AI is only as good as its data and its context, so we built both first.
Point a model at raw machine logs and it'll answer confidently, and often wrong, because the logs never say why anything stopped. MACH captures the reason the operator gave, and what that number actually means in your plant. Virgil, the assistant on every dashboard, works from both.
Data
Stops carry the operator's reason, rejects are tied to their order, and changeovers are timed as they run, all in one history.
Context
Every plant has its own definition of OEE, and its own KPIs for the processes it actually runs. You settle what each one means here, and Virgil reads that before it answers.
Ask Virgil
Ask a question on any dashboard and get an answer from your own production data. Open the query behind any number it gives you.
Governance
Virgil reads. It never edits your production history, and if it's ever allowed to act, a person approves first.
Last updated: September 16, 2026
Machine data a model can actually answer from.
A machine log will tell you the press stopped at 9:12. It won't tell you why, which order was sitting behind it, or who was running it that shift. Those are the parts a model needs, and MACH captures them at the machine, while the job is still running.
Pieces, tons, linear feet, whatever the process runs in.
Output is counted in the machine's own units rather than forced into a generic part count, so the answer fits the process that produced it.
Classified at the machine while it's still stopped.
The operator taps a category and a reason before the machine is back up, tied to the shift and the order it held up.
Logged against the order it came off.
Scrap carries the order, the material, and the machine, so yield by product is something you look up rather than reconcile at month end.
Timed while they run.
The clock starts when the last order finishes and stops when the next one is up to speed, with nobody holding a stopwatch.
Every operation defines its own numbers. Virgil answers in yours.
Every plant has its own OEE formula and its own idea of what "on pace" means, and a general-purpose model knows none of it, so it guesses. In MACH you choose every formula and definition. We bring recommendations from years on the floor, but you make the call, and the reasoning behind it stays with the KPI. Virgil reads those lines when it's looking at that number.
It knows what Press 2 is.
Virgil knows the plants, work centers, machines, stations, and shifts, so it can tell which line a machine sits on without being told.
Your formula, in your words.
Each KPI carries the reasoning behind it, not just the math.
It starts where you're looking.
The plant, machines, and date range on your dashboard are already the scope when you start typing.
How the plant actually runs.
A few paragraphs about the plant, like which lines matter most and what a normal week looks like. Virgil reads them before every answer.
Ask the question you'd ask your best analyst, and have the answer before the meeting.
Virgil is MACH's assistant, a continuous improvement analyst on every Hub dashboard. It already knows the plant, machines, and date range you're looking at. Ask which machine improved the most last month, or why night shift keeps trailing day shift on Press 2.
Press 1 improved the most on OEE, up 6.3 points against July. Press 3 improved less but did it on a harder mix.
| Machine | OEE Jul | OEE Aug | Δ |
|---|---|---|---|
| Press 1 | 66.9% | 73.2% | +6.3 |
| Press 3 | 72.0% | 75.1% | +3.1 |
| Press 2 | 70.4% | 68.8% | −1.6 |
Unplanned downtime fell 11.2 hours, with 8.4 of that under "material feed jam." Order mix was 88% the same SKUs as July. Changeover count was flat.
Press 3 ran 14 more changeovers than July and still gained 3.1 points, so its improvement is harder-earned. Press 2 is the one to look at: schedule adherence dropped to 81%.
Here's what you're seeing: one question, answered from the same KPI definitions the dashboard uses, with every figure traced to a query. Notice it also brought up the machine nobody asked about.
- When the data isn't there, Virgil says so rather than estimating around it. Conversations stay with the dashboard they're about, so Monday picks up where Friday left off.
Some questions you ask once. Others you want answered every Monday at 5am.
Virgil can run on a schedule, so the question you'd normally type at 7am on Monday is already answered and waiting. There are two kinds, because "send me the report" and "tell me if something's wrong" are different asks.
A report delivers every time.
The prior-week ranking lands in your inbox before the production meeting, findings on top with the charts and tables underneath.
A monitor speaks up only when something is wrong.
It stays quiet while things are on track. When they aren't, it tells you once and then holds off, instead of repeating itself every hour.
- Results land on the dashboard they're about, with the conversation open and ready for the obvious next question.
Your data science team keeps its models and finally gets labeled inputs.
Read-only query APIs expose the same history to your BI tools, data warehouse, data lake, or in-house models. And if you'd rather your own model do the asking, it can call MACH's agents over MCP and inherit the context and governance.
Read-only query APIs.
Machine-day metrics, downtime with reasons, order steps, and schedule performance flow into the tools you already run. You build the models, and MACH supplies the labels they never had.
Your own model, over MCP.
Connect your model, copilot, or agent platform to MACH's MCP server and it calls the same agents your team uses in Hub, with the plant context attached and the same rules in place.
It reads. It never writes.
Virgil earns trust the way a new engineer does: it shows its work and doesn't touch anything it wasn't asked to.
It can't change anything.
Virgil can't edit a schedule, change a KPI, or touch a record. If it's ever allowed to propose an action, a person confirms first.
Your definitions apply on every query.
Every question goes through the same query engine your dashboards use, so your OEE formula and your plant's dates apply every time.
Every step is recorded.
Each query it ran and what came back stays with the answer. Open "How I got this" on any message.
MACH is building its own AI on this foundation.
Virgil is the first piece, live today as the dashboard assistant and the scheduled reports and monitors above. Everything else on the roadmap reads from the same foundation, so none of it needs a new data project first.
Predictive failure alerts
Virgil learns the patterns that show up before a stop from downtime history with the operator's reason attached, and flags them before it happens.
AI-triaged work orders
The operator escalates from the kiosk. The work order arrives at the right technician with the stop history already summarized.
Auto-scheduler
Durations come from what each job actually took on each machine, instead of the estimate in the routing.
Executive intelligence layer
Leadership gets insight cards and a weekly briefing built from the plant's own data.
Questions about MACH AI.
What does it mean that MACH is AI native?
MACH was built so labeled production data exists and so you can put it to work: every stop classified by the operator as it happens and linked to the machine, shift, and order it delayed, rejects logged against the order, changeovers timed as they run, and corrections kept with an audit trail. The data and the plant context both came before the model.
Does MACH have an AI assistant?
Yes. Virgil sits on every Hub dashboard, answers from your own KPIs and production data, cites every number to a query you can open, and can run on a schedule as a report or a monitor. It's read-only.
Why not point a general-purpose AI model at our data?
Because the data alone isn't enough. Raw machine logs don't carry reasons, and a general model doesn't know your OEE formula or your reason codes, so it gives you a confident answer you can't check. MACH holds the plant context, runs every question through typed queries that use your definitions, and cites each number to a query you can open.
Is Virgil safe to run in a production plant?
Yes. Virgil is read-only. It can't edit a schedule, change a KPI, or modify a record, every step is recorded so you can open the trace on any answer, and an admin can turn it off with one toggle.
Can I connect ChatGPT, Copilot, or my own model to MACH?
Yes, through MACH's MCP server. Your model, copilot, or agent platform can call the same agents your team uses in Hub, with the plant context attached and the same governance in place. If you'd rather pull the raw history, read-only query APIs expose it to your BI stack, warehouse, lake, or in-house models.
The data is already being captured. See what it says about your floor.
We start with a Shop Floor Diagnostic. One-hour call to understand your situation, then a written assessment of where the losses are and what closing them would be worth. Where we can't help, we say so.