Machine Monitoring

An ERP knows what work should be happening. A monitoring system knows what is happening. Almost no ERP does both — so a shop buys two systems, then pays a third time to connect them.

That integration cost is the whole argument.

Several early-access candidates were days from signing with competing vendors. One — Hill Manufacturing — paused a MachineMetrics evaluation because they wanted monitoring inside ProShop, not in a second vendor's interface. I designed against that sentence more than any persona document. They weren't asking for a better dashboard. They were asking not to leave ProShop.

Machine Monitoring didn't exist in ProShop, so every convention had to be decided rather than inherited. Our PM's POC answered what and why credibly, but not how best — and with one developer and five weeks, getting that wrong meant a bad foundation under everything the platform builds next.

The second mandate on this project wasn't a feature: work out how a team collaborates using AI agents, and keep the method.

What AI did, and what it didn't

I used AI to break the PRD into its constituent decisions, run competitive research, and synthesize two years of Customer Advisory Board notes and product specs into the handful of findings that bore on a design decision. But the IA and workflow were mine, in Figma, first. The prototype's job was to test concepts I'd already formed — generating first and rationalizing after is how you end up defending an output you didn't choose.

The POC had one flat grid, and I first read this as a sizing problem: the same view needs to work at a desk and across a shop. That's the shallow version. The real difference is job.

An ops manager asks "is the floor running well, and what do the numbers say?" Someone walking past a wall display asks "is anything wrong right now?" — and will never drill into anything, because they're on their way somewhere. One view serving both fails both.

What the shop floor taught me about the wall

So Floor Status ships with card customization and a re-arrange mode — a shop orders the tiles so the board mirrors its actual layout. The display stops being a list and becomes a map, which is what turned that view from a smaller dashboard into a spatial one.

The question that made a supervisor click was comparative — VMC-03 is at 48%, how are the others doing? A detail view that answers it by destroying that comparison is worse than none.

Where the detail opens

I argued the panel from the user's question rather than from aesthetics, which is why it stuck.

How the panel is organized

One machine holds more than a panel can — five MTConnect domains plus work order data the machine doesn't know about. Rather than one long scroll, four tabs, each answering one question, ordered to how attention moves across a shift.

When the data isn't there

A monitoring tool spends real time off the happy path. When the collector stops reporting, a machine flags OFFLINE rather than freezing on its last known status — a supervisor trusting a stale green card is worse off than one looking at a gray one, so absence of data is itself a status.

Before any machine reports, the grid shows named machines with blank cards and help text, which makes the empty state a configuration checklist wearing the interface of the finished product.

The most consequential thing I did here wasn't a screen.

Showing the active job is a display problem. Proving it's the right job spans two modules: the controller knows a program name, the schedule knows the work order. Get it wrong and every number it feeds is wrong too.

A developer works inside a module, a PM inside a feature. I was walking the whole path. I escalated it in week two as scoping, not a design workaround.

Why Schedule Sync became a state machine, not a prompt

So the match is proposed rather than assumed, and it holds a pending state until someone resolves it: one click to confirm when it's right, and a work order lookup rather than a free text field when it isn't — with the correction recorded.

Labeling where every number came from

Quantity Progress splits one count into three, each labeled with where it came from.

No machine can verify a part meets spec — only a person can. That is why the conforming count is entered rather than read.

So the panel carries both counts rather than reconciling them into one — 16 counted at the spindle, 14 validated as conforming, 2 still waiting on a person. Total complete is also the only one an operator can edit: where a number comes from decides who is allowed to change it.