Plant data put to work for the engineers who need it
Built by an instrumentation and control engineer who also writes production Python. The tools run inside your network, read the systems you already have, and answer the questions your engineers actually ask.
Service packages
-
Plant Data Readiness Assessment
An inventory of your historian, DCS, maintenance and laboratory data, what each source can and cannot answer, and a short list of use cases worth building first.
Deliverables
- Data source inventory
- Target architecture
- Ranked use-case list with effort estimates
$4,800Fixed price, two to three weeks
See an example deliverable
Source Coverage Quality Verdict Historian, process values 1,240 tags, 1 s to 1 min 3 % gaps, 2 frozen tags Ready Laboratory results 4 years, all grades Sample times entered by hand Fix timestamps first Maintenance work orders 6 years Free-text failure codes Needs mapping Alarm journal 12 months 18 % chattering alarms Clean before use Illustrative excerpt written for this site; tags and values are invented, not client data. -
Engineering Document Assistant
A retrieval-augmented assistant over your manuals, operating procedures and design documents, running on a local language model. Every answer cites the document and page it came from.
Deliverables
- Pilot on your hardware
- Evaluation question set with scored answers
- Operating and handover documents
from $12,000Pilot, four to six weeks
See an example deliverable
Item Content Question What is the trip setpoint and proof test interval of the separator high-pressure function? Answer Trip at 18.0 barg with 2oo3 transmitters; proof test every 12 months. Sources SRS-012, section 5.2, page 14 · Proof test procedure PT-SIF-12, step 2 When not found "Not in the indexed documents." No answer is invented. Illustrative excerpt written for this site; tags and values are invented, not client data. -
Plant Analytics Portal
Web modules that bring historian and DCS data together with maintenance work orders and lab results, with scheduled reports by email.
Deliverables
- Portal MVP with first modules
- Data pipelines
- User guide
from $15,000Minimum viable portal
See an example deliverable
Equipment Running hours Vibration trend, 30 days Next step Pump A 2,180 h since overhaul Stable, 2.1 mm/s None Pump B 6,920 h Rising, 3.4 → 5.8 mm/s Work order suggested Compressor 11,400 h Stable Oil analysis due Illustrative excerpt written for this site; tags and values are invented, not client data. -
Computer Vision for Site Safety
Detection models trained on your own camera footage, for example vehicle speed monitoring on plant roads, with event snapshots and a simple review screen.
Deliverables
- Trained model and service
- Event review screen
- Accuracy report on your footage
from $9,000One use case, one camera view
See an example deliverable
Time Zone Speed / limit Event 08:14:22 Gate road 34 / 20 km/h Over limit, plate read, image stored 08:21:05 Tank farm road 18 / 20 km/h No event 09:02:47 Gate road 27 / 20 km/h Over limit, plate unreadable, image stored Illustrative excerpt written for this site; tags and values are invented, not client data. -
Quality Statistical Modelling
Statistical models that relate process conditions to laboratory quality results, to find which variables move a quality parameter and by how much.
Deliverables
- Model and code
- Findings report
from $6,000One product line and one set of quality parameters
See an example deliverable
Driver Effect Confidence Reactor temperature +1 °C → colour index +0.4 High Residence time +10 min → +0.3 High Additive feed rate Effect only above 120 ppm Medium Model fit R² 0.81 on unseen data; mean error 0.35 Holds across grades Illustrative excerpt written for this site; tags and values are invented, not client data.
Prices in US dollars. Hardware and software licences, where needed, are bought by the client.
Reference build: a plant intelligence platform
The packages above come from a platform designed, built and run in a world-scale process plant.
The portal
More than 15 modules on one read-only data layer: a live process overview by unit, maintenance work from the ERP, live spare-part stock and withdrawals, design against field values with tags read from P&ID files, capacity and budget planning, daily quality reports and energy use by consumer. Bilingual throughout.
The assistant
A local language model on one GPU inside the network, with more than fifty read-only tools for process, maintenance, stock, planning, quality and documents. Each module has a scoped profile, engineers teach it by correcting it in the chat, and every answer lists the data points it used. It is tested by scripted question rounds and a weekly regression set.
Reference architecture
Everything shown runs inside the plant network. No plant data is sent to an outside AI service.
Data stays on site
Language models run on your own hardware, so documents and process data never leave the network. Connections to historians, DCS and business systems are read-only.
A model that answers without a source is treated as a defect. Assistants are tested against a question set agreed with your engineers before anyone relies on them.
Built to be maintained
Plain Python, standard databases and documented interfaces. Your IT team can run and extend what is delivered without a proprietary platform.
Questions
What hardware does a local LLM assistant need?
A single workstation-class GPU server is usually enough for a pilot. The readiness assessment sizes it for your document volume and number of users.
Can it read P&IDs and scanned documents?
Text-based PDFs and office documents index directly. Scanned drawings need text recognition first, and the assessment states what quality to expect from your scans.
Do you use cloud AI services?
Not with plant data. Cloud services are only used if you ask for it and your policies allow it.