Porcelio
Book a pilot
Architecture

Physical AI, built for the plant floor

Porcelio is an industrial physical-AI platform: GPU-accelerated vision, time-series models, digital twins, and bounded agents that reduce fired defects, tone and caliber drift, breakage, and kiln energy, where GPUs are required, not optional.

The stack

GPU-essential, end to end

Every layer of the loop is a GPU workload, perception at line speed, simulation of firing physics, model serving with cited reasoning, and large-scale optimization.

Perception

Jetson edge nodes run DeepStream and TensorRT for green and fired inspection at line speed.

Simulation

Omniverse simulates firing, shrinkage, glaze, and planarity before a batch is burned.

Serving

Triton, NIM, and NeMo serve plant models and cited process reasoning.

Optimization

cuOpt and RAPIDS optimize kiln energy, sorting, and schedules at scale.

Robotics

Isaac drives robot-cell inference for glazing, sorting, and packing.

Grounding

Bounded agents act within limits, citing telemetry and prior outcomes.

Perception

Line-speed vision at the edge

High-speed multi-camera and profile sensing runs on Jetson Orin with DeepStream and TensorRT, grading every green and fired piece at line speed. Inference happens on the plant floor, so grading keeps working independent of the network.

Jetson OrinDeepStreamTensorRTMulti-cameraProfile sensing
TONE & CALIBER GRADINGVITREO
V0 · on-toneV1V2 · downgrade
Caliber deviation+0.18 mm
Planarity0.4 mm/m
Fired-defect risk1.2%
Pieces graded / min420
Digital twin

Firing physics, simulated before the burn

Omniverse-class simulation models the firing curve, shrinkage, vitrification, tone, and glaze behavior for a specific recipe on DGX/HGX plus RTX, so the plant can hit spec before committing a single piece to the kiln.

  • Physics-informedModels shrinkage and vitrification, not just temperature.
  • Recipe-specificCalibrated per body-and-glaze recipe from real outcomes.
  • Closed-loopTrained on Vitreo labels; feeds Thermiq and Glazio.
Kiln 3 · firing curve ● LIVE
1300 600 preheat soak cool
1284°PEAK °C
+0.2%CALIBER Δ
V0TONE GRADE
Agents & serving

Bounded agents with cited reasoning

Triton, NIM, and NeMo serve plant models and process reasoning that cites its sources. Agents are bounded, they act within limits your engineers define, and every decision is grounded in telemetry, simulation, and prior fired outcomes.

TritonNIMNeMoBounded agentsCited reasoning
SERVING & AGENTSSTACK
Model servingTriton
MicroservicesNIM
ReasoningNeMo, cited
Agentsbounded
Groundingtelemetry + twin
Why NVIDIA

Aligned with Physical AI priorities

Physical AI GPUs required for the core workload, not acceleration
Digital twin Omniverse simulation of firing and shrinkage
Edge inference Jetson at line speed on the plant floor
Agentic AI bounded agents with cited process reasoning
Data & integration

One data lake, one model router

The five products share one plant-edge runtime, one firing-and-quality data lake, and one model router, so data compounds instead of fragmenting across point tools.

Shared data lake

Firing and quality data from every line feeds one compounding memory.

One model router

Models are served and routed centrally across products.

Deep integration

Reads telemetry from presses, dryers, glaze lines, and kilns.

Questions

Answers for plant leaders

Why are GPUs essential, not optional?
The core workloads, line-speed vision, firing-physics simulation, and large-scale optimization, are GPU-bound. Porcelio is a physical-AI platform where GPUs are required to run the loop, not just to speed it up.
How is process reasoning kept trustworthy?
Reasoning is served with citations to the telemetry, twin simulations, and prior outcomes behind it, and agents act within bounds your engineers define.
Does data stay on-prem?
Inspection and robot inference run plant-edge. Deployment options for the data lake and simulation are scoped per customer, including on-prem and hybrid.
Reads your plant

Instruments the line you already run

Porcelio ingests telemetry from presses, dryers, glaze lines, and kilns, and runs alongside your existing controls.

Kiln & press OEMs

Roller kilnsTunnel kilnsHydraulic pressesExtruders

Glaze & decoration

Glaze linesDigital decorationEngobe applicationRobotic cells

Inspection & data

Profile scannersSCADA / PLCMESHistorian

Ceramics groups

Multi-siteR&D labsSustainabilityQuality
The workload
GPU-bound

by design, not by choice

Line-speed vision, firing-physics simulation, and plant-scale optimization are all GPU workloads. This is physical AI, where GPUs run the loop.

Go deeper on the architecture

We'll walk your technical team through the stack, the twin, and how bounded agents stay grounded and auditable.