ARCHITECTURE / LAYERS// 2026

Two thin layers
over every
system you run.

The architecture is deliberately simple: a semantic layer that makes each system machine-readable in your vocabulary, an agent-application layer that acts on it, and a cross-stack layer that lets agents reason end-to-end. Everything else is plumbing.

Signal path · sensor → action

SENSORHISTORIANSEMANTIC LAYERAGENTACTION
01Per-stack layers

What sits over
each system of record

SCADA / Historians

/ L3

tag→asset, quality scoring, unit normalization

Sensor Cleanse · Alarm Rationalizer · Asset Health

MES

/ L2

batch genealogy, operator ontology, loss taxonomy

OEE Analytics · Quality Deviation · Scheduling

ERP / CRM

/ L1

order-to-fulfil graph, margin model, customer map

Order Intelligence · Margin · Inventory

Cross-stack

/ L∞

Shared vocabulary across all three layers: asset ↔ batch ↔ order linkages, cross-system joins

Root Cause · Demand-to-Production · Compliance

02Semantic layer

What lives in
the context layer

The thin strip between your source systems and the agents. Invisible when it works, painful when it doesn't. Everything here belongs to you.

/ 01

Tag-to-asset graph

Every raw tag in your SCADA/historian mapped to the physical asset it represents. Assets link to MES work-centers and ERP equipment masters. One source of truth for what-is-what.

/ 02

Unit & time harmonization

Engineering units normalized, timestamps aligned, sampling rates reconciled. Agents downstream don't have to care that one line reports °F and another °C.

/ 03

SOP & document corpus

Procedures, manuals, change-control records, and ticket history indexed and retrievable in-context. Agents cite the exact paragraph they drew on.

/ 04

Ontology of the operation

Shifts, routes, batches, SKUs, lines, and sites — modeled as first-class entities so agents reason in your vocabulary, not the source system's.

/ 05

Incident & decision memory

Past investigations, CAPAs, and engineer notes kept searchable. New incidents arrive with their own precedent pack.

/ 06

Quality scoring per tag

Every incoming tag carries a live quality score. Agents know what they can trust and refuse to reason on poor-quality inputs.

03Deployment

Three modes,
same platform

/ 01Air-gap supported

On-premises

K8s or bare-metal, no internet egress required
Models, inference, and semantic layer all local
SSO via on-prem IdP (AD, FreeIPA)
Fit for OT-networked plants and regulated sites
/ 02Most common

Customer VPC

Deployed inside your AWS/Azure/GCP tenant
Data never crosses your account boundary
SSO via cloud IdP (Azure AD, Okta, Google)
Infra-as-code handover at end of engagement
/ 03Fastest to start

Managed SaaS

ContextWeaver-operated multi-tenant control plane
Per-customer encryption and tenancy isolation
Region-pinned storage + inference
Upgrade path to VPC or on-prem any time
04Integrations

Connectors you
won't have to build

OSIsoft PI
Aveva
Ignition
Wonderware
Rockwell FT
GE iFIX
Siemens WinCC
OPC-UA / DA
MQTT / Sparkplug B
SAP S/4HANA
Oracle ERP
MS Dynamics
NetSuite
Salesforce
ServiceNow
Kafka / Pulsar

Want an architecture walkthrough?

Share your current stack, source systems, and deployment constraints. We'll map the semantic layer, the first-wave agent pack, and the migration plan together.