes
  maintenance:       15 * 60 * 1000,  // 15 minutes
  defaut_produit:     5 * 60 * 1000,  //  5 minutes
  manque_ressource:  10 * 60 * 1000,  // 10 minutes
};

AI Assignment Weights

// lib/services/ai_assignment_service.dart
const double workloadWeight    = 0.35;
const double expertiseWeight   = 0.30;
const double locationWeight    = 0.20;
const double successRateWeight = 0.15;

Voice Auth Sensitivity

// lib/services/voice_auth_service_io.dart
const double similarityThreshold = 0.80;   // cosine similarity
const double minSNRdB            = 6.0;    // minimum signal quality
const int    minSpeechMs         = 600;    // minimum speech duration (ms)

Project Structure

Smart Industrial Alert - SIAapp/
│
├── lib/
│   ├── main.dart                          # App entry, Firebase init, auth gate
│   ├── theme.dart                         # AppTheme light/dark + context extensions
│   │
│   ├── models/                            # Data models
│   │   ├── alert_model.dart
│   │   ├── user_model.dart
│   │   ├── collaboration_model.dart
│   │   └── hierarchy_model.dart
│   │
│   ├── providers/                         # State management
│   │   └── alert_provider.dart            # Central alert + supervisor state
│   │
│   ├── services/                          # 30+ service classes
│   │   ├── alert_service.dart
│   │   ├── ai_service.dart                # Claude API proxy
│   │   ├── ai_assignment_service.dart     # Scoring + auto-assignment
│   │   ├── predictive_intel_service.dart  # Morning briefings, risk curves
│   │   ├── voice_service.dart             # Platform-conditional entry
│   │   ├── voice_service_io.dart          # Android native impl
│   │   ├── voice_service_stub.dart        # Non-Android stub
│   │   ├── voice_auth_service_io.dart     # TFLite speaker verification
│   │   ├── sherpa_stt_service_io.dart     # Offline ASR
│   │   ├── voice_command_parser.dart      # Intent + number parsing
│   │   ├── voice_command_dispatcher.dart  # Command → AlertProvider bridge
│   │   ├── fcm_service.dart
│   │   ├── offline_db_service.dart
│   │   └── ...
│   │
│   ├── screens/                           # 24+ screens
│   │   ├── dashboard_screen.dart
│   │   ├── alerts_tree_screen.dart
│   │   ├── voice_claim_screen.dart
│   │   ├── voice_enrollment_screen.dart
│   │   ├── alert_scan_screen.dart
│   │   ├── admin_escalation_screen.dart
│   │   ├── collaboration_screen.dart
│   │   └── ...
│   │
│   └── widgets/                           # Reusable components
│       ├── voice_command_button.dart
│       ├── ai_logs_panel.dart
│       ├── predictive_heatmap_widget.dart
│       └── ...
│
├── android/
│   └── app/src/main/kotlin/com/example/Smart Industrial Alert - SIAapp/
│       ├── MainActivity.kt
│       └── VoiceLockRecorderActivity.kt   # Native audio + keyguard bypass
│
├── assets/
│   └── models/
│       └── conformer_tisid_small.tflite   # Speaker embedding model
│
├── functions/                             # Firebase Cloud Functions (Node.js)
│   └── index.js
│
├── cloudflare_worker.js                   # Edge AI scoring + escalation cron
├── WORKER_UPDATE_FILTER_CLAIMED.js        # Claimed filter sync worker
├── pubspec.yaml                           # Flutter dependencies
├── firebase.json                          # Firebase config
└── shorebird.yaml                         # Code push configuration

Built With Conviction

"Factory floors don't wait. Neither should your alerts."