AgentMesh and AgentNexus run the same agents, models, safety lanes and Evidence Log on Unovie's custom-built AMD and NVIDIA hardware accelerators, engineered for agentic performance, low latency and high availability. AMD Ryzen AI Max+ 395 is the reference build for both. NVIDIA Jetson AGX Thor brings AgentMesh inside the machine; NVIDIA DGX Spark runs AgentMesh on a desk or a small AgentNexus for a site; Qualcomm QCS6490 extends sensing to the far edge.
The AgentMesh Sovereign Pod (2 × 10GbE, NVMe) runs today; the AgentNexus rack is a reference design on the same server.
Robotics & DeepVision at the tool; the AI engineering lab — brought to NVIDIA per project.
Remote and mobile assets, feeding the nearest AgentMesh pod.
Every option runs the same Unovie platform. Pick by where the work happens and how much of it there is.
| AMD Ryzen AI Max+ 395 | NVIDIA Jetson AGX Thor | NVIDIA DGX Spark | Qualcomm QCS6490 | |
|---|---|---|---|---|
| Industry 4.0 purpose | Sovereign factory intelligence | Robotics & DeepVision for fabs | AI engineering lab | Mobile & remote assets |
| Role | AgentMesh Sovereign Pod · AgentNexus rack | AgentMesh at the tool or robot | Training · test stands · site AgentNexus | Far-edge sensing for AgentMesh |
| Where it sits | Factory network · central cabinet | Inspection tool · robot cell · AMR | Engineering lab · server room | Robots, AGVs · remote assets |
| Agentic jobs | Robot & motion health · battery line safety · fab utilities · factory network watch · quality records | Wafer & die inspection · robot-cell quality · wafer handling & AMRs · tool health | Skill adapter training · test-stand analytics · digital-twin experiments · site agents | Mobile robots · AGVs · drone docks · battery containers |
| Memory | 64 GB per server (128 GB builds) | 128 GB unified | 128 GB shared | On-device |
| Form | 4 servers / 3U · 48 servers / 42U | Module · 40–130 W | Desk-sized unit | Fanless box |
| Status | Runs today | Available on | Available on | Far edge |
Built for sovereign factory intelligence: AI on the signals your lines and equipment produce, kept inside your site. Four of these quiet, compact servers make an AgentMesh Sovereign Pod in 3U next to the control system; twelve pods make an AgentNexus rack that runs agents across the fleet. Each server has sixteen Zen 5 cores, a Radeon 8060S graphics processor and an XDNA 2 AI accelerator sharing one pool of memory.
Four servers in 3U with 2 × 10GbE networking and NVMe storage, on the factory network: Signal Encoders, the Unovie Decision Engine, the Unovie Reasoning Model and the Kiln — air-gapped by default.
Twelve pods in one 42U rack: agents on the processors; models, skill adapters and training on the graphics processors; data copied across shelves.
The XDNA 2 accelerator (51 TOPS, measured on our servers) is being tested for small helper models, to free graphics time for the main model.
One box, three compute engines.
Four servers in 3U, linked.
Twelve pods, one system.
A pod per plant, one AgentNexus.
The same server runs both systems: four in a Sovereign Pod at each line or site, forty-eight in one AgentNexus rack.
Robot & Motion Health on robots, motion axes, spindles and conveyors: fault class, severity and a drafted work order.
Weld and formation signatures in line; on storage sites, runaway precursors raised as an alarm into the fire panel.
Pumps, chillers, scrubbers, drives and tool power at fabs and their suppliers: early warnings with a drafted work order.
Unauthorized writes and unusual commands on the factory network, with evidence mapped to NIST SP 800-53.
Agents group findings across lines and sites, draft non-conformance reports and quality records, and plan repairs around production.
The same fault pattern seen on one line is recognised on the next; tested skill adapters roll out fleet-wide.
Four servers on a sliding shelf in a compact cabinet, with 2 × 10GbE networking and NVMe storage, on the factory network.
Twelve trays of four servers, two switches and dual power feeds, in a central cabinet.
| Where it sits | Factory network (pod) · central cabinet (rack) |
| Form | 4 servers in 3U · 48 servers in 42U |
| Per server | 16 Zen 5 cores · Radeon 8060S graphics · XDNA 2 accelerator · 64 GB shared memory · NVMe storage |
| Network | 2 × 10GbE per server to two switches · fast cables inside each shelf |
| Environment | Climate-controlled rack, 0–35 °C |
| Status | Pod runs today · rack is a reference design on the same server |
| Unovie platform component | Sovereign Pod | AgentNexus rack |
|---|---|---|
| Unovie Gatewayread-only plant connections | ● | ◐ |
| Signal Encoderssignal fingerprints | ● | — |
| Vector Memory"seen this before?" | ● | ● |
| Unovie Decision EngineLane A / Lane B | ● | — |
| Unovie Reasoning Modelroot cause, work orders | ● | ● |
| Kilntested improvements | ● | ● |
| Evidence Logsigned record | ● | ● |
| Unovie Agent Runtimefleet agents in sandboxes | ◐ | ● |
| Unovie Model Routermodel + skill adapter per request | ◐ | ● |
| Mesh Arbiterfailover | ● | ● |
● runs here · ◐ light or partial role · — not used. The pod forwards events to AgentNexus; the rack never reaches back into factory control.
Industrial PC makers now ship compact systems built on the Ryzen AI Max+ 395 with dual 10GbE, aimed at running language models on-premise in plants.
Industrial AI software partners run local models on Ryzen AI processors to detect anomalies, predict failures and trigger guided actions from live plant data — without cloud inference bills.
AMD's embedded Ryzen AI line extends the same architecture to robotics, industrial PCs and machine vision, with industrial lifecycles.
AgentMesh and AgentNexus put these servers to work as a complete system: asset signals, safety lanes, evidence and fleet agents — not just a box that runs a model.
Public announcements and vendor material, November 2026 · vendor-reported, not Unovie measurements
Built for robotics and DeepVision in the fab. Jetson AGX Thor puts a Blackwell-class graphics processor and 128 GB of memory into a module that lives inside a robot cell, an inspection tool or a mobile robot — running high-resolution vision models next to the camera and the motion controller, with AgentMesh's safety lanes and Evidence Log around every decision.
Thor ships as a compact, fan-cooled edge node: a dense I/O wall of USB, networking, display and capture, with a Blackwell GPU and 128 GB of unified memory behind it. Mount it on the line, in the cab or at the cell — and run the models where the data is born.
High-resolution vision models find defects, particles and misalignment on wafers, dies, panels and assemblies at line speed — on the tool, without sending images off the floor.
Perception and checks for robot cells, wafer-handling robots and autonomous mobile robots, combined with motor, drive and vibration signals.
A reject or a hold is Lane A; anything that touches people or process safety is Lane B — an alarm and a request to the cell controller and the operator. Every decision lands in the Evidence Log.
Cameras and sensors stream into Thor.
Vision models score every part.
Lanes apply the safety boundary.
Findings flow to AgentNexus and the Kiln.
Thor is where Unovie uses vision first: the cameras and models sit on the tool, inside the same safety lanes as every AgentMesh decision.
DeepVision models find defects, particles and scratches on wafers and dies at the tool, holding or rejecting a lot through Lane A.
Solder, placement and surface defects on boards, panels and assemblies, checked at line speed.
Vision plus process signatures — weld current, torque, vibration — confirm every robot cycle.
Perception and health checks for wafer-handling robots and autonomous mobile robots moving lots between tools.
Vacuum pumps, chillers and motion stages watched through their own vibration and electrical signals.
Images and models stay on the tool; only findings and evidence move to AgentNexus.
| Where it sits | Inside an inspection tool, a robot cell, a wafer-handling robot or an AMR |
| Form | Compact module · 40–130 W configurable (NVIDIA spec) |
| Memory | 128 GB unified (NVIDIA spec) |
| Connectivity | Works disconnected · findings sync to AgentNexus |
| Status | Available on — the AgentMesh software is brought to Thor per project |
| Unovie platform component | Jetson AGX Thor |
|---|---|
| Unovie Gatewayread-only plant connections | ● |
| Signal Encoderssignal fingerprints | ● |
| Vector Memory"seen this before?" | ● |
| Unovie Decision EngineLane A / Lane B | ● |
| Unovie Reasoning Modelroot cause, work orders | ● |
| Kilntested improvements | ◐ |
| Evidence Logsigned record | ● |
| Unovie Agent Runtimefleet agents in sandboxes | — |
| Unovie Model Routermodel + skill adapter per request | — |
| Mesh Arbiterfailover | — |
● runs here · ◐ light role (nightly playbook only; training happens on AgentNexus or DGX Spark) · — not used.
Industrial Thor systems are being deployed for multi-camera quality inspection, automated optical inspection and defect detection on the line.
Industrial robot makers and heavy-equipment makers are adopting Thor for manufacturing cells and autonomous equipment that fuse many sensor streams on the device.
A major industrial computer maker reports running an agentic “factory brain” on Thor in its own manufacturing plants.
Thor with a sensor bridge ingests camera, LiDAR, IMU and encoder data for real-time pipelines in harsh, vibrating environments.
Unovie focuses Thor on DeepVision in fabs — wafer, die and panel inspection and wafer handling — inside AgentMesh's safety lanes. Public fab deployments on Thor are still emerging.
Public announcements and vendor material, November 2026 · vendor-reported, not Unovie measurements
Built for the AI engineering lab. DGX Spark pairs a Grace Blackwell superchip with 128 GB of shared memory on a desk: train and test new skill adapters, run test-stand analytics and digital-twin experiments on recorded data, and host a small AgentNexus for one site — then move the same models to the plant unchanged.
Spark is a desktop-sized chassis with a perforated cooling top and a full I/O wall — a GB10 Grace Blackwell Superchip and 128 GB of coherent memory inside. Develop, fine-tune and serve large models locally; promote them to the edge unchanged.
The Kiln trains new skill adapters here, which then go through the same tests before they reach production.
Analyse test-rig runs and replay recorded factory data against new models before anything touches the factory network.
One to four units host agents, the Unovie Model Router and skill adapters for a single site or a pilot.
Bring in recorded plant data.
The Kiln trains a skill adapter.
Prove it on unseen cases.
Same model to the plant.
When a pilot grows, the same agents, models and adapters move to the AMD rack unchanged.
The Kiln trains new skill adapters on recorded plant outcomes, ready for testing.
Pressure, strain and acoustic data from test rigs analysed run by run, with drafted test reports.
Replay recorded plant data against new models and rules before anything reaches the plant.
A small AgentNexus for one site: planner, shift-handover and quality agents working across lines.
| Where it sits | Engineering office, server room or a shelf near the line |
| Form | Desk-sized unit · two can be linked |
| Memory | 128 GB shared (NVIDIA spec) |
| Roles | Single-unit AgentMesh · site AgentNexus · training node |
| Status | Available on — brought to Spark per project |
| Unovie platform component | As AgentMesh | As site AgentNexus |
|---|---|---|
| Unovie Gatewayread-only plant connections | ● | ◐ |
| Signal Encoderssignal fingerprints | ● | — |
| Vector Memory"seen this before?" | ● | ● |
| Unovie Decision EngineLane A / Lane B | ● | — |
| Unovie Reasoning Modelroot cause, work orders | ● | ● |
| Kilntested improvements | ● | ● |
| Evidence Logsigned record | ● | ● |
| Unovie Agent Runtimefleet agents in sandboxes | — | ● |
| Unovie Model Routermodel + skill adapter per request | ◐ | ● |
| Mesh Arbiterfailover | ◐ | ◐ |
● runs here · ◐ light or partial role · — not used.
Industrial DataOps platforms now run on DGX Spark, bringing language-model capability to factory data with lower latency and data kept on site.
Demonstrations stream real camera feeds into DGX Spark to keep a simulation-based digital twin of a cell continuously in step with the physical world.
A supported software stack with lifecycle management and security updates is offered for smart-manufacturing and inspection deployments.
Unovie uses Spark as the AI engineering lab: training and testing skill adapters, replaying recorded plant data, and hosting a small AgentNexus for a pilot.
Public announcements and vendor material, November 2026 · vendor-reported, not Unovie measurements
Built for mobile and remote assets. Where wires and watts are scarce — a mobile robot, an AGV, a drone dock, a battery container — a fanless QCS6490 box collects and pre-processes signals on site and forwards compact fingerprints to the nearest AgentMesh pod.
QCS6490 reference hardware brings a full I/O wall — USB-C, USB 3.0, dual Ethernet, 10GbE and HDMI — to a compact, fanless box. Premium-tier on-device AI without the power bill, deployed where wires and watts are scarce.
Reads sensors where running cables would be costly or impossible.
Turns raw signal into compact fingerprints before sending anything.
Sends findings to the nearest pod, where the safety lanes and Evidence Log apply.
Sensors on remote assets.
Fingerprints on the device.
Wireless to the nearest pod.
AgentMesh applies the lanes.
QCS6490 senses and prepares; the decisions, lanes and Evidence Log stay on the nearest AgentMesh pod.
Motor, wheel and battery health on robots that move, pre-processed on board.
Motor, charge and environment signals at docks and outdoor equipment, with only compact findings sent back.
Battery and off-gas data from remote containers, forwarded to the nearest pod for the safety lanes.
Sensors on moving and hard-to-cable equipment, without running new cable.
| Where it sits | Next to remote assets, in a small fanless enclosure |
| Power | Low power, fanless (Qualcomm spec) |
| Connectivity | Wi-Fi 6E and Bluetooth · forwards to the nearest pod |
| Status | Far-edge sensing that feeds AgentMesh |
| Unovie platform component | QCS6490 |
|---|---|
| Unovie Gatewayread-only plant connections | ◐ |
| Signal Encoderssignal fingerprints | ● |
| Vector Memory"seen this before?" | — |
| Unovie Decision EngineLane A / Lane B | — |
| Unovie Reasoning Modelroot cause, work orders | — |
| Kilntested improvements | — |
| Evidence Logsigned record | — |
| Unovie Agent Runtimefleet agents in sandboxes | — |
| Unovie Model Routermodel + skill adapter per request | — |
| Mesh Arbiterfailover | — |
● runs here · ◐ light role (local sensor connections) · — runs on the nearest AgentMesh pod instead.
Qualcomm targets rugged handhelds and tablets, industrial scanners, industrial PCs and operator panels with this processor.
Autonomous mobile robots, companion robots and site-inspection drones are listed robotics targets.
Third-party inspection robots built on QCS6490 use multimodal sensing to flag anomalies and support predictive maintenance in factories and hazardous sites.
Part of a product-longevity programme with extended software support, and runs Linux, Ubuntu, Android and Windows IoT.
Public announcements and vendor material, November 2026 · vendor-reported, not Unovie measurements
Turnkey Edge-AI — fixed time, fixed cost, full responsibility.
