Agentic AI Architecture for Manufacturing Intelligence
The industrial sector is transitioning from deterministic Operational Technology (OT) to a "Cognitive Industrial Edge," where probabilistic Generative AI is integrated directly into factory operations.
Address the limitations of siloed data and fragmented legacy protocols by deploying autonomous AI agents within secure, air-gapped environments.
Small Language Models (SLMs) and quantised LLMs on robust edge compute enable machinery to reason about its state and autonomously optimise production.
Air-gapped "Fortress Factory" architecture ensures AI reasoning occurs locally, isolated from public internet threats.
The result: A resilient, secure ecosystem that transforms manufacturing from passive data collection to active, agentic orchestration — delivering significant improvements in Overall Equipment Effectiveness (OEE) and quality yield.
The Cognitive Industrial Edge bridges the gap between low-level industrial protocols ("Southbound") and high-level Enterprise Resource Planning systems ("Northbound").
ERP Systems • SAP • Oracle • Dynamics 365
Agentic AI • SLMs • Vision Models • Decision Engine
NVIDIA IGX • Jetson • Ruggedised Servers
OPC UA • Modbus • MQTT • PROFINET • EtherNet/IP
PLCs • Sensors • Actuators • Vision Systems
The "Southbound" layer manages the heterogeneous communication environment of the factory floor, normalising diverse data dialects into a semantic model usable by AI agents.
Unified Architecture
The modern standard for interoperability. Provides semantic richness, allowing agents to "browse" and self-discover machine capabilities via metadata and information modeling.
RTU and TCP/IP
The foundational standard for industrial communication. Because it lacks semantic context (transmitting only raw 16-bit registers), the AI requires a mapping layer to interpret data.
High-Frequency Telemetry
Optimised for high-frequency telemetry. Sparkplug B provides structured payloads and "Birth/Death Certificates," allowing the AI to maintain a real-time model of the device fleet.
PROFINET & EtherNet/IP
Essential for motion control and safety. Integrated via gateways that expose cyclic data to the AI layer for real-time decision making.
A specialised software component normalises all protocol inputs into a unified format (CloudEvents), enabling AI agents to work with consistent, semantic data regardless of the underlying industrial protocol.
Security requirements for critical infrastructure necessitate a "Fortress Factory" model where AI reasoning occurs locally, isolated from the public internet.
AI models at the edge require high parallel processing power and industrial-grade durability.
NVIDIA IGX / Jetson AGX Orin
Used for autonomous machines and cobots. Bridges the gap between AI (Linux) and Safety (RTOS).
Dell PowerEdge XR8000 / HPE Edgeline
Hosts the central Cognitive Core for a production cell. Capable of running quantised 70B parameter models.
NVIDIA Jetson Orin Nano
Dedicated to single-task agents like visual inspection or quality control.
To maintain air-gapped integrity while allowing enterprise visibility:
Hardware-enforced unidirectional gateways that allow telemetry to flow out to corporate IT but prevent any external data from entering the secure zone.
Updates performed via a "Decontamination Station" (kiosk). Encrypted, signed containers transferred via virus-scanned media and verified against a local TPM public key.
The Cognitive Core moves beyond rigid logic to "Agentic AI," where models perceive environments, plan sequences, and execute actions using a Supervisor-Worker pattern.
Optimised for efficiency (Phi-3, Gemma, Mistral 7B) to handle log interpretation, anomaly detection, and code generation at the edge.
Enable the system to "see" and interpret visual anomalies using models like LLaVA for quality inspection and defect detection.
Models compressed to 4-bit or 8-bit precision to fit within edge GPU VRAM limits while maintaining inference quality.
Decomposes high-level objectives into sub-tasks and routes them using state machines. Orchestrates the overall workflow.
Translates natural language requests into specific industrial protocol commands (e.g., OPC UA NodeIDs, Modbus registers).
Detects anomalies by monitoring telemetry streams against statistical baselines. Predicts failures before they occur.
The only agent authorised to write commands to PLCs. Operates within strict, hard-coded safety guardrails.
The AI system is strictly separated from functional safety. This is a fundamental architectural principle.
Principle: The Safety PLC remains solely responsible for emergency protection. AI commands are restricted by deterministic boundaries that cannot be overridden by probabilistic outputs.
The Cognitive Core integrates with ERP systems to ensure production actions are reflected in financial and inventory records.
The architecture prioritises REST-based APIs for modern integration.
Large Enterprise
OData (REST)
Production Orders, Material Documents
Mid-Market
SuiteTalk REST
Work Orders, Inventory Adjustments
Enterprise/Mid
OData v4
Warehouse Journals, Production Orders
Manufacturing
REST API v2
Job Entry, Material Queue
To manage air-gapped constraints and latency:
A lightweight database (SQLite/Redis) mirrors ERP data (BOMs, schedules) locally for real-time access.
Data is pulled from the ERP at shift-start and pushed back in batches (Uplink) at defined intervals.
Discrepancies (e.g., insufficient stock) trigger alarms for human intervention with full audit trail.
The user interface follows the ISA-101 standard to provide high-performance graphics and situational awareness.
"Talk to Your Factory"
Uses Text-to-SQL technology. CXOs can query complex metrics using natural language, which the SLM translates into SQL queries for real-time visualisation.
Digital Twin Forecasting
The system uses a Digital Twin powered by historical data to forecast the impact of operational changes before implementation.
Simulation to Execution
If a simulation is approved, the Control Agent orchestrates the setpoint changes across the PLC network automatically.
Reduced Unplanned Downtime through AI-driven predictive maintenance and anomaly detection.
Visual agents reduce scrap rates by catching defects at the source before they propagate.
Democratised data access allows for faster, data-driven executive decision-making.
A structured approach to building trust and demonstrating value:
Read-only mode to validate AI accuracy and build trust. The system monitors, learns, and generates recommendations without taking action.
AI suggests changes that require human approval on the HMI. Operators validate recommendations before execution.
AI manages low-risk optimisation loops and automatic quality-based line stops within strict safety guardrails.
Our team has hands-on experience with OPC UA, Modbus, PROFINET, and EtherNet/IP in production environments. We understand the nuances of real-time industrial communication.
We design AI systems that operate within strict security constraints. Our "Fortress Factory" approach ensures AI reasoning occurs locally, isolated from external threats.
Experience deploying quantised SLMs and VLMs on NVIDIA edge hardware. We understand the constraints of industrial-grade compute environments.
We architect AI systems that respect functional safety boundaries. Our systems are designed to enhance operations without compromising safety integrity levels.
Proven integration patterns for SAP, Oracle, Dynamics 365, and manufacturing-specific ERPs like Epicor. We bridge the IT/OT divide.
Our phased rollout approach—Silent Observer → Co-Pilot → Bounded Autonomy—builds trust and demonstrates value before granting AI control.
Book a discovery call to discuss how agentic AI can transform your manufacturing operations.