Measure · Plan · Explain
OEE you can defendin a meeting.
AI on the plant floor that cannot invent a number. Understand the problem, what it costs, and what to fix first.
The AI tool for your production team.
14 days, one line, no card.
- No hardware
- No OT network access
- One CSV to start
Denim Sewing 2 · Tánger · last 30 days
Loss found and priced
$164,605
Priced at the contribution margin this line earns per hour — not a generic rate card.
$15,500 of it is upstream starvation — the fix is not a mechanic on this line.
Live measurement
Plan 75.0% · gap −10.2 pts · 14 days
Fourteen days, every one of them below plan. The accent segment is the most recent day — the one still worth acting on.
The artifacts
What lands on Monday morning.
Every figure below comes from one file, so the arithmetic agrees across all three screens. Check it with a calculator.
Plan 75.0% (−10.2)
Against all calendar hours, not just scheduled ones.
- 1Minor Stops$67,695 · 41%
Hatched: $15,500 the line sat starved, not broken — upstream supply.
- 2Breakdowns$41,220 · 25%
- 3Setup & Changeover$22,110 · 13%
- 4Production Rejects$21,400 · 13%
- 5Startup Rejects$12,180 · 7%
Top three categories carry 79.6% of the money.
- 1Minor Stops is the biggest loss$67,695needs investigation
503 of those minutes were the line sitting starved rather than broken — $15,500 of it. That share is upstream supply, not maintenance.
- 2Changeovers run 10 minutes long$22,110fast / low-cost fix
34 minutes against a 24 minute baseline across 40 events. Changeover practice is the lever here, and it is this line's own work.
Handing this to whoever signs the check? Read the two-minute version.
What the AI does
It reads what your plant already writes down.
Eight places where somebody was already doing tedious work by hand. Pick one to see what the engine guarantees about it.
Capability 01 of 08
It writes the report
A plain-language account of the period — the biggest losses and which of them moved — written for the person reading it.
Every number is checked back against the computed payload before storage; a sentence citing a figure that is not in the data is discarded whole.
Capability 02 of 08
It reads your stop codes
Groups the six spellings your plant uses for one failure into a named mode, glossed from your own equipment manual.
It may only quote the manual word for word — a code the document does not cover stays unglossed rather than guessed at.
Capability 03 of 08
It suggests why the line stopped
An unlabelled stop gets a suggested reason from what this station has been called before. Confirm fourteen similar stops in one action.
The engine forms the group from this line's own history; every label is attributable and individually reversible.
Capability 04 of 08
It reads a photograph
Point a phone at an HMI or the handwritten downtime log, and the paper is compared against what the historian recorded.
The model is held to its own transcription, the photo is kept beside the reading, and nothing from a photograph is ever written into your production record.
Capability 05 of 08
It explains the losses
Builds cause chains between failures, and flags the finding that came back after somebody said it was fixed.
The statistics are the engine's, multiplicity-corrected; the model names and orders findings, never decides whether one is significant.
Capability 06 of 08
It sets the plant up by asking
A new plant describes itself from its own line names and order book instead of a configuration form.
Everything is staged as a draft a person confirms — it cannot write a rule that silently changes what your history means.
Capability 07 of 08
It answers questions about the plan
Propose a scenario in words and see it costed against the real order book. It learns the trade-offs your planners keep making.
The engine computes every schedule; the model only argues for it, and a planner approves or rejects.
Capability 08 of 08
It reads what people already wrote
Six sources of free text — stop notes, handovers, remedies, refusals — mined for the theme nobody was assigned to notice.
Held to the strictest rule of all: no figures, because a corpus of sentences has no computed payload to check one against.
Why you can leave it switched on
The model proposes. The engine proves.
Four rules the software is built around rather than reassurances written afterwards.
1
Your rows
The export your historian already writes.
2
Engine computes
Every figure, in code, before any model runs.
3
Model writes
Prose, ordering, naming. No arithmetic.
4
Checked, then stored
Cited figures matched against the payload.
Fails the check
Discarded whole, not trimmed.
No figure on any screen was written by a model
Every number is computed in code from your rows before any model runs.
It proposes; a person decides
Groupings, loss rules, schedules and mappings all arrive as proposals with a name on them.
It refuses rather than guessing
"No data", "not measured" and "not enough to score" are three different facts, and it will not merge them.
Nothing acts on its own
Calendars, loss rules, thresholds and mappings all re-score numbers you already published, so none changes autonomously.
Ingestion
Bring the export you already have.
Optora reads three shapes and detects which one you've sent. Download a sample and check your own export against it — the same files the app's test suite imports through the real pipeline on every run.
The samples run on a Mon–Fri 08:00–16:00 day shift in Europe/Madrid — set the plant timezone and a shift calendar to match, and the numbers will line up.
The mapping is config-driven rather than a parser per vendor, so any historian that exports CSV works.