Audiences AITriage · TRIAGE-type decision system · built first for Manufacturing
Your plant knows it is losing time. Nobody decides which loss to attack.
Monitors measure. Solution suites fix a loss you already picked. Triage is the layer in between: a costed loss map and a recommendation your team accepts or rejects.
Available for pilot. The pilot runs read-only against your own data, and the first deliverable is scored for sufficiency before anything else is promised.
The downtime log tells a confident story
Improvement budget gets allocated off the labels your operators had time to type. In one documented case, the log said seventy percent equipment failure. Automated capture against the same period found eighteen.
Documented case from the product’s concept-validation research: automated capture measured against manual logs. Directional industry context. Not a claim about your plant.
Every lost hour in a Triage loss map is priced, carries a cause, and is computed from your events.
The output is a card your analytics team can audit
Triage delivers a short queue of cards. Each is a single decision waiting on your accept or reject, and nothing runs until you give it. Every figure on a card traces back to the underlying events.
Worth is sized against what your plant has already achieved, never an industry average. Where an input is missing, the card says so and narrows its claim: hours instead of currency when no margin data was supplied.
Avoidable changeover
confidence: 0.81- recommendation
- Group the four A-family orders on Line 2 into one run, before the switch to B.
- why this
- Avoidable changeover is the second largest recoverable loss on this line, and the order book was loaded throughout, so demand is real.
- evidence
- 41 A-to-B transitions last month. Demonstrated baseline on this transition: 34 minutes. The line averaged 95. Operators coded 38 of 41 as SETUP, which agrees with the computed picture.
- worth
- About 68 machine-hours a month on Line 2. Hours, and deliberately so: no margin data was supplied, so no currency figure is claimed.
Illustrative card with synthetic figures, shown to demonstrate the reasoning and the audit trail.
The same code, twice
Two three-hour stops, both coded NO MATERIAL. Only the computed picture separates them.
what separated it
Work was due to run. The idle sits against a late receipt from a vendor that has been late before.
verdict
Readiness failure. Machine ready, demand real, material absent. Recoverable. routed → vendor diversification
what separated it
Nothing was loaded and demand is soft. The missing material is a consequence, and treating it as the cause would mean making stock nobody ordered.
verdict
Demand-driven. Correctly idle. not routed · on purpose
Challenge a claim and it is re-diagnosed on the evidence. A claim that cannot survive is withdrawn, never defended.
The gates it runs before it diagnoses
Most runs come back as a full loss map with routed recommendations. The gates exist for the runs that should not. Every run starts with a data-sufficiency score against fixed floors. The thresholds are set in advance and cited to the product specification; the run cannot tune them to please anyone.
full diagnostic→ caveated→ partial→ refusal
Each tier is a defined product state with its own consumer set, and a category one gate withholds can never be quietly re-added by another. Below forty percent non-generic coded data, Triage refuses to diagnose. The refusal ships a deliverable: a data-gap report and a logging prescription stating exactly what to record and for how long. Log it, resubmit, and the run is re-scored.
Designed and tested in the product codebase. No field outcome is claimed here.
Explore a sample loss map
One month of losses on a synthetic manufacturing plant, priced in machine-hours. Pick a loss; the card that Triage would issue for it appears alongside.
A synthetic manufacturing plant with illustrative figures. This is the product surface, and only the surface.
Avoidable changeover
confidence: 0.81- recommendation
- Group the four A-family orders on Line 2 into one run, before the switch to B.
- worth
- About 68 machine-hours a month. Hours, because no margin data was supplied.
- route
- scheduling
Readiness failure
confidence: 0.74- recommendation
- Qualify Vendor 7 as a second source for material M. Vendor 3 was late on 9 of 22 receipts.
- worth
- About 41 machine-hours a month across two lines.
- route
- vendor diversification
Speed loss
confidence: 0.68- recommendation
- Take the B run to five days of cover. The line has already demonstrated the faster rate.
- worth
- About 37 machine-hours a month on Line 1.
- route
- lot-sizing and batching
Preventable stoppage
confidence: 0.71- recommendation
- The stops cluster at shift change on Line 3. Stagger the handover procedure before buying anything.
- worth
- About 29 machine-hours a month.
- route
- scheduling
Unattributed
held- finding
- Reason-code coverage for this slice sits below the floor. No diagnosis is issued.
- deliverable
- A logging prescription: which codes to add and how long to run them before re-scoring.
- route
- held · data-gap report
Six routing families
| Scheduling | Re-sequences what runs when. Run the four A orders back to back before switching to B. | solver |
|---|---|---|
| Lot-sizing | Re-sizes the runs. Take the B run to five days of cover and changeovers fall. | solver |
| Vendor diversification | Flags supply that keeps failing, using only vendors already in your own receipt history. | rules |
| Forecasting | Fixes demand-side readiness where an intermittent SKU is driven off the wrong method. | solver |
| Pricing | Flags thin margin on scarce capacity, for your commercial judgment. Any price change stays a human call. | rules |
| Maintenance | A flag, and only a flag. Predicting failures needs sensors on the machine, and Triage will not claim it from flat files. | flag |
The routing engines are off the shelf: open source and published. If your scale outgrows one, the engine swaps out without changing the card.
What it will not do
You approve every card. Write-back to your ERP is off by default, scoped per family, with dry-run and rollback if it is ever enabled.
Every code is treated as generic until the data proves otherwise. Codes corroborate the computed picture; they never drive the routing.
It refuses, and tells you what to log instead.
Every inferred number carries a confidence score, and inference is never presented as ground truth.
It runs containerized inside your own network. Nothing leaves.
Worth is sized only against what your plant has already hit.
How a pilot runs
-
Base integration against your existing systems, read-only. Your data is scored for sufficiency, honestly. First loss map, costed where you supply margin data, within five working days of data receipt, delivered with its data-sufficiency report.
-
Your operations leadership confirms the top loss categories against floor reality: a yes or a no. Disputed categories are re-diagnosed on the evidence. Then a routing preview: cards, with no execution.
-
The loss map becomes a living monthly artifact: continuous re-diagnosis, drift alerts when changeovers creep or bottlenecks migrate, tracked results on every card you act on, and re-calibrated baselines.