03 Intelligence
All your plant data, for everyone, in one sentence
A plant AI operations copilot that runs on top of the IIoT platform. Beyond a chatbot, it drafts prioritized task suggestions every morning, detects anomalies and proposes root causes and actions, and — for any natural-language question — orchestrates live queries, forecasting, and manual search to return a full report in 10 seconds. With a proprietary industrial LLM (AURA) deployed on-premises, your data never leaves the plant.
10 sec
NL query to full report
A Capabilities · In depth
The brain of PlantPulse AI is AURA, an industrial LLM we developed in-house on open-source foundations. A 35B MoE architecture with a 262K long context and vision-capable multimodality reads equipment manuals and chart images alike. It is fine-tuned on industrial standards — occupational-safety SOPs, ISO 9001 quality procedures, OSHA/NIOSH safety practices — and runs on-premises on a single GPU such as GB10 (DGX Spark-class) or RTX 6000.
Ask "analyze the injector" and the built-in AI agent decides for itself to query live sensors, run anomaly detection and forecasting, map equipment relationships, and search manuals — then combines it all into one report. The analysis streams in real time, and results are visualized as tables, charts, and diagrams. Multimodal questions with attached images are supported.
Hourly automatic anomaly detection and real-time alarms are promoted into a single incident, and the AI diagnoses the cause and generates recommended actions automatically. Incidents are tracked through acknowledgment and closure, with duplicates deduplicated. Alerts reach the right person over three channels.
The AI analyzes plant status daily and automatically drafts prioritized task suggestions assigned to the right people. Weekly briefings, next-day forecasts, and correlation analysis across suggestions help even non-experts see what to look at first.
Time-series analytics run on a dedicated insight server. Industrial time-series foundation models are served on your own GPUs (with automatic CPU fallback), covering anomaly detection across 9 domains and forecasting across 8 — sensors, OEE, energy, alarms, and work orders. Anomaly scores combine time- and frequency-domain reconstruction, and forecasts ship with conformal-prediction confidence bands. A feature store auto-builds hourly and daily features per tag and asset to power similar-asset search and LLM context, while five alarm diagnostics — frequency, duration, sensor correlation, failure prediction, and statistics — narrow down root causes.
Asset documents like equipment manuals and SOPs are searched with knowledge-graph RAG. Four search modes combine vector search with entity and relation graphs to find evidence that matches the question, and every answer cites which document of which asset it came from. Documents registered on the platform are kept in sync by a 10-minute incremental indexing cycle that reconciles additions, changes, and deletions by itself — and the search engine is isolated on the internal network, so documents never leave your plant.
B Architecture
AI operations copilot
AURA LLM (on-prem)
Unified data tools
live & historical · 116
TimeSeries-Insight
anomaly & forecast · 27
Knowledge graph
asset relations · 14 nodes
Document search (RAG)
manuals · SOPs
One question flows through the built-in AI agent, which automatically combines 116 data tools, time-series analytics, a knowledge graph, and document search — and the on-premises AURA LLM generates a full report with citations.
C Use Cases
Outcome
30–50% less unplanned downtime (estimate)
Outcome
90% faster situation awareness (estimate)
Outcome
Tribal knowledge captured · 40–60% faster handover (estimate)
E More Capabilities
Ask with natural language and images; get answers as tables, charts, and diagrams. Watch the analysis stream in real time.
The AI automatically composes 116 tools covering every real-time and historical dataset on the platform — all read-only by design.
Industrial time-series AI models provide 27 analysis tools spanning 9 anomaly-detection and 8 forecasting domains.
Equipment, tags, and documents linked in a 14-node-type graph give answers real context.
Searches equipment manuals and SOPs, answering with citations. Four search modes supported.
The industrial LLM is served on your own GPUs. Sensitive process data never leaves your network.
F Technical Specifications
G Product Screens
D FAQ
No. It runs fully on-premises and works even in air-gapped environments with no internet. AURA LLM is served on your own GPUs, so process data never leaves your organization.
All 116 data tools are read-only by design. The AI only queries and analyzes — it never writes control values — so it doesn't interfere with existing operations.
AURA uses an MoE architecture and 4-bit quantization to run a large model on low-spec GPUs (NVIDIA DGX Spark-class). You can deploy without a large GPU cluster.
It's a Korean-native model trained on field terms like degradation, interlock, and trip, plus safety and health regulation context. The UI supports Korean/English switching.
It's an on-premises fixed-cost model, not cloud pay-as-you-go. There is no per-token billing as query volume grows.
A 30-minute demo walks you from ingest to AI. Check the fit for your plant with an expert, right away.
