AURA Models
AURA is the family of industrial AI models built in-house by KOPENS. It is the layer that actually does the thinking when PlantPulse AI produces an answer — the part that reads equipment data, calls tools, and responds in the language of the field.
General-purpose models know a great deal about the world but know nothing about the compressor in your plant. They will happily explain what a bearing is, and just as happily invent a vibration value that was never measured. In a plant, it is the second behavior that is dangerous. AURA was trained for the opposite — answer from evidence, and say "not measured" when something was not measured.
Three models in the language track
| Model | Size | What it does | Current status |
|---|---|---|---|
| AURA-Lite | approx. 26B (MoE, approx. 4B active · 256K context · image input supported) | Diagnosis · calculation · tool calling · report writing | Running in production |
| AURA-Embedding-4B | 4.0B | Retrieval across procedures, standards, and failure history | Running in production |
| AURA-Reranker-0.6B | 0.6B | Selecting the retrieved items that actually answer the question | Running in production |
The three work together. The embedding model casts a wide net across the documents, the reranker picks out the passages that genuinely answer the question, and AURA-Lite reads them, calls whatever tools are needed, and writes the answer.
What the gateway exposes is not the base model but the version trained by KOPENS. That said, performance figures are not published yet — these models have only just entered the product path, and the numbers will be re-measured on that path before being released in a revised technical white paper. Not quoting numbers that have not been measured is a rule of this product.
Waveforms are treated as a «different kind of thing» — the time series track (in research)
Language models are strong at explaining why something happened, but finding what is abnormal in a waveform arriving every second is a different kind of problem. So the waveform side is handled by a separate family at the million-parameter scale — small enough to run without a GPU, so it can sit on the Edge and be tuned per site at almost no cost.
It is not yet on the product path. It will not go in until its value is proven.
There are cases where not calling a model is the better choice. In some situations, simply looking at the statistical deviation from the equipment's normal baseline was enough to separate the abnormal intervals. Those intervals are not solved with a model — distinguishing where a model is needed from where statistics suffice comes before "make the big model bigger."
What makes an AI «industrial»
It calculates — it does not estimate. Bearing defect frequencies (BPFO/BPFI), L10 life, NPSH margin, ASME shell thickness, orifice flow, OEE, process capability, welding heat input. These values come from formulas. And scoring is done against those same formulas — not against the opinion of another model.
It distinguishes 0 from «unknown». "0 work orders this week" is a measurement. "The CMMS connection dropped" is not. Reporting the second as if it were the first makes operators believe that equipment nobody has looked at is "normal." When a value was not collected, AURA says so plainly.
It reasons by failure mechanism, not by keyword. There are at least four reasons a temperature rises — bearing wear, insufficient lubrication, cooling failure, electrical fault. The four are distinguished by what moves along with it. Did vibration rise before temperature? Did flow drop first? Did current rise at the same time? AURA learns that structure, not "symptom → answer" pairs.
It stays in its lane. Equipment, process, quality, safety. It defaults to Korean and answers in English on request, and keeps politics and diplomacy out of scope. This is not censorship — it is the scope a plant assistant should have.
How it is deployed
AURA models are delivered as quantized checkpoints (NVFP4) and run behind the same OpenAI-compatible endpoint used by the rest of the platform. The retrieval models keep their embedding dimensions stable across versions, so swapping a model does not require rebuilding the vector index.
Nothing leaves the premises. The models run on the customer's hardware, against the customer's data.
How it is measured
There is no leaderboard for industrial AI, so KOPENS built its own instrumentation. Scoring is not delegated to another model; answers are checked mechanically against formulas, tool schemas, and operational records.
- Industrial calculation — 30 families of engineering formulas. Seven of them were excluded from training entirely. The point is to see whether the model can actually do it, not whether it memorized it.
- Tool calling — using the real tool schemas of the running platform: which tool was chosen, whether the arguments are correct, whether multi-step chains hold together.
- Staying grounded — whether the answer uses only what was actually present in the given context.
- Field diagnosis (158 questions) — based on six months of real compressor measurements and recorded failures. This is the only one of the five axes where the ground truth comes from outside our own work. It includes «hard normals» — high-load intervals that look abnormal but are not. False alarms are counted separately from misses, because an assistant that cries wolf gets switched off, and a switched-off assistant catches nothing.
Going deeper
There is an AURA technical white paper for technical reviewers (20 A4 pages, Korean/English). It covers how the training data was built, the recipe for teaching failure mechanisms, quantization, how the scoring rules are mechanized, the criteria used to select the field diagnosis questions, and how a reviewer can reproduce the same measurements.
Licensing
Deployment terms are negotiated with KOPENS on a per-contract basis. Licenses and notices for open source components are delivered together with the model artifacts.