Perception at line speed
Metropolis / DeepStream + TensorRT and Holoscan drive high-speed print-defect, color, register, structural/crush, warp and glue/fold vision.
Technology
A plant-edge autonomy stack, grounded and auditable
Perception models, process agents, bounded control policies, a physics-informed twin and a closed-loop data flywheel, built for the physics and brands that converting lives and dies on.
Agent architecture
Seven agents under one plant orchestrator
Fine-tuned models
Fine-tuned defect/register/color, structural/warp and fold-quality vision plus corrugator and waste/OEE optimization models at the edge; time-series prognostics for machines.
Reasoning where it pays
A model router serves the best/cheapest model per step; high-volume steps use fine-tuned open models, premium root-cause reasoning uses frontier models.
Bounded control
Process agents issue setpoint moves within approved limits, with human-in-the-loop checkpoints and graduated autonomy.
Accelerated computing
NVIDIA physical-AI stack, edge to twin
Foldeon runs a closed perceive-decide-act loop across the converting plant, from Jetson at the edge to Omniverse for the twin.
- Isaac + Jetson for robotic feeding, stacking, palletizing and changeover
- Metropolis / DeepStream + TensorRT + Holoscan for high-speed vision
- Triton + NIM serving; NeMo-fine-tuned reasoning models
- cuOpt for scheduling, nesting, makeready and palletizing
Edge cell · JetsonAUTONOMOUS
Defect p95<50 ms
Register advisory<20 ms
Camera streams4–12
ServingTriton + NIM
Loop latencyon target
The digital twin
A converting-line-and-box twin that runs the job before you do
An Omniverse twin simulates corrugation, print and die/fold to hit target quality and yield before the run, auto-optimizing board, print and fold and validating every change first.
- Physics-informed twin of corrugation, print, die/fold and handling
- Pre-run job simulation and makeready auto-tuning
- Cosmos / Omniverse Replicator synthesize rare defect scenarios
- Isaac Sim/Lab validates robot cells before deployment
Perceive → decide → act
The closed AI workflow
Perceive → plan → run → sense/predict → optimize → flag → log → retrain.
-
01
Perceive
Sense board, sheet and box across every stage with fused vision and sensors.
-
02
Plan
Plan corrugate/bond, print/color, die-cut/crease and fold/glue for the job.
-
03
Run adaptively
Execute with adaptive, bounded closed-loop control.
-
04
Sense & predict
Predict defects, color, structural strength, warp and fold quality.
-
05
Optimize & flag
Optimize waste, makeready and OEE; flag defect and claim risk.
-
06
Log & retrain
Log immutably, then retrain the fleet on production and synthetic data.
The stack
What powers each layer
| Capability | Technology |
|---|---|
| Edge vision & fusion | Jetson Orin / Thor · TensorRT · Holoscan · Metropolis / DeepStream |
| Serving | Triton + NIM at plant edge and private cloud |
| Reasoning & RAG | NeMo + Retriever + Guardrails · pgvector · citations enforced |
| Training | DGX / HGX · 10k–40k GPU-hours per major release |
| Simulation | Omniverse / OVX twin · Cosmos & Replicator synthetic data · Isaac Sim/Lab |
| Optimization | cuOpt: sequencing, nesting, makeready order, energy, pallet patterns |
The stance
Trust, safety and evaluation, built in
Grounded outputs are non-negotiable on a plant floor.
Grounded & cited
Citation and grounding checks on every output; RAG enforces traceable sources.
Human-in-the-loop
Graduated autonomy with approval checkpoints for high-risk changes.
Continuously evaluated
Golden datasets + LLM-as-judge gate every model and prompt change in CI.
Isolated & auditable
Per-tenant isolation, strict IP/artwork/recipe protection, immutable audit log.
<50msDefect decision p95
<20msRegister advisory p95
1MSynthetic scenarios / qtr
4–12Camera streams / line
Memory & the flywheel
Per-plant history and per-craftsman memory
Foldeon captures the scarce craft of press operators, corrugator crews and die-makers as versioned, tenant-scoped memory that improves over time without cross-tenant leakage.
- Per-plant converting-quality-and-waste history
- Per-operator/die-maker performance memory
- pgvector defect/color-image and job/recipe retrieval
- Time-series store for corrugator, press and gluer telemetry
The router and the twin are the two ideas that make this economical: cheap models where volume is high, and a simulation that pays for makeready before the run.
Answers
Questions converters ask
Can we run fully on-prem?
Yes. Optional on-prem deployment is available for sensitive converters, with per-tenant isolation and strict IP, artwork and recipe protection throughout.
How do you control inference cost?
Models are abstracted behind a router so high-volume steps move to fine-tuned open models to control COGS, while premium reasoning steps use frontier models only where they add value.
How do you handle rare defects with little data?
Cosmos and Omniverse Replicator synthesize 100k–1M rare defect, material, color, fold and handling scenarios per quarter to augment production data.
Go deep with our engineers
Bring your line topology and machine list. We’ll map the edge, twin and integration plan.