Does any plant data leave the site?

No. TruAdaptiv Copilot runs inference on the plant LAN using open-weight models hosted on your own hardware. There is no API call to an external provider, which is what makes it viable in plants where sending process data to a cloud service is not permitted.

Which models does it use?

Open-weight models, adapted to the plant with a LoRA adapter trained on that site's own vocabulary, equipment names and process context. The base model can be updated without discarding the plant-specific adaptation.

What does "process, predict, prescribe" mean in practice?

Process is making plant data answerable in plain language. Predict is flagging where a trend is heading before it breaches a limit. Prescribe is recommending the specific action a shift team should take, grounded in that plant's own procedures rather than generic advice.

Does it need IndustrialBridge and SyncPlant first?

Copilot is Tier 03 of the TA Factory stack and works best on the data foundation the lower tiers provide. It is not a hard dependency, but an AI layer over data that is inconsistent or unreachable produces confident answers to the wrong question — the data layer is what makes it trustworthy.

Can it run in an air-gapped plant?

Yes. Because inference is local and there is no external service dependency, it deploys in fully air-gapped environments. Model and adapter updates are applied through your existing controlled-media process.

What hardware does it need?

An on-premise server sized to the deployment — the requirement depends on model size, how many concurrent users you expect and response-time targets. This is scoped during the pilot rather than quoted generically, because over-specifying GPU hardware is one of the easiest ways to make an on-prem AI project uneconomic.

TA-FACT · TIER 03 · INTELLIGENCE

The plant-floor AI co-pilot

TruAdaptiv Copilot
Your plant gets smarter every day it runs.

The first plant-floor system where the data, the model, and the decision live in the same box on the plant LAN. A tightly-bounded embedded AI co-pilot that does three things — Process, Predict, Prescribe — without anything leaving your building.

Request a demo Back to TA Factory

TRUADAPTIV COPILOT · PLANT LAN ON-PREM · AIR-GAPPED

Mill 2 bearing temp keeps creeping up. What should I do?

Bearing temp on Mill 2 has risen 4.2°C in the last 28 minutes. Three similar drifts in the last 18 months — each resolved by load reduction within 45 min. Two candidate actions:

Reduce Mill 2 load by 15% for 30 min87% success

Switch to feed bin 4 (cooler ambient)62% success

Sources: maint_log#4471 · tag_history.mill2_bearing · shift_entry#8821

Ask in English or Hindi…

On plant LANNever leaves your perimeter

EN + HINatural-language Q&A

4–48 h aheadPredictive alerts before failure

You own itPlant-tuned LoRA adapter

The tier nobody else offers

Every other vendor sells acquisition and storage

Ignition, AVEVA PI, Wonderware, the Indian challengers — they sell acquisition and storage. Decisions still happen in human heads, hours after the data was collected, with no memory of how similar situations resolved last time. TruAdaptiv Copilot closes that loop — without anything leaving your building.

PROCESS

Explain what just happened

An alarm fires. The operator taps it. The Copilot reads the last 30 minutes of tags, the equipment's maintenance history, recent logbook entries — and produces a plain-language explanation in English or Hindi, with every claim cited.

Plus · Fill the form for me.Handovers, permit-to-work, deviation reports, daily energy logs, LOTO checklists — pre-filled from the historian, KPI engine, alarm log, and CMMS. 70% of typing eliminated. 100% of audit-trail rigour preserved.

PREDICT

What to watch

Per-equipment forecasting models predict tag values 5/15/30/60 minutes ahead. When a leading indicator drifts outside its envelope, you get a predictive alert — not an alarm.

Predictive alert · example"Over the next 30 minutes, Mill 2 bearing temperature is likely to exceed 85°C unless load is reduced." Time-to-impact and confidence shown up front.

PRESCRIBE

Help me decide

The operator asks "what should I do?" The Copilot retrieves similar historical situations from your plant's own past, ranks them by how each one resolved, and offers 2–3 candidate actions with historical success rates.

It gets better over time.The operator picks one (or none); the outcome is logged and feeds the next recommendation. The system genuinely improves the longer it runs in your specific plant.

How it runs

Open-weight models on a fanless industrial PC — owned by you

The Copilot runs on a fanless industrial PC on the plant LAN. Open-weight models — Gemma, Phi, Qwen class — fine-tuned to your specific plant via a 30 MB LoRA adapter that encodes your equipment naming, your failure modes, your operators' question patterns. The adapter is generated from data that never leaves the plant. You own it. We operate it.

01 · HARDWARE

Fanless industrial PC

Runs on cabinet-mounted industrial hardware. No GPU farm, no cloud round-trip, no internet dependency during operation. Sits on the plant LAN alongside SyncPlant and IndustrialBridge.

02 · MODELS

Open-weight base models

Gemma, Phi, Qwen-class small open models running locally. No proprietary cloud API. Models can be audited, replaced, upgraded — you're not locked into anyone's stack.

03 · ADAPTER

30 MB plant-tuned LoRA

Your plant's equipment, failure modes and question patterns encoded in a small LoRA adapter. Generated from data that never leaves the plant. You own the artifact — we operate the engine.

04 · LANGUAGE

English & Hindi natively

Operators ask questions in English, Hindi, or code-mixed speech. The Copilot responds in the language asked. Designed for the realities of Indian plant floors.

05 · RETRIEVAL

Grounded in your plant's records

Every answer is built from real retrieved records — tags from the historian, entries from the logbook, work orders from CMMS, similar past situations. Not invented from training data.

06 · LEARNING

Outcomes feed back

Every recommendation, decision and outcome is logged. The next "what should I do" question is answered using everything the plant did before. The longer it runs, the better it gets.

Engineering stance

Hallucinations are an engineering problem, not magic

A plant-floor AI that occasionally invents facts is worse than no AI at all. TruAdaptiv Copilot is engineered with hard guardrails around what it can say and how it can say it.

Citation guard

Every numeric value and entity in a Copilot response must be cited to a retrieved record. Uncited claims are rejected before they reach the operator.

Strict JSON schema

The model fills a strict JSON schema for every response — never free-forms. If it can't fill the schema cleanly, it refuses.

"I don't know" is first-class

The Copilot is rewarded for saying "I don't know" when retrieval comes up empty. Better an honest refusal than a confident invention.

Never autonomously acts

The Copilot proposes — the human disposes. No autonomous setpoint changes, no automated control loop, no machine actions. Always human-in-the-loop.

Deployment

Built to receive the work you did in Tiers 1 and 2

TruAdaptiv Copilot is the natural destination of the TA Factory journey. Every artifact built in IndustrialBridge and SyncPlant becomes context for the Copilot — historian tags, KPI definitions, alarm thresholds, equipment hierarchy, logbook entries, work orders.

On-prem

Plant LAN

Lives on a fanless industrial PC on the plant LAN. No internet egress during operation. Air-gapped option available for sensitive sites.

Owned

Your LoRA

The plant-tuned LoRA adapter is your intellectual property, generated from data that never left the plant. You own the artifact.

Integrated

Tier 1 + 2 ready

Consumes IndustrialBridge tag streams and SyncPlant operational records directly. No new integration layer needed if those are in place.

Audited

Every answer cited

Every Copilot response includes citations to the underlying retrieved records. Forensic-grade traceability for regulated environments.

The decisions, not just the dashboards

TruAdaptiv Copilot is the AI tier of TA Factory — for plants that have outgrown dashboards and want decision support that stays on the LAN. Book a discovery call.

Request a demo ← Back to Tier 1 · IndustrialBridge

Questions we get asked

Common questions

Does any plant data leave the site?: No. TruAdaptiv Copilot runs inference on the plant LAN using open-weight models hosted on your own hardware. There is no API call to an external provider, which is what makes it viable in plants where sending process data to a cloud service is not permitted.
Which models does it use?: Open-weight models, adapted to the plant with a LoRA adapter trained on that site's own vocabulary, equipment names and process context. The base model can be updated without discarding the plant-specific adaptation.
What does "process, predict, prescribe" mean in practice?: Process is making plant data answerable in plain language. Predict is flagging where a trend is heading before it breaches a limit. Prescribe is recommending the specific action a shift team should take, grounded in that plant's own procedures rather than generic advice.
Does it need IndustrialBridge and SyncPlant first?: Copilot is Tier 03 of the TA Factory stack and works best on the data foundation the lower tiers provide. It is not a hard dependency, but an AI layer over data that is inconsistent or unreachable produces confident answers to the wrong question — the data layer is what makes it trustworthy.
Can it run in an air-gapped plant?: Yes. Because inference is local and there is no external service dependency, it deploys in fully air-gapped environments. Model and adapter updates are applied through your existing controlled-media process.
What hardware does it need?: An on-premise server sized to the deployment — the requirement depends on model size, how many concurrent users you expect and response-time targets. This is scoped during the pilot rather than quoted generically, because over-specifying GPU hardware is one of the easiest ways to make an on-prem AI project uneconomic.