Industrial AI · positioning
Industrial AI is six markets — one is wide open
Industrial AI isn’t one thing — it’s six distinct markets.
Predictive maintenance. Quality inspection. Process optimisation. Anomaly detection. Supply-chain forecasting. And the newest arrival: operations assistants.
The first five are mature. GE Digital, AspenTech, PTC, Siemens MindSphere — deeply entrenched, hard to displace.
The sixth is wide open.
What the plant actually wants
A chat you can ask “why is line 3 running hot?” and get the answer in seconds — grounded on the plant’s own telemetry, its maintenance manuals, and the last ten years of incident reports. On-prem. Audit-logged. No cloud exfiltration.
That’s the wedge. Existing dashboards tell you the temperature is 94 °C. They don’t tell you why. They don’t cite the SOP that says your next action is to bypass valve 7B. They don’t link the four prior occurrences to the root cause someone wrote up in a PDF in 2021. A chat interface that knows the plant does.
The six markets at a glance
| Use case | What it does | Incumbents |
|---|---|---|
| Predictive maintenance | Sensor data → “this bearing fails in 18 days” | GE Digital, Uptake |
| Quality inspection | Vision AI replaces line inspectors | Cognex, Landing AI |
| Process optimisation | Push setpoints that trade yield vs energy vs throughput | AspenTech, Imubit |
| Anomaly detection | “This run is not like the others” — precursor to failure / leak / breach | SparkCognition, C3 |
| Demand / supply-chain | Order → production → logistics optimisation | Blue Yonder, o9 |
| Operations assistant | LLM answers with citations, grounded on live + historical plant data | Wide open |
Why now
Three things converged in the last 18 months:
- Open-weight LLMs good enough to ground-and-answer on technical documents (Llama 3, Qwen 3, DeepSeek).
- Consumer-grade GPUs that run those models with enough throughput for a plant-sized workforce (an RTX 4090 serves ~40 concurrent technicians on a 14B model).
- Reasoning-based retrieval (vectorless RAG over structured documents) that beats embedding search on the kinds of content plants actually have — hierarchical manuals, standard operating procedures, incident write-ups.
The first two mean you can deploy this on-prem without a hyperscaler. The third means the retrieval quality is good enough for a plant engineer to trust. That’s the new thing.
How Eldric fits
Eldric AI OS is a single-binary AI operating system — 15 role modules in one process, dnf install eldric-aios on any Fedora 42+ or RHEL 9+ host. For this use case it provides:
Terminology. Every “Worker” below is a daemon that does the work — ingests sensor data, stores vectors, runs inference, delivers messages. In 4.x each Worker was its own systemd unit; in 5.0 they’re role modules inside the single eldric-aios binary. Full list on the glossary page.
- iiotd — the IIoT Worker. Ingests from PLCs / SCADA / DCS / field devices over the protocols they actually speak (OPC-UA, Modbus TCP/RTU, MQTT Sparkplug B). Store-and-forward buffering so a dropped WAN link doesn’t lose telemetry.
- Data Worker — historian storage, RAG over maintenance manuals and SOPs, matrix memory for learned patterns.
- Data-plugin fan-out — a single chat query fans out to the historian, the manuals RAG, the live sensor stream, and the structured-document retriever in parallel. The LLM sees citations from all four.
- On-prem LLM — embedded llama.cpp with optional CUDA add-on for GGUF inference. No data leaves the plant.
- Audit trail — hash-chained record of every retrieval + LLM call. Plant managers get the “why did the AI say that” trace for free.
No new dashboard. No rip-and-replace of MES / SCADA / CMMS. A chat window that knows the plant, sitting next to the systems you already run.
Where this goes next
Alpha.3 shipped on 2026-04-23 — the Phase-3 kernel: identity system with real users / tenants / workgroups, projects + skills CRUD, 127 API routes all wired, 14 dashboards, and the modular webchat shell. The same extension model makes plant-specific sensor schemas or proprietary instrument protocols a manifest drop, not a code fork.
If your technicians spend more time looking things up than fixing things — this is the shape of the fix.
Questions, plant-pilot interest, or just want to push back on any of this — DMs open on LinkedIn or juergenp@core.at.