A11 / A12 Reliability Desk · Unilever, Leeds

MaintAIn — Solution Overview

A reliability dashboard and AI assistant for the A11/A12 "Pester" shrinkwrapper line — built to help maintenance and reliability teams see breakdown and preventive-maintenance trends at a glance, and get grounded, cited answers to questions about the equipment, its history, and how to troubleshoot it, by typing or by voice.

Own your intelligence.

01 · A11 / A12 Reliability Desk

Capability Overview

What the dashboard and assistant do, how the assistant finds and grounds its answers, and how to use it — by text or by voice.

MaintAIn brings together two things reliability engineers usually have to hunt for separately: a live view of breakdown and preventive-maintenance data, and an assistant that can answer questions about the equipment in plain English (or by voice), grounded in the plant's own documentation and history.

Reliability Dashboard

  • Reactive Maintenance — breakdown counts, fault categories, chronic issues, filterable by line, machine, category and time period.
  • Preventive Maintenance — cleaning/inspection/lubrication task status (overdue / due soon / on track), filterable by machine, responsible role and reason.

AI Assistant — Text & Voice

  • Ask questions by typing, or by speaking to it directly.
  • Answers are grounded in maintenance manuals, operational procedures and troubleshooting guides — not the model's general knowledge.
  • Every answer shows which document(s) it came from, with the source excerpt available to view alongside the answer.
  • Can also reason over root-cause questions ("why does X keep failing?") by combining breakdown history, preventive-task records and the documentation together.
Both the dashboard and the assistant sit behind the same sign-in used across the rest of the tool — nothing here requires a separate login or account.

How It Works

The assistant doesn't rely on general AI knowledge to answer plant-specific questions — it retrieves the relevant passage from the actual source documents first, then answers from that passage, and shows its source. This is what keeps its answers accurate and checkable.

Your question
Find relevant documents
Answer grounded in what was found
Answer + cited source

At a very high level, every question — typed or spoken — follows the same four steps end to end:

  1. The question is sent to the agent (Agent Engine for text, or the in-app voice agent for speech).
  2. The agent searches the relevant document store(s) for the passages that best match the question, and — for recurring-failure questions — also pulls matching records from the breakdown/preventive maintenance data.
  3. The agent composes an answer using only what it retrieved, never from general knowledge.
  4. The answer is returned with its source(s) attached, so it can be checked against the original document.

Where it runs: everything runs on Google Cloud, inside Unilever's own cloud project — no plant data or documents leave that environment to answer a question.

The documents (embeddings)

Each source document is broken into passages and stored, along with a numeric representation of its meaning (an "embedding"), in Google Vertex AI's managed document-retrieval store — one store per document area (maintenance / operational / troubleshooting). This is what the assistant searches to find the passage most relevant to a question, before it ever generates a reply.

The assistant (agent)

The text assistant runs as a managed AI agent on Google Vertex AI Agent Engine, hosted separately from the rest of the application. The voice assistant runs on Google Cloud Run, alongside the dashboard and its API, since real-time voice needs a continuously-open connection rather than a request/response call.

The documentation the assistant draws from is organized into three areas, so it can pull from the right one for the type of question being asked:

Maintenance Operational procedures Troubleshooting

For questions about recurring failures ("why does the gripper sensor cable keep breaking?"), the assistant additionally looks at the historical breakdown log and the preventive-maintenance task list, so its answer reflects what has actually happened on this line — not just what the manuals say should happen.

Current Architecture

End to end, the PoC is five stages: the plant's own source records and documents go in, get prepared and indexed, are made searchable by a Google Cloud AI stack, and come back out through the dashboard and assistant a user actually sees.

1
PoC Sources

Selected knowledge documents, structured & unstructured intake

  • Breakdown records
  • Machine logs & fault categories
  • CIL task register
  • Manuals & SOPs
2
Data Preparation

Structured & unstructured intake

  • Breakdown/CIL records loaded into the reliability database
  • Manuals & SOPs indexed into the knowledge corpus
  • Metadata & filters — source, page, machine
  • Chunking with source metadata for text & tables
3
AI Stack

Google Cloud · Vertex AI, grounded reasoning

  • Vertex AI RAG Engine — corpus management & retrieval orchestration
  • Embeddings — multilingual, per-chunk
  • Vector search — high-scale similarity search
  • Gemini on Vertex AI — grounds every answer in retrieved context
4
Application Layer

What users see

  • PoC web application — connects users securely to the assistant & dashboard
  • AI Maintenance Assistant — troubleshooting, RCA, operational & maintenance Q&A
  • Unified Maintenance Dashboard — single view of status, incidents, schedules & performance
5
PoC Outputs

Handled automatically

  • Fault categories & historical incidents surfaced automatically
  • RCA evidence & analysis, preventive-maintenance tasks
  • Source citations with linked evidence — every match traced to its source, page & machine

Included knowledge

  • Breakdown / resolution history & fault categories
  • Operational & SOP guidance, operating & safety procedures
  • Maintenance schedules & preventive-maintenance (CIL) tasks
  • Fault codes, changeover & operating guides

Single view of

  • Breakdown / resolution history
  • Troubleshooting Q&A
  • Machine status
  • Performance insights

Using the Assistant

By text

Open the Assistant page, type a question, and get a written answer with source citations you can click to see the original document excerpt.

By voice

Say the wake phrase "Hello Maintain" to start talking. The assistant stays listening for the rest of the conversation — no need to repeat the wake phrase for follow-up questions — and it will let you know before it goes back to sleep after a period of silence.

What kind of questions can I ask?
Anything about the A11/A12 line's equipment, procedures, fault history, or preventive-maintenance tasks — e.g. "What's the CIL frequency for the heater controller?" or "Why does the gripper sensor keep failing?"
What happens if it doesn't know the answer?
It says so, rather than guessing — it only answers from what it can find in the plant's documentation and data, and declines questions outside that scope.
Can I trust the sources it shows me?
Yes — every citation links back to the real document and passage the answer was drawn from, so you can verify it yourself.

02 · A11 / A12 Reliability Desk

Data Model & ER Diagram

The tables behind the dashboard and assistant's data lookups — what's stored, and how it relates.

The dashboard and the assistant's data-driven answers are both backed by one shared data model — a small "star schema" of reference tables (lines, machines, sub-assemblies, fault categories, roles) around two record tables: every reactive-maintenance breakdown, and every preventive-maintenance (CIL) task. The same physical machine is sometimes named differently across the two source systems (e.g. "Pester" vs. "Wrapper") — the model resolves that with an alias, so both sources land on one consistent machine record.

DIM_LINE A11 / A12 DIM_MACHINE name + alias DIM_SUBASSEMBLY CIL only DIM_THEME fault category DIM_ROLE Operator/Electrician/Craft FACT_BREAKDOWN one row per reactive incident FACT_CIL one row per preventive task DIM_USER illustrative BRIDGE_USER_ROLE illustrative

Solid boxes and arrows are populated directly from the plant's own breakdown log and CIL task export. Dashed boxes are an illustrative access-control layer (who can view/edit this data) — not derived from plant data.

DIM_LINE

One row per production line.

ColumnKeyDescription
idPKGenerated ID, one per distinct line.
line_codeThe line code as it appears in both source files — "A11" or "A12".

Sample rows

idline_code
L1A11
L2A12

DIM_MACHINE

One row per physical machine on a line.

ColumnKeyDescription
idPKGenerated ID, one per physical machine per line.
line_idFKWhich line this machine sits on → DIM_LINE.
machine_namePrimary name — the two source systems sometimes spell the same machine differently.
aliasThe other source system's name for the same physical machine — e.g. "Pester"/"Wrapper", "Deo"/"Deo Filler", "Coder"/"Inkjet".

Sample rows (5 of 38 real machines)

idline_idmachine_namealias
M30L1PesterWrapper
M13L2DeoDeo Filler
M22L1LineGeneral
M2L2AP Filler
M12L2Crimper

DIM_SUBASSEMBLY

One row per sub-assembly of a machine — present only in the preventive-maintenance (CIL) data.

ColumnKeyDescription
idPKGenerated ID, one per sub-assembly per machine.
machine_idFKWhich machine this sub-assembly belongs to → DIM_MACHINE.
subassy_namee.g. "Outfeed conveyor", "Capper Outfeed", "Change Parts".

Sample rows (6 of 521 real sub-assemblies)

idmachine_idsubassy_name
S11M30Outfeed conveyor
S24M30Air Motor
S12M30Infeed Conveyor x 4 Lane Conveyor
S10M9Capper Outfeed
S6M17Change Parts

DIM_THEME

One row per fault category used on the Reactive Maintenance dashboard.

ColumnKeyDescription
idPKGenerated ID, one per fault category.
theme_nameOne of six categories: Safety & guarding, Electrical & controls, Software & comms, Pneumatic & fluid, Mechanical, Other/unclassified.
keyword_rulesThe keyword-matching rule used to assign a breakdown's free-text description to this category.

All 6 rows

idtheme_name
T1Electrical & controls
T2Mechanical
T3Other / unclassified
T4Pneumatic & fluid
T5Safety & guarding
T6Software & comms

DIM_ROLE

One row per job function. Used both by preventive tasks (who is qualified to do the work) and, illustratively, by system access control.

ColumnKeyDescription
idPKGenerated ID.
role_nameOperator, Electrician, or Craft — the three job functions used across the preventive-maintenance task list.

All 3 rows

idrole_nameusage
R1Operator637 of 1,466 CIL tasks require this role — the most common.
R2Electrician422 of 1,466 CIL tasks.
R3Craft407 of 1,466 CIL tasks.

FACT_BREAKDOWN

One row per reactive-maintenance incident — powers the Reactive Maintenance dashboard tab.

ColumnKeyDescription
idPKOne row per incident.
date / year / monthWhen the breakdown occurred.
line_id / machine_id / theme_idFKWhich line, machine and fault category this incident belongs to.
breakdownFree-text description of the fault symptom.
resolutionFree-text description of how it was resolved.

Sample rows (2 of 222 real incidents)

iddatemachine_idtheme_idbreakdownresolution
BD12026-03-05M30T5 Panel had been blown out and cardboard dust blew up a power supply. On power supply replacement guard door circuit would not reset. Relay auxiliary sticking and Sick safety module needed a complete power-down to reset.
BD102025-02-28M18T3 Overfill setting on A11 Heuft was not rejecting the overfilled test can. Changed setting from 227 to 190. Tested OK.

FACT_CIL

One row per preventive-maintenance (cleaning/inspection/lubrication) task — powers the Preventive Maintenance dashboard tab.

ColumnKeyDescription
idPKOne row per task.
line_id / machine_id / subassy_id / tbc_by_role_idFKWhich line, machine, sub-assembly and required role this task belongs to.
last_done / next_dueWhen the task was last completed, and when it's next due.
frequency_daysHow often the task recurs, in days (0.5 = twice daily).
statusOverdue / due soon / on track — computed live from next_due, unless a manual override has been set for that task.
task_description / reasonWhat the task is, and why it's done (e.g. Housekeeping, Minor Stops, Product Quality).
people_required / running_static / expected_min / annual_minStaffing and time-on-task figures for the task.
completed_by / ppe_needed / tools_and_products / procedureWho typically completes it, and what's needed to do so.

Sample rows (3 of 1,466 real tasks)

idmachine_idtask_descriptionreasonnext_duestatus
CIL7M17 Clean all change parts using all purpose cleaner and a dry cloth. Housekeeping (5S)2026-08-10overdue
CIL869M17 Conduct a basic panel check — water/dust ingress, panel secure, no foreign objects. Breakdowns2026-11-17on_track
CIL1045M9 Check all push buttons / lamps / indicators are working correctly. Breakdowns2026-09-15due_soon (manual override: "Completed")

DIM_USER & BRIDGE_USER_ROLE (illustrative)

An access-control layer for who can view/edit this data in the tool itself — not derived from the plant's own maintenance data.

TableColumnKeyDescription
DIM_USERidPKOne row per user account.
DIM_USERuser_name / email / activeAccount details.
BRIDGE_USER_ROLEuser_id, role_idPK FKComposite key — lets one user hold multiple roles → DIM_USER, DIM_ROLE.
BRIDGE_USER_ROLEassigned_dateWhen the role was assigned.

Illustrative sample rows (not real accounts)

iduser_nameemailactive
U1J. Smithj.smith@example.comtrue
U2R. Patelr.patel@example.comtrue
user_idrole_idassigned_date
U1R12026-09-01
U1R32026-09-01
U2R22026-08-15

03 · A11 / A12 Reliability Desk

Evaluation Report

An 80-question accuracy run against the assistant, with every result and every failure explained.

A set of 80 realistic questions — covering safety procedures, maintenance tasks, changeover planning, drive troubleshooting, and root-cause/breakdown-history analysis — was run against the assistant, and every answer was checked against a verified reference answer from the source documents and data.

80
Total Questions
67
Passed
13
Failed
83.8%
Overall Accuracy
8–45s
Avg. Latency

Accuracy by category

Operations & Safety
16 / 18 · 88.9%
Maintenance
13 / 17 · 76.5%
Planning & Changeover
13 / 14 · 92.9%
Troubleshooting
6 / 7 · 85.7%
RCA & Breakdown History
19 / 24 · 79.2%
Where accuracy is weakest — RCA & Breakdown History (5 of 13 fails): mostly false negatives, where a real matching record existed but the data lookup didn't surface it, plus two cases of the assistant drawing conclusions from too small a slice of the historical data instead of the full record set. See the Failure Analysis tab for each case.

All 80 test questions, their category, and the pass/fail outcome. Failure reasons are summarized here — see the Failure Analysis tab for the full detail on each.

#CategoryQuestionResultNotes
1Operations & Safetywhere are the estop buttons on this machinePassMatches documented reference answer.
2Operations & Safetywhats the door release button for near doors 1 and 6PassMatches documented reference answer.
3Operations & Safetyif the guard door lamp is on but the hmi still shows closed, am i ok to startPassMatches documented reference answer.
4Operations & Safetygas alarm cleared, how do i restart from the main panel, step by stepPassMatches documented reference answer.
5Operations & Safetywhos supposed to investigate a gas alarm, me or the electricianPassMatches documented reference answer.
6Operations & Safetydo i need to fully evacuate for every gas alarm or just somePassMatches documented reference answer.
7Operations & Safetybefore i open a guard door what am i supposed to check on the hmiPassMatches documented reference answer.
8Operations & Safetycan i put my whole body inside the machine guard to fix something quickPassMatches documented reference answer.
9Operations & Safetytheres a job that isnt covered by any written procedure, what do i doPassMatches documented reference answer.
10Operations & Safetygas card is reading 4, is that ok to restart onPassMatches documented reference answer.
11Operations & Safetywho approves changes to the sop document itself, and is that the same as the document ownerFailReal SOP cover page has the owner field blank - only 'Approved by: Process Engineer' is filled. Answer asserts an equ...
12Operations & Safetycan new operators use this sop as their only training or do they need someone with themPassMatches documented reference answer.
13Troubleshootingwhats a safety open fault on the drive and whats the first thing i should checkPassMatches documented reference answer.
14RCA & Breakdown Historywe've got a pester guard door 12 alarm that wont reset even after craft used the override switch, whats fixed this beforePassMatches documented reference answer.
15RCA & Breakdown HistoryA12 pester keeps throwing false RHS safety switch faults, we've had this before right, what did they findPassMatches documented reference answer.
16RCA & Breakdown Historyhold-down motor's throwing a bus fault on A11 pester again, is it actually the motor or the same issue as beforePassMatches documented reference answer.
17RCA & Breakdown Historyjust had an axis A pac robot exceeded limit alarm on the pester, then axis B did the same right after, what did we do last timePassMatches documented reference answer.
18RCA & Breakdown Historysplicer 1 ropex temp controller's in alarm and wont reset on A11 pester, whats the fix from last timePassMatches documented reference answer.
19RCA & Breakdown Historyis there a logged 'category 5 gearbox' failure on the pester pick and place, or have we not seen thatFailReal matching incident exists (ID450, gearbox bearing seizure) - false negative on SQL lookup.
20RCA & Breakdown Historylast time we had a coder EHT fault trip repeatedly, what did we end up doing, swap the whole thing or just clean itFailReal breakdown record exists describing repeated EHT trips and ink leaks - false negative.
21RCA & Breakdown Historyout of the 29 breakdown records analysed, whats the biggest category and whats the actual repeat theme behind itFailSource document's own stated breakdown shows Mechanical (13 records, 45%) as the clear biggest category - answer invented ...
22RCA & Breakdown Historywhats our single biggest downtime hit across all the breakdowns thats been logged, and how long did it cost usFailDoesn't answer the question asked - repeats the wrong 'biggest category' answer instead of stating a downtime du...
23RCA & Breakdown Historyany record of us actually having a phase-to-ground fault, like F038 or F039, happen on the pester specificallyPassMatches documented reference answer.
24Maintenancehow often do the pester sealing unit seals actually need changing, and whats the colour setupPassMatches documented reference answer.
25Maintenancewhat tools do i need before i check the toothed belts on the pesterPassMatches documented reference answer.
26Maintenancewhat do i need to grab before doing a seal change on the sealing unitPassMatches documented reference answer.
27Maintenancewhats involved in isolating the machine before i start the toothed belt checkPassMatches documented reference answer.
28Maintenancebefore i fit the axis zeroing jig, do i need to calibrate anything firstPassMatches documented reference answer.
29Maintenancewhich axis do i zero first on the pester axis zeroing jig, and is there anything to be careful ofPassMatches documented reference answer.
30Maintenanceafter fitting new seals on the sealing unit, how do i know the job's actually good, and any tips for getting them inFailResponse contained an unrelated sealing-bar validation answer instead of B-axis speed settings - answer/question misalignment.
31Maintenancewhats the actual jogging procedure for zeroing the b axis, like the speed settingsPassMatches documented reference answer.
32RCA & Breakdown Historywe're getting double packs or a bad seal seam after fitting new seals on the sealing unit, whats the known causePassMatches documented reference answer.
33Maintenancewhats the smp procedure for cleaning the probe and coil on the pesterPassMatches documented reference answer.
34Maintenancewhats the procedure for the thermography check on the servo motor drivesPassMatches documented reference answer.
35Planning & Changeoverwhat film width do we run for the posh cap 250ml formatPassMatches documented reference answer.
36Planning & Changeoverwhats the collation pattern for the carla cap 150ml format, like how many go in a packPassMatches documented reference answer.
37Planning & Changeoverwhats the product dimension for polaris 100ml, just the can sizePassMatches documented reference answer.
38Planning & Changeoveroutside case dimensions for the carla cap 150ml formatPassMatches documented reference answer.
39Planning & Changeoverdoes the incoming speed at infeed change between formats or is it always the samePassMatches documented reference answer.
40Planning & Changeoverwhats the drawing number used for all these formatsFailReal answer: film width does change (190mm to 220mm). Repeat of a previously-flagged failure on this exact question.
41Planning & Changeoverwhich formats actually run at a different film width than most of the othersPassMatches documented reference answer.
42Planning & Changeoveris the collation ever anything other than 2x3 across all the formats we runPassMatches documented reference answer.
43Planning & Changeoverwhats the output rate at discharge, and does it change by formatPassMatches documented reference answer.
44Planning & Changeoverwhat machine types is this format list actually forPassMatches documented reference answer.
45Planning & Changeovercompare the outside case dimensions for carla cap 150ml and posh cap 250mlPassMatches documented reference answer.
46Planning & Changeoverhow many of the format settings actually change going from format 1 to format 2 on the changeover list, roughlyPassMatches documented reference answer.
47Planning & Changeoverwhats the outside case dimensions for vic 250ml, straight off the format list tablePassMatches documented reference answer.
48Operations & Safetywhat does the risk assessment say the residual risk rating is for the guard interlock systemPassMatches documented reference answer.
49RCA & Breakdown Historyhow many breakdown records are in the full A11/A12 logPassMatches documented reference answer.
50Operations & Safetyis the wrapper the same thing as the pester machine on this linePassMatches documented reference answer.
51Maintenancewhy do we have to isolate both pneumatically and electrically before the toothed belt check, isnt electrical isolation on its own enoughFailContradicts the real documented reason (compressed air stays pressurized after electrical isolation) - reads as a fabricated te...
52MaintenanceWhat's the procedure for zeroing the A/B axis on the Pester?PassMatches documented reference answer.
53MaintenanceHow do I replace the sealing unit seals on A12?PassMatches documented reference answer.
54MaintenanceWhat does the thermography check on the Pester servo drives involve?PassMatches documented reference answer.
55MaintenanceHow do I clean the Pester temperature probes and heater coils?PassMatches documented reference answer.
56MaintenanceWhat tools are needed to replace the conveyor link on the infeed belt?FailOnly lists tools; omits the isolation and post-repair function-check steps the documented SMP procedure requires.
57MaintenanceHow is the missing-cap sensor set up on A11/A12?FailDescribes the sensor's general location/function but gives no calibration steps - wrong document/agent used, not the actua...
58Operations & SafetyWhat does the Pester Pick & Place area's Gripper Head Safety Sensor do?FailDeclines the actual question, then substitutes an answer about a different, unrelated sensor (Safety Laser Scanner).
59Operations & SafetyHow long does the Therm tunnel take to reach operating temperature during startup?PassMatches documented reference answer.
60Planning & ChangeoverWhat film width is used for the 'Carla Cap 150ml' product changeover?PassMatches documented reference answer.
61Operations & SafetyWhat should I check before starting the machine according to the startup SOP?PassMatches documented reference answer.
62Operations & SafetyWhat is documented about disconnecting electrical energy on the PEWO-pack 600?PassMatches documented reference answer.
63TroubleshootingWhat does fault code F004 mean on the PowerFlex 525 drive?PassMatches documented reference answer.
64TroubleshootingWhat does fault code F005 indicate and what should I check?PassMatches documented reference answer.
65TroubleshootingWhat does the DC Bus Voltage parameter (b005) show on the drive?FailStates b005 shows 'Process Fract' - this is incorrect; b005 is the DC Bus Voltage reading. Retrieved the right docume...
66TroubleshootingWhat causes an F007 Motor Overload fault and how do I resolve it?PassMatches documented reference answer.
67TroubleshootingWhat does an F013 Ground Fault mean on the drive?PassMatches documented reference answer.
68TroubleshootingWhat does fault code F029 Analog In Loss mean?PassMatches documented reference answer.
69RCA & Breakdown HistoryHow many breakdowns has A11 had in total?PassMatches documented reference answer.
70RCA & Breakdown HistoryHow many breakdowns has A12 had in total?PassMatches documented reference answer.
71RCA & Breakdown HistoryHow many breakdown records exist for the Pester machine?PassMatches documented reference answer.
72RCA & Breakdown HistoryHow many breakdown records exist for the Capper machine?PassMatches documented reference answer.
73RCA & Breakdown HistoryWhich machine has the most CIL preventive-maintenance tasks?PassMatches documented reference answer.
74RCA & Breakdown HistoryHow many CIL tasks are marked Safety Critical?PassMatches documented reference answer.
75RCA & Breakdown HistoryHow are CIL tasks split across Operators, Electricians, and Craft?PassMatches documented reference answer.
76RCA & Breakdown HistoryHow many CIL tasks fall under the 'Breakdowns' reason category?PassMatches documented reference answer.
77RCA & Breakdown HistoryWhy does the Pester gripper keep breaking down?PassMatches documented reference answer.
78RCA & Breakdown HistoryWhy does the Pester tunnel keep tripping into emergency exhaust?PassMatches documented reference answer.
79RCA & Breakdown HistoryWhy does Capper keep breaking down?PassMatches documented reference answer.
80RCA & Breakdown HistoryWhat's the most common root cause of breakdowns on the Kuka robot?FailOnly 1 incident used to build a full RCA conclusion, when the Kuka robot has significantly more breakdown records in the full l...

All 13 failed test cases, with a recommended fix for each.

#CategoryQuestionWhy it failedRecommended fix
1Operations & Safetywho approves changes to the sop document itself, and is that the same as the document ownerThe response states the approver 'is also identified as the document owner' - this specific equivalence is not written anywhere in the source. The owner field being blank and the approver field being filled are two separate facts; treating a blank field as 'the same as' a filled one is an inference the document does not support.Restrict the answer to facts explicitly stated per field; don't let the model infer that two separate document fields (approver, owner) refer to the same person.
2RCA & Breakdown Historyis there a logged 'category 5 gearbox' failure on the pester pick and place, or have we not seen thatThe response states no matching records exist. A real record does exist for this machine and fault type (gearbox); it was likely missed because the query searched for the literal fake term 'category 5' rather than the actual fault description.Match breakdown-log searches on the described symptom (e.g. "gearbox", "bearing") rather than a literal invented category label; add fuzzy/synonym matching to the SQL lookup tool.
3RCA & Breakdown Historylast time we had a coder EHT fault trip repeatedly, what did we end up doing, swap the whole thing or just clean itThe response states the data-gathering step returned nothing usable for this question. A real, directly matching record exists in the same source document used successfully for other questions in this set.Same fix as #2 - broaden the breakdown-log search so a real matching record (by machine + fault keyword) isn't missed when the exact phrase isn't in the text.
4RCA & Breakdown Historyout of the 29 breakdown records analysed, whats the biggest category and whats the actual repeat theme behind itThe response reports a 3-way tie at 7 incidents each between different category names (Robot/Gripper/Axis, Safety Interlocks, Heating/Temperature) - this doesn't match the document's own stated category breakdown at all. It appears to have invented its own grouping instead of reading the document's existing Category Pareto table.Have the RCA tool read the source document's own existing summary table (e.g. "Category Pareto") directly instead of re-deriving a category breakdown from raw records.
5RCA & Breakdown Historywhats our single biggest downtime hit across all the breakdowns thats been logged, and how long did it cost usThe response doesn't answer the question at all - it repeats the same wrong 3-way-tie category answer from the previous question and never states a downtime duration. The real answer (480 minutes, a specific named event) is explicitly stated in the source document's own summary section.Add an answer-relevance check so the agent verifies it actually answered the question asked (a duration) before returning a reused answer from the prior turn.
6Maintenanceafter fitting new seals on the sealing unit, how do i know the job's actually good, and any tips for getting them inThe response describes checking the gap between a sealing bar and clamping bar (0.2-0.5mm) and spraying a silicone separating agent before insertion - this is content from the general Translation manual's sealing-BAR servicing section, a different component and different document to the one this SMP-specific question needed.Improve document disambiguation between similarly-named procedures (sealing bar vs. sealing unit) - tighter chunk labeling/metadata so retrieval doesn't cross-match the wrong component.
7Planning & Changeoverwhats the drawing number used for all these formatsThe response gives "21385-10000" - this is the Changeover List's own project/document reference number (from its title "Changeover list: 21385-10000"), not the Format List's drawing number that the question asked for. Two different documents' reference numbers were confused.Same fix as #6 - tag retrieved passages with their source document name and prefer an exact document-title match for numeric/reference-number lookups.
8Maintenancewhy do we have to isolate both pneumatically and electrically before the toothed belt check, isnt electrical isolation on its own enoughThe response states electrical isolation 'causes the air system to exhaust via an air dump valve' and that pneumatic isolation is only about parts 'coming to rest in a safe position' - this is the opposite of what the source says. The real hazard is that air pressure is NOT automatically exhausted by electrical isolation alone; it stays stored, which is exactly why pneumatic isolation is a separate required step.Increase retrieval coverage (top-k) for safety-critical sections and add a fact-check pass comparing the generated answer against the retrieved passage before returning it.
9MaintenanceWhat tools are needed to replace the conveyor link on the infeed belt?The tools listed are correct and properly cited. It's marked as a fail because the question is being graded against the full expected procedure, which also requires isolating the machine first and checking the repair afterward - both omitted here.Prompt tuning: for procedure questions, require the full documented step sequence (isolation, task, verification) rather than stopping at the first matching sub-section (tools).
10MaintenanceHow is the missing-cap sensor set up on A11/A12?The response describes the sensor's general location and function (mounted above the scroll, stops the machine if a cap is missing) - this is accurate as far as it goes, but it's a general SOP description, not the specific SMP calibration steps the question asked for. Wrong document was used.Same fix as #6/#7 - improve document routing so calibration-specific SMPs are preferred over general SOP descriptions when the question asks "how is X set up".
11Operations & SafetyWhat does the Pester Pick & Place area's Gripper Head Safety Sensor do?The response says it doesn't have information on this specific sensor, then substitutes a description of a completely different sensor (the Safety Laser Scanner, from the Machine Safety section) as if it were relevant. The real Gripper Head Safety Sensor is documented elsewhere in the same SOP and was not surfaced.Add a guardrail against substituting a different, unrelated entity when the exact one asked about isn't found - the agent should say it doesn't know rather than answer about something else.
12TroubleshootingWhat does the DC Bus Voltage parameter (b005) show on the drive?The response states b005 shows "Process Fract" - a different parameter/value entirely. This is a direct factual contradiction, not just an incomplete answer: the right document was retrieved, but the wrong field or row was read from it.Improve table/row extraction accuracy for tabular fault-code documents so retrieval returns the correct field value, not an adjacent one.
13RCA & Breakdown HistoryWhat's the most common root cause of breakdowns on the Kuka robot?The response builds its entire root-cause conclusion from just 1 incident record, when 29 real records exist for this machine. This is a severe data undercount - the SQL query likely returned a small, unrepresentative slice of the real data, and the RCA synthesis presented that single record as if it were the complete picture.Fix the SQL query's row limit/aggregation for RCA lookups so root-cause synthesis is built from the full matching record set, not a small unrepresentative slice.

04 · A11 / A12 Reliability Desk

Scope & What's Next

What this proof of concept covers today, and what's coming next.

This is a proof of concept scoped to the A11/A12 "Pester" shrinkwrapper line at the Leeds site. It's built to demonstrate the approach — grounded, cited, voice-and-text assistance layered on top of a real reliability dashboard — rather than as a finished, plant-wide rollout.

  • Covers A11/A12 breakdown history and preventive-maintenance records currently available for this line.
  • Answers are only as complete as the documentation and data provided — gaps in the source material mean gaps in what the assistant can answer.
An 80-question accuracy evaluation has been run against this build — see the Evaluation Report page for the full results, category breakdown and failure analysis.

How this scales beyond one line

This PoC proves the pattern on a single line (A11/A12). The same three workflow types generalize to any line or plant, because none of them are hard-coded to A11/A12 — the agent, the retrieval layer and the write-back logic are all driven by configuration (which corpus, which database schema, which line) rather than line-specific code. Scaling to more lines or sites means re-pointing the same architecture at more data, not rebuilding it.

1

Fix & Close a Maintenance Issue

Resolve faster. Close with confidence.

1
Ask
Technician asks by voice or text, any language
2
Classify & Route
Question is classified and routed to the right agent
3
Agent Answers
Combines SOPs, breakdown history and live context
4
Evidence & Action
Grounded answer with source, confidence and next step
5
Technician Confirms
Answer is complete and the action is approved
6
Write Back
Updates the maintenance record and closes the loop
7
Logged & Visible
Audit trail and cross-line insights over time
Outcome: Faster resolution, fewer repeat callbacks, and a closed loop with full auditability.
2

Predict & Prevent Failure

Prevent downtime. Plan with confidence.

1
Sense
Machine telemetry and sensor data, where available
2
Predict
Failure-risk scoring from history and live signals
3
Risk Assessment
Agent explains drivers, impact and confidence
4
Scheduling
Agent proposes the optimal maintenance window
5
Approval
One-tap approval in the manager workspace
6
Task Created
Preventive task created and assigned to the team
7
Monitor & Feedback
Execution tracked, outcome captured, model improves
Outcome: Reduced unplanned downtime and longer asset life through proactive maintenance.
3

Investigate & Learn Across Lines

Learn once. Improve everywhere.

1
Ask Across Lines
A trend or root-cause question, asked in plain language
2
Cross-Line Search
RCA agent searches incidents, SOPs, logs and patterns
3
Correlate
Agent correlates across lines and time
4
Explain
Root-cause explanation with an evidence trail
5
Approve Action
Manager approves the recommended improvement
6
Roll Out
Update published as a new SOP / work standard
7
Govern & Measure
Impact, compliance and continuous improvement tracked
Outcome: Learning gets institutionalized and performance becomes consistent across every line.

Shared architecture foundation (one across every line)

Every line runs the same underlying stack — only the data underneath changes. That is what makes rollout to a new line fast: no new agent, no new integration pattern, just a new corpus and a new database connection plugged into the same architecture already proven here.

Enterprise Systems & Sources

Maintenance/CMMS records, SOPs, manuals, breakdown logs, sensor data, more

AI-Ready Data Products

Clean, governed, structured data with quality and lineage

Context & Semantic Layer

Per-line document corpora, classification, multi-language normalization

Agent Layer / Shared Intelligence

Troubleshooting, RCA, scheduling, action agents — same code across lines

MaintAIn Control Tower

Operator, technician, manager and executive workspaces

Built once. Reused everywhere. Continuously improved.

AI OPS & GOVERNANCE — ACROSS EVERY LAYER
RBAC & line isolation
Human-in-the-loop approvals
Action audit trail
Model & agent monitoring
Cost & token optimization
Deployment & versioning