Foceon / Platform / AI architecture

The physical-AI stack, step by step

Foceon closes a real-time loop on the shop floor. This is how perception, prediction, simulation, and action are engineered on accelerated compute so the loop runs where the glass is.

Latency budget across the control loop
The loop

Four stages, one budget

Each stage has a latency budget so the whole loop closes fast enough to steer a live process.

  • Perceivesensor fusion
  • Predictlearned models
  • Simulatetwin validation
  • Actbounded control

How each stage is built

The architecture is deliberate about where each computation runs.

Edge-fused surface state from raw metrology

Perceive at the edge

Metrology and vision are processed at the edge on Jetson and Holoscan, so raw data becomes surface state without a round trip.

  • Jetson and Holoscan at the sensor
  • DeepStream vision pipelines
  • Low-latency surface state
Predicted move validated in the twin

Predict and simulate on OVX

Learned models and the optics twin run on OVX and Omniverse, close enough to advise the next move in real time.

  • TensorRT inference
  • Omniverse twin on OVX
  • Triton and NIM serving
Design choices

Why it is engineered this way

Latency

Loop on the floor

The loop closes at the machine, not in a distant cloud, so it can steer a live process.

Locality

Data stays put

Perception and inference run on-prem, so proprietary optics data never leaves the fab.

Portability

Edge to cloud

The same stack scales from one Jetson to an OVX cluster without a rewrite.

Safety

Bounded action

Every action passes through interlocks and is logged against the twin.

The stack

What runs where

ARCHITECTURErepresentative
PerceiveJetson HoloscanAt the sensor
PredictTensorRT TritonInference
SimulateOmniverse OVXTwin
OptimizecuOptScheduling
Perception

Metrology becomes surface state

Interferograms and profilometry are fused at the edge into a single surface state the rest of the stack reasons about.

Fused surface state from multiple instruments
Serving

From model to machine

  1. Train with NeMo

    Models are trained and adapted with NeMo on your process data.

  2. Serve with Triton and NIM

    Inference is served at the edge with Triton and NIM microservices.

  3. Simulate with Omniverse

    The twin validates the move on OVX before it acts.

  4. Schedule with cuOpt

    Line-level decisions are optimized with cuOpt.

Language

Stack terms

Holoscan
NVIDIA's sensor-processing platform for low-latency edge pipelines.
TensorRT
An inference optimizer that accelerates trained models.
Omniverse
A simulation platform used here to run the optics digital twin.
cuOpt
An accelerated solver for routing and scheduling optimization.
Comparison

Where the loop closes

Foceon architecture
PropertyCloud-only AIFoceon
Loop locationRemoteOn the floor
Data localityLeaves siteStays on-prem
Real-time controlNoYes
Edge-to-cloud scalePartialYes
Built on

The full NVIDIA stack

Foceon is engineered end to end on accelerated compute.

JetsonHoloscanDeepStreamTensorRTOmniverseOVXTritonNIMNeMocuOptCosmosReplicator
Questions

Architecture questions

Do we need NVIDIA hardware?

Foceon is engineered for it and deploys on edge and cluster configurations we scope with you.

Can it run air-gapped?

Yes. The full loop runs on-prem without external connectivity.

How is it updated?

Models and services update through controlled, auditable releases.

Put the loop on your floor

Book a pilot and we will scope the stack to your cells and metrology.