One model of your plant, served over i3X. PLC tags, MES events and vision detections become a single governed model — so your systems, your people and your physical AI act on the same meaning.
Tag names only a controls engineer can decode. Historian trees that disagree with the MES. Every project re-derives the same plant context by hand. Edge makes that context a governed, queryable asset — kept current from the buses you already run.
Assets, signals, materials and events modeled once — versioned and governed like code. Ask it questions in plain language, or wire systems to it; every consumer sees the same plant.
Plant structure, production, material genealogy, quality and reliability served over the open CESMII i3X standard. Publish once, consume everywhere — no custom integration per system.
Robots, AGVs and people pick up work described by the capability it needs — torque, transport, inspect — not the vendor that supplies it. Factory process logic stays in the model, out of any one fleet system.
Detection alone doesn't fix anything. Praxis raises countermeasures on the same model, dispatches them, and records whether the fix worked.
Twelve steps, one thread: context, observations, a named situation, countermeasures, task attempts, an evaluated closure. Payloads illustrative; links explicit.
i3X evidence → situation → countermeasures → evaluated closure
Downtime is charged to the machine that stopped. The plant pays for the loss that propagated. Praxis reconstructs each event across machines, buffers and logistics, prices the propagated loss, and ranks countermeasures by counterfactual — simulated before anyone acts.
| Countermeasure, ranked | Loss avoided / event | Cost | Payback |
|---|---|---|---|
| Tugger schedule −4 min | 11 min | ~$0 | Immediate |
| +2 buffer positions | 14 min | $8k | 6 weeks |
| Additional tugger | 17 min | $90k | 14 months |
Each row is a counterfactual run against the operational twin, not an estimate. Illustrative planning figures — a pilot replaces them with measured ones.
Three kinds of number, kept provably separate. Every dollar traces back to its events and its assumptions.
PLC, robot, conveyor, tugger and andon signals over i3X — raw, not pre-typed.
Typed events — cycle, blocked, starved, faulted, changeover, KPI drift — each citing its signal.
A live object per disruption: scope, cascade, propagated loss — not the stopped machine's downtime.
Ranked by counterfactual, dispatched by capability, closed with computed effectiveness.
PLC, MES, historian and logistics events reconstruct the loss and price it. No 3D required.
Queues, buffers, timing, routing, failure distributions — the counterfactual engine under the ranking.
Omniverse / Isaac for robot reach, layout, AMR traffic and synthetic vision. 3D is an engine, not the product.
Pilot metric: recovered production minutes per shift — not model accuracy, not twin fidelity.
One named problem; three takers. A robot cell, an AGV and a technician each pick up the piece that matches their capability — and the closure is recorded, whoever did the work.
detect → name → dispatch by capability → record
Diagnosis without a war room. The assistant reads live signals, walks controller logic upstream, pulls event history and Praxis countermeasures — and shows the reads behind every answer. Illustrative exchange; no live connection.
false · PART AT FIXTURE = true · TOOL TT-47-2 READY = falseTT-47-2 during WO-8841, both on units from LOT-4471-B. That matches the lot-variance situation Praxis raised this morning — countermeasure verify-part is already queued at gate QG-3. Want the tool recalibration raised as well?
Sized by model class and concurrency, not by the model file — KV cache grows with every user and every token of context. We size from lab telemetry before a production bill of materials exists.
Machine vision, cycle classification, anomaly detection, sensor inference, local buffering. Sub-second, beside the equipment.
Edge Chat, RAG over manuals and drawings, root-cause and starvation agents, semantic layer and knowledge graph, MES/MOM and twin queries — multi-user. Dual 10 GbE, 2–4 TB NVMe, TPM 2.0, secure boot.
Large reasoning models, enterprise search, fleet learning, training, long-horizon analytics. Only embeddings, events and curated clips leave the plant.
Simultaneous plant users, 8B–70B target, and whether line vision shares the box.
The same stack on two modest machines; CPU, RAM, GPU, IOPS and latency measured under load.
Specified from telemetry, not guesses. Never from the fleet plane's minimums.
Three concrete configurations per line or site — costed in a conversation, not on a web page.
Assets, signals, materials and events modeled once — versioned and governed like code.
Every system and agent consumes the same context through the open standard.
Work goes to whoever has the capability — robot, AGV or person, from any vendor.
Countermeasures run to an evaluated outcome. The plant learns from what it did.
Bring one line and one hard question. We'll show you how VitalFusion Edge would model it, and what your team could ask of it.