Isaac + Jetson
Robotic ply handling, trimming, drilling, bonding and part-movement robots.
Physical AI, engineered for the factory edge
Curelum is GPU-essential: high-rate sensor fusion, large-scale training, digital-twin simulation and low-latency bounded control at the edge.
0
AI depth (Inception)
boosted score
0
Sensing modalities
0+
GPU workloads
Edge
Deployment
bounded control
Architecture
From connectors to autonomous agents
Agents
Layup, defect, cure, NDT, trim, handling and yield agents close the loop from perception to action.
Digital twin
Placement, cure kinetics, consolidation and strength, validated in Omniverse.
Perception & fusion
Vision, laser profilometry, ultrasonic and thermographic fused on the edge.
Factory-edge runtime
Jetson, Holoscan, TensorRT and Triton run bounded control at low latency.
Connectors
AFP / ATL, autoclave / OOA, NDT, MES via API, OPC-UA and OT.
NVIDIA stack
GPU-essential, end to end
Metropolis + Holoscan
In-process AFP / layup vision, porosity / wrinkle / FOD sensing, profilometry, NDT fusion.
Cosmos
Synthesizes rare porosity, wrinkle, FOD and cure-fault scenarios for training.
cuOpt
AFP tool-path / layup sequencing, autoclave-batch scheduling, cure-energy dispatch.
Omniverse
Composites twin: placement, cure kinetics, consolidation and strength.
Triton · TensorRT · NIM · NeMo
Low-latency serving, optimized inference and model management on the edge and in the core.
Perception & fusion
See the laminate as it’s built
High-rate machine vision and laser profilometry sense tow geometry, gaps, overlaps and wrinkles during placement; ultrasonic C-scan and thermography certify internal integrity after cure. Multimodal fusion links every indication to the ply and action that produced it.
- Vision + laser profilometry in-process
- Ultrasonic + thermographic NDT fusion
- Sub-defect-class prediction pre-cure
- Defect-to-cause traceback
Simulation
Physics-informed surrogates
GPU-trained surrogates for placement, compaction, cure kinetics, exotherm, residual stress and strength let the twin predict as-built, as-cured behavior — and auto-optimize recipes — orders of magnitude faster than build-and-test.
Learning
A compounding, federated corpus
Each record links sensor evidence, agent recommendation, engineer correction, machine action, cure trajectory, NDT result, disposition and final strength. That closed-loop action-response corpus improves models with every part — shared across sites in a privacy-, IP- and ITAR-preserving way.
Edge & control
Bounded, low-latency, safe
| Edge compute | Jetson / industrial GPU at the cell |
|---|---|
| Serving | Triton, TensorRT, NIM |
| Control | Bounded envelopes with exotherm / motion guards |
| Latency | Real-time sensor fusion & inference |
| Training | DGX / HGX, RAPIDS, NeMo in the core |
| Optimization | cuOpt scheduling & dispatch |
Security posture
Isolated, auditable, controlled
Curelum deploys in ITAR-ready, isolated configurations with SSO / RBAC and per-site data isolation. Every model change and machine action is captured in an immutable audit trail for airworthiness and qualification.
- ITAR-ready, isolated deployment
- SSO / RBAC access control
- Immutable per-action audit log
- Federated, IP-preserving learning
- Model-change evidence for qualification
Frontend & platform
The engineering under the hood
Edge runtime
Rust / C++ factory-edge runtime for deterministic, bounded control.
Control plane
Go + Python (FastAPI) services and a NestJS control plane.
Interfaces
Next.js, React and TypeScript review consoles and dashboards.
Go deep with our engineers
Architecture review, integration scoping and a security walkthrough for your stack.