Now live: a hybrid process digital twin for amine gas sweetening — it names the fault and the root cause, not just the alarm. Explore the platform →

Plant Assistant

qemIntel

Live in the platform

Ask the plant a question. Get an answer it can prove.

qemIntel sits beside every tool. It reads the screen you are on and the plant knowledge base behind it, then answers in plain language. If the answer is not in the data, it tells you that instead of guessing.

See it in the platform How it works

11Knowledge sets searched

65kTokens in the knowledge base

~2kTokens used per question

0Values it will invent

What it does

Four things it does that a dashboard does not.

It reads the screen you are looking at

Every tool publishes a structured snapshot of exactly what it is displaying. Ask which KPI is out of band, or why a fault fired, and the answer comes from the numbers in front of you — not from a general impression of how amine plants behave.

It knows the plant, not just the data

Faults and what drives them, root causes, indicators, sensor tags, equipment, recorded fault episodes, engine notes, open items — and the blind spots, so it can tell you what the model cannot see.

Retrieval, because cost is an engineering problem

The full knowledge base runs to roughly 65,000 tokens. Searching first and sending only the cards a question needs keeps a typical query near 2,000. That is the difference between a few dollars a month and a few hundred — at one train.

It refuses to invent

If the snapshot does not carry the answer, it says so and points at the tool that would show it. It never estimates a figure the data does not hold, and it never invents a tag, a fault or a timestamp.

How it works

The pipeline, end to end.

1
Ask from any tool page
The assistant panel is always there, next to the work.
2
The page publishes its facts
Each tool reduces what is on screen to structured values — the same numbers you can see, without the chart geometry.
3
The knowledge base is searched
Records are flattened into short cards and scored against the question. Only what is needed is retrieved, under a fixed budget with a per-card cap so one long fault record cannot consume it.
4
The answer is composed
Screen snapshot first, knowledge base second. General process-engineering knowledge is allowed, but it is flagged as such rather than blended in silently.

Under the hood

The engineering, stated plainly.

Every figure on this page is countable in the delivered system. If you want to see where a number comes from, ask us and we will show you the record it is read from.

Grounding: Page snapshot first, plant knowledge base second
Knowledge sets: Faults · causes · indicators · tags · equipment · edges · episodes · blind spots · notes · open items · results
Retrieval budget: ~2,000 tokens per question, with per-card caps
Injection guard: Page text is treated as data, never as instructions
Provider: Model-agnostic — swapping provider is configuration, not code

What it does not do

The limits, published on the same page as the promise.

To answer about a screen, that screen has to be open. It is a co-pilot beside the work, not a reporting service you email.
It will not give you a number that is not on the page or in the knowledge base. That is a design decision, and it is the reason its answers can be checked.
It reads. It does not run the model, change an input, or act on the plant.

Where it applies

Built for amine systems today. Not limited to them by design.

The platform is focused on acid gas removal — gas sweetening and post-combustion carbon capture — because that is where the model is calibrated and where the fault library is real. The method underneath is not amine-specific: a physics reference model, residuals against expected-healthy, fuzzy diagnosis over graded indicators, and costed consequences. Extending it to another unit means new datasheets and a new calibration, not a new platform.

Gas sweeteningLive today
Post-combustion carbon captureLive today
Other acid gas removalMethod applies
Wider process unitsFuture

The rest of the platform

One model. Seven ways to use it.

RunThe process digital twin that says what a healthy train should be doing.DetectNames the fault, the root cause, and what it costs per hour.PrognoseProjects the slow-wear faults against the threshold that declares them.DiscoverThe whole train in KPIs, trends, profiles and correlations.OTSIn developmentTrain crews on faults they have never seen. In development.TuneIn developmentFind the valves and loops working against you. In development.

The fastest way to judge this is to open it.

The platform runs on a real amine train with thirteen days of one-minute history behind it. Nothing on these pages is a mock-up.

Open the platform