Maintenance on the bottleneck line
The press fails on the night shift; its vibration had changed weeks before.
- Kinds of AI
- ModelsSLMsAgentic AI
- Who decides
- Maintenance planner
- Where it runs
- Your servers or your cloud account
Model · risk of unplanned stop, next 7 days
SLM · manual
- §7.3Shaft seal replacement, summarized in Turkishpump service manual p. 41
Agent · draft work order (not released)
- TaskInspect shaft seal, replace if leaking; check coupling alignment
- PartsSeal kit in stock (2), reserved
- Estimate2.5 h, 2 technicians
- Model
- SLM
- Agent
DecidesMaintenance planner
1. The desk
Today
- DetectMaintenance learns of a failure from the operator; the whole plant waits on the bottleneck.
- DiagnoseThe fix is in an equipment maker's manual in German or a work order written three years ago, and nobody finds it at 3 a.m.
- PrepareParts are looked for, or bought in a rush, only after the line has stopped.
- ScheduleThe maintenance planner sees planned windows (changeovers) pass unused, because nobody knew the pump was degrading.
With the system
- DetectThe failure-risk model watches each asset's signals and flags the risk of an unplanned stop, with the signals that moved.Model
- DiagnoseThe SLM reads the manuals and past work orders and returns the likely cause and the procedure, its page cited, summarized in Turkish.SLM
- PrepareThe agent checks spare-part stock, the production plan, and the next planned window, drafts the work order, and reserves the parts without issuing them.Agentic AI
- ScheduleThe maintenance planner receives the draft with a proposed slot, releases the work order and schedules the stop, as before; the shift lead still owns any stop.Person decides
What lands on the desk
The maintenance planner gets the pump's risk and the signals behind it, the repairs that fixed it before, and the manual page in Turkish, in a work order that is drafted but not released. The parts are reserved and the slot is a proposal until the planner decides.
Model · risk of unplanned stop, next 7 days
Agent · similar past cases
- Seal leak (2.5 h)work orders
- Bearing (9 h)work orders
SLM · manual
- §7.3Shaft seal replacement, summarized in Turkishpump service manual p. 41
Agent · draft work order (not released)
- TaskInspect shaft seal, replace if leaking; check coupling alignment
- PartsSeal kit in stock (2), reserved
- Estimate2.5 h, 2 technicians
- SlotThursday 06:00 planned changeover (no lost output)
- NeedsMaintenance planner approval
- Model
- SLM
- Agent
DecidesMaintenance planner
2. The system
For your technical team — skip it if you like.
How it works
In
- Vibration, temperature, pressure, current
- Cycle time
- Plant historian, supervisory control (SCADA)
- Line controller (PLC) alarms
Model · estimates
Failure-risk model, per asset
Anomaly detection; supervised only where failures are labeled
Reads
- Everything under In
Out
- Risk of an unplanned stop in the horizon
- The signals that moved
SLM · understands
Manual and work-order reader
Retrieval over manuals and work orders; tuned on shorthand
Reads
- The risk and the asset ID
Out
- The likely cause
- Similar past cases
- Manual procedure, page cited
- Summarized in Turkish
Agent · prepares
Work-order agent
Tool loop
Reads
- The cause and the procedure
Out
- Spare-part stock from the ERP
- Production plan, next planned window
- Draft order in the maintenance system (CMMS)
- Parts reserved, not issued
Decision
Maintenance planner and shift lead
Reads
- The draft work order
Out
- The maintenance planner releases the work order and schedules the stop
- The shift lead owns any stop
How the agent's work is arrangedTool loop
Tool loop
What to check next depends on the last answer (part in stock or not, window this week or not), so the agent picks from approved tools until the draft is complete or a stop rule fires.
How the models are trained and tested
Your expert writes the test first. A step the test does not need is skipped, and we say why.
Show the training steps
Anomaly and failure-risk model
Target
An unplanned stop within the horizon where labels exist; otherwise an anomaly against a learned normal.
Thresholds
Set by weighing the cost of a false alarm against a missed failure; alerts sized so one shift can act on them.
Validation evidence
Alert precision on past failures, and how far ahead of each failure the alert came.
Monitoring
Drift per sensor, after go-live.
Small language model (SLM)
Test set
100–200 past failures with the technician's actual cause and procedure, graded by senior maintenance staff.
Dataset
Work orders, failure codes, and manual sections.
Continued pre-training (CPT)If needed
Only if the base model fails the test on the plant's shorthand.
Supervised fine-tuning (SFT)
On work-order shorthand, part codes, and the answer format.
Reinforcement learning (RL)Skipped
Skipped: retrieval with cited pages and fine-tuning cover the task.
Where the agentic system stops
The maintenance planner schedules every stop and releases every work order; the shift lead owns any stop on the line.
Show its tools, limits, and record
- Tools and access
- Read sensors, the CMMS, ERP stock, and the production plan; create a draft work order; reserve parts up to a set amount.
- May not
- Stop a machine, change a PLC parameter, release a work order, or order parts above the set amount.
- Who decides and the written limit
- Maintenance plannerThe maintenance planner schedules every stop and releases every work order; the shift lead owns any stop on the line. The agent drafts and reserves only.
- Stops when
- A sensor data gap or model drift: an alert only, with no draft work order. A part is not in stock and its lead time runs beyond the risk window: the planner at once. Five tool calls without a complete draft: handed to the planner.
- Record
- The risk score, signals, documents cited, tool calls, and the planner's decision.
3. Value and conditions
Where the time goes, and what your institution will measure.
Where the value comes from
- Fewer unplanned stops on the constraint.
- Shorter diagnosis, and repairs moved into planned windows.
- Less rush buying of parts.
What your institution measures
- Unplanned downtime hours per month
- Mean time between failures (MTBF)
- Mean time to repair (MTTR)
- Availability, as part of overall equipment effectiveness (OEE)
- Share of maintenance planned
- False-alarm rate
Measured in the Decision Blueprint
Baseline downtime, and how far ahead of past failures the alerts would have come, set the starting point; the plant sets the target.
Conditions
The rules, the data, and the known error, with its guard.
- Few labeled failures: an anomaly is not a prediction, and the file says so.
- A high false-alarm rate can erase the savings.
- Equipment makers' manuals are for internal use.
- Shift data names people: the model watches machines, not operators (KVKK, the Turkish data protection law).
Start here
We take the sensor history and past failures of your bottleneck line, check how early the signals moved before each one, and test the reader on your manuals and work orders before anything is built.
Similar desks
- ModelsSLMsAgentic AI
TodayInspectors check every part under a lamp; night shift misses more, and each nonconformance report takes half a day.
With the systemA draft nonconformance report with its evidence, ready for a hold-or-release call.
Quality engineer decides
- SLMsAgentic AI
TodayContainment is due tomorrow, and the evidence is in five systems.
With the systemEvidence from five systems in one 8D draft.
Quality engineer decides
Equalize the gap.
Bring one decision. Leave knowing the first step.
A short description of the decision is enough; no confidential records.
https://eqlx.ai/use-cases/predictive-maintenance · Printed copy; the website is the current version.