AURA LLM (Industry-Specific Model)
Overview
AURA is an industrial-purpose large language model (LLM) developed in-house by PlantPulse.
Built on top of the best-performing open-source models and further trained (fine-tuned) on manufacturing/industrial domain data, it accurately understands the language, context, and technical nuances of the factory field.
Available Models
| Model | Purpose | Characteristics |
|---|---|---|
| AURA LLM | Conversation/analysis/report generation | Industrial fine-tuning, multimodal |
| AURA Embedding | Vector embedding | Optimized for RAG retrieval |
| AURA Reranker | Reranking of search results | Precise relevance evaluation |
Fine-Tuning Data
Official Standards Body Data
| Source | Content |
|---|---|
| KATS (Korea Occupational Safety and Health Agency) | Mandatory SOP guides based on the Occupational Safety and Health Act |
| ISO/IEC | Quality management SOP templates based on international standards (ISO 9001, etc.) |
| OSHA/NIOSH | US industrial safety standards, industry-specific safety procedures |
Field-Validated Data
| Data Type | Content | Effect on AI Capability |
|---|---|---|
| Regulatory/safety documents | Internal safety manuals, MSDS, legal precedents | Improved accuracy in regulatory compliance and hazard response |
| Q&A/discussion data | Q&A among field technicians, in-house forums, incident reports | Better contextual understanding and ability to propose concrete solutions |
| Unstructured logs | Equipment failure records, maintenance logs, operator notes | Stronger predictive maintenance and troubleshooting inference |
Multimodal Support
- Text: Natural-language conversation, technical document comprehension, report generation
- Images: Equipment photo analysis, instrument panel reading, drawing interpretation
- Tables/charts: Data table analysis, chart interpretation
AURA vs. General-Purpose LLMs
| Item | General-purpose LLM (GPT, Claude, etc.) | AURA |
|---|---|---|
| Industrial terminology comprehension | General level | Manufacturing-expert level |
| SOP/regulation interpretation | Lacks context | Interpretation grounded in legal basis |
| Data location | Cloud (sent externally) | On-premises (kept internal) |
| Cost structure | Per-token billing | Fixed cost, unlimited use |
| GPU requirements | Large GPU or API | Runs on low-spec hardware such as DGX Spark |
| Multimodal | Supported | Supported (specialized for industrial imagery) |
| Embedding/reranker | Requires separate models | Provided as an integrated package |
| Customization | Limited | Additional training on field data possible |
Technical Characteristics
Operation on Low-Spec GPUs
AURA can run on low-spec GPU environments such as NVIDIA DGX Spark. It enables on-premises deployment without large cloud GPUs, and is lightweight enough to run on edge servers in the factory field.
Full On-Premises Support
By eliminating dependency on external cloud APIs, sensitive factory data never leaves the site. AI services continue to operate even when the network is disconnected.
Native Korean
AURA accurately understands the terminology used in Korean manufacturing plants. It interprets field terms such as "degradation," "interlock," and "trip" according to context, and understands the context of Korean occupational safety and health regulations.
Use Cases
| Scenario | How AURA Is Used |
|---|---|
| Copilot conversation | Core conversation engine for Copilot Ask (chat) |
| Document search optimization | Improves RAG retrieval accuracy via AURA Embedding |
| Anomaly cause interpretation | Infers root causes from sensor data plus industrial knowledge |
| Automatic report generation | Automatically writes daily/per-shift operation reports |
| Safety regulation checks | Instant reference to KATS/ISO/OSHA-based regulations |
| New-hire training support | Expert-level answers to field questions |