Now in production · System One decision model dual-decision layer active

Detect. Decide.
Protect.

AI-powered HSE surveillance for oil & gas — PPE compliance, zone access, and safety violation detection with calibrated dual-AI decisions in under 154ms.

Request Demo Explore the Technology
Scroll

The Platform

Three layers. One decision.

vAIgilant combines GPU-accelerated computer vision, calibrated AI decision models, and a real-time command center into a unified HSE intelligence platform.

🎯

DeepStream Vision AI

NVIDIA DeepStream 9.1 with TensorRT-optimized YOLO models processes every camera frame with GPU acceleration. Detects PPE violations, fall events, zone breaches, vehicle proximity, and 12+ other safety scenarios simultaneously across 64+ camera streams.

🧠

Decision Model Layer

A dual-model System One AI layer replaces rule-based threshold logic with calibrated probability distributions. System One decision model classifies alert metadata in 114ms; vision decision model visually verifies from the camera frame in 40ms. Together they cut false positives by 40–60%.

🖥️

Real-Time Command Center

SvelteKit frontend with WebSocket-driven live plant map, trajectory tracking, and role-based alert dashboards. HSE Managers see auto-escalated critical events; Representatives review medium-confidence cases with probability breakdowns for fast, informed decisions.

Decision Model Intelligence

"Detect and Decide" — a new category

Traditional surveillance detects events. vAIgilant goes further: it makes calibrated, auditable safety decisions with probability distributions that are statistically meaningful — enabling automated escalation thresholds that would be irresponsible with LLM confidence scores.

0
× faster than LLMs
0ms
End-to-end decision latency
$0
For 64 cameras (cloud decision model API)
0%
False positive reduction

Per-event System One decision model questions

Q1 Noul: "Genuine safety violation?" → 0.94
Q2 Score: "DEESI severity [1–5]" → 3.7
Q3 Choice: "Category [ppe/fall/fire/zone]" → ppe:0.71
Q4 Noul: "Escalate to HSE Manager?" → 0.88

Cost: ~$0.000004 · Latency: ~114ms · All parallel

Dual Decision Pipeline

📹
DeepStream Detection
YOLO + TensorRT → Redis pub/sub
<1ms
↓
🤖
Decision Model — Metadata
4 questions, 1 request, calibrated probs
~114ms
↓ if ≥80% confidence
👁️
Vision Decision Model
Frame verification, self-hosted GPU
~40ms
↓
⚡
Confidence Router
Both ≥90% → auto-alert · 60–90% → review · <60% → suppress
IOGP aligned
↓
🚨
RBAC Alert Dispatch
WebSocket → HSE Manager / Representative
DEESI mapped
Decision Model
Cloud API · $0.042/M tokens · 40–200× cheaper than LLMs · text/metadata only
Vision Model
Self-hosted · Apache 2.0 · 5B params · 12GB VRAM · shares DeepStream GPU

How It Works

From frame to decision in 154ms

Every camera event travels through a disciplined pipeline: GPU vision detection, AI metadata classification, visual frame verification, and confidence-gated escalation — fully automated for high-confidence events, human-in-the-loop for everything else.

64+
Camera streams
5.5M
Events/day processed
~$2.30
Total AI cost per day (64 cameras)
Step 01
Camera Feed Ingestion
RTSP streams from plant cameras are ingested into NVIDIA DeepStream via GPU-accelerated NVDEC decoding. Up to 64 concurrent streams at full resolution.
Step 02
DeepStream Detection
TensorRT-optimized YOLO models run on GPU to detect persons, PPE items, vehicles, zone violations, and anomalies. Detection metadata is published to Redis pub/sub in real time.
Step 03
Decision Model Triage
The FastAPI backend sends detection metadata to the System One decision model. Four questions — severity, category, genuine violation, escalation — are answered in one 114ms API call with calibrated probabilities.
Step 04
vision decision model Frame Verification
For events where System One decision model confidence ≥80%, the actual camera frame is sent to the self-hosted vision decision model. It visually confirms the violation in ~40ms using the same GPU as DeepStream.
Step 05
Confidence-Gated Dispatch
Both signals are fused. High confidence (≥90% both models) → automatic alert to HSE Manager. Medium → human review queue with probability breakdown. Low → suppress and log only.
Step 06
DEESI Lifecycle Alert
Alerts are tagged with their DEESI stage and dispatched to the appropriate RBAC role via WebSocket. The frontend shows a full probability audit trail for every alert — critical for IOGP compliance reporting.

DEESI Framework

Aligned with IOGP Life-Saving Rules

Every alert follows the DEESI lifecycle — a structured approach to safety event management aligned with IOGP standards. decision model score questions map directly to DEESI stages, enabling automated workflow routing.

01
🔍
Detect
System identifies a potential safety event from camera feeds and sensor data
decision model score 1.0–1.5
02
⚖️
Evaluate
System One decision model classifies severity, category, and probability of genuine violation
decision model score 1.5–2.5
03
📣
Escalate
Responsible HSE personnel are automatically notified via role-based WebSocket alerts
decision model score 2.5–3.5
04
🚑
Support
Response resources are deployed; incident management workflows are triggered
decision model score 3.5–4.5
05
📈
Improve
Post-incident learning triggers model recalibration and procedure updates
decision model score 4.5–5.0

By the Numbers

Performance that scales with your plant

64+
Camera Streams
Simultaneous RTSP feeds processed on a single GPU node
<154ms
End-to-End Decision Latency
Detection → System One decision model triage → vision decision model visual verify → alert dispatch
94%
Detection Confidence
System One decision model Noul score for genuine safety violations in production
60%
False Positive Reduction
Dual-model confidence gating vs rule-based threshold logic
86%
Gross Margin (Option B)
Decision model layer at $499/mo for 64 cameras — $69/mo underlying cost
~$0
vision decision model Additional Cost
Runs on existing DeepStream GPU — no new hardware required for vision AI

Use Cases

Built for industrial environments

Designed for oil & gas refineries, petrochemical plants, LNG terminals, and offshore facilities — environments where a missed safety violation has life-or-death consequences.

🦺

PPE Compliance

Detect missing hard hats, high-visibility vests, safety gloves, steel-toed boots, and protective goggles across all monitored zones. System One decision model scores context — a violation in a live hydrocarbon area at 2 AM is weighted differently than one in the cafeteria. Automated escalation at configurable confidence thresholds.

Hard Hat Detection Vest Monitoring Gloves & Boots IOGP Life-Saving Rule #1
🚧

Restricted Zone Access

Define restricted, limited-access, and vehicle-exclusion zones by drawing polygons on live camera views. Detect unauthorised entry, maximum occupancy breaches, and prohibited vehicle movement. System One decision model compounds zone type, time of day, personnel count, and authorization level into a single severity score.

Zone Polygon Drawing Occupancy Limits Vehicle Exclusion Night-Shift Weighting
👤

Lone Worker Safety

Monitor workers operating alone in high-risk areas. Detect stationary persons, abnormal postures, fall events, and man-down scenarios using skeletal pose estimation. Lone worker + restricted zone + no PPE + no recent check-in triggers a critical compound System One decision model severity score requiring immediate escalation.

Fall Detection Pose Estimation Man-Down Alert Compound Risk Scoring
🔥

Leak & Flame Detection

Visual detection of liquid leaks, abnormal vapor plumes, and open flame events with thermal cross-validation.

🚗

Vehicle & LPR

License plate recognition, speed monitoring, and vehicle-pedestrian proximity alerts with live plant map tracking.

🚭

Behavioural Detection

Smoking detection, abnormal gathering behaviour, and unauthorized equipment interaction in live hydrocarbon zones.

Who We Serve

Built for high-hazard industrial environments

vAIgilant is purpose-built for facilities where safety violations have life-or-death consequences and compliance is non-negotiable.

🛢️

Upstream Oil & Gas

Drilling platforms, wellhead facilities, and production sites where lone workers and permit-to-work compliance are critical.

🏭

Refineries & Petrochemical Plants

High-occupancy process areas where PPE compliance, zone access control, and hot-work monitoring protect against major accident hazards.

🚢

LNG Terminals & Offshore Platforms

Air-gapped deployments for data sovereignty. Full on-premise stack — no cloud dependency, no sensitive footage leaving the facility.

⚗️

Chemical & Specialty Plants

Multi-hazard environments needing simultaneous monitoring of PPE, chemical handling procedures, spill detection, and restricted-area access.

⚡

Power Generation

High-voltage exclusion zones, arc-flash PPE verification, and contractor workforce monitoring across large generation facilities.

⚓

Port & Terminal Operations

Vehicle-pedestrian conflict zones, cargo handling PPE compliance, and dangerous goods proximity monitoring in busy terminal environments.

Technology Stack

Enterprise-grade, open foundations

Built on industry-proven open-source foundations with commercial GPU acceleration — deployable on-premise, air-gapped, or hybrid cloud.

NVIDIA DeepStream 9.1 TensorRT YOLO Object Detection System One Decision Models Vision Decision Model (Apache 2.0) FastAPI · Python 3.11 SvelteKit PostgreSQL + TimescaleDB Redis 7 Pub/Sub WebRTC · mediamtx Three.js Docker Compose IOGP Aligned RBAC Security

Decision Approach Comparison

Approach Cost / 5.5M events/day Latency Calibrated?
vAIgilant (dual decision model layer) ~$2.30/day ~154ms ✓ Yes
Claude Haiku (per event) ~$46/day ~800ms+ ✗ No
GPT-4o-mini (per event) ~$30/day ~600ms+ ✗ No
Rule-based thresholds $0 <1ms ✗ N/A

FAQ

Frequently Asked Questions

Everything you need to know about AI HSE surveillance for oil & gas.

vAIgilant is an AI-powered HSE (Health, Safety & Environment) surveillance platform designed for oil and gas, refineries, petrochemical plants, and LNG terminals. It combines NVIDIA DeepStream GPU-accelerated computer vision with a dual decision model layer to detect safety violations, classify severity, and dispatch calibrated alerts — all in under 154ms. Unlike rule-based systems, vAIgilant returns probability distributions, not binary flags.
vAIgilant uses a dual-model approach: a System One metadata decision model classifies detection metadata (object class, zone type, time of day, confidence score) and asks 4 calibrated questions in 114ms. For events above 80% confidence, a vision decision model visually verifies the actual camera frame in ~40ms. Both models must agree before auto-alerting — reducing false positives by 40–60% compared to single-model, rule-based threshold logic.
System One decision models are a new category of AI that return calibrated probability distributions over typed questions — not free text. They are 40–200× faster and 20–444× cheaper than LLMs for classification tasks, and critically, their confidence scores are statistically meaningful. vAIgilant uses a System One decision model metadata classifier (cloud API, ~$2/day for 64 cameras) for alert triage and a vision decision model (self-hosted, Apache 2.0) for visual frame verification.
Yes. vAIgilant supports fully air-gapped, on-premise deployments where no data leaves the facility. The vision decision model runs locally on the same NVIDIA GPU as DeepStream — no cloud API is required for visual decisions. For fully air-gapped deployments, a text-only decision model variant handles metadata classification as well, making the entire decision layer self-contained on-site.
vAIgilant ingests standard RTSP streams from any IP camera system — no camera replacement required. It deploys as Docker containers alongside NVIDIA GPU hardware (minimum 24 GB VRAM recommended for the vision model). The platform connects to your existing network and adds the AI decision layer on top of your current CCTV investment.
DEESI is a structured safety event lifecycle aligned with IOGP Life-Saving Rules: Detect, Evaluate, Escalate, Support, Improve. vAIgilant's decision model Score questions map directly to DEESI stages — a score of 1.0–1.5 triggers awareness (Detect), while 3.5–4.5 deploys response resources (Support). Every alert carries a full probability distribution for each question, creating an auditable compliance trail for IOGP reporting and post-incident investigation.

Get a Free Demo

Request Your Free Demo

See vAIgilant running on live camera feeds from a simulated process plant. We'll walk through PPE detection, the System One decision model dual-decision layer, DEESI alert dashboard — and size the right deployment for your facility.

✅
Live detection demo
Watch PPE violations, zone breaches, and lone worker alerts trigger in real time
✅
System One decision model probability walkthrough
See calibrated probability distributions for every alert — full audit trail for compliance
✅
Deployment architecture review
On-premise, air-gapped, or hybrid — we'll size the right setup for your plant
✓

Request received

Our team will reach out within one business day to schedule your demo.