Jetson Orin / Thor
On-line and on-robot real-time perception and control at the edge.
The physical-AI stack
Edge to cloud, built on NVIDIA
Circuex runs perception at the line, training in the data center, and simulation in Omniverse — a full physical-AI stack for the factory floor.
Design principle
Inference where the boards are
Drift is caught in 20–50 ms per tile at the edge, so a bad deposit or profile is corrected before more boards fail.
Hardware
Right-sized silicon per job
IGX + Holoscan
Medical-grade edge for high-rate AOI/AXI and coating inference.
DGX / HGX
Training multimodal defect, thermal, and yield models over millions of samples.
OVX / RTX
Omniverse twins, recipe optimization, and engineer-facing visualization.
SDKs & libraries
The software that makes it move
TensorRT · DeepStream · Metropolis
High-rate, low-latency inspection vision at the edge.
Omniverse · Replicator · Cosmos
Digital twins and synthetic rare-defect generation.
Isaac Sim · Isaac Lab · Isaac ROS
Robot simulation, skill learning, and deployment for Roboq.
cuOpt · RAPIDS · NeMo
Line balancing, telemetry analytics, and IPC-grounded reasoning.
MODEL SERVING
20–50msLatency
INT8TensorRT
LiveOn line
The twin
Physics-informed, factory-calibrated
Modulus physics models and learned simulation predict deposition, thermal reflow, and joint formation — calibrated to each factory's real outcomes.
Explore Twinex
Data & the loop
Why it compounds
Every action writes to a shared closed-loop corpus — the moat that grows with every board.
01
Ingest
Multimodal telemetry: SPI, AOI, AXI, coating, oven, and test.
02
Ground
Every call cited to IPC standards and process context via NeMo.
03
Learn
Models retrain on outcomes; rare defects synthesized in Cosmos/Replicator.
04
Serve
Updated models pushed to the edge under version control.
0Edge inference / tile
0Cameras per line
0Agents on one stack
0Actions logged
They didn't hand us a model — they handed us a stack that trains, serves, simulates, and closes the loop on our floor.
Do you run in our data center or the cloud?
Both. Edge inference runs on-line; training and twins run where you prefer — on-prem DGX/HGX/OVX or cloud.
How do you keep models honest?
Continuous calibration to real outcomes, confidence scoring, human-in-the-loop on low-confidence cases, and full audit logging.
Is our data used to train other customers' models?
No. Your closed-loop corpus is yours. See our security and trust page for details.
Every board, perfectly built.
Go deep with our engineers
Book a technical session on architecture, edge deployment, and the twin.