Skip to content

Manufacturing and industrial operations

AI for plants that ship good parts, shift after shift

We work across the whole span of a plant: from the two-minute micro-stop nobody logs and the stack light with no controller behind it, to the OEE charter two shifts still argue about and the schedule your planner rebuilds by hand every morning. Most of it is measurement before modelling — what counts as a stop, where the ideal cycle time came from, and which asset is actually the constraint.

Book a callSee our projects →

30 minutes with the engineer who would do the work — not a salesperson. No obligation, and you keep whatever we work out on the call.

Machine operator reading a tablet at a control panel on the production floor

The numbers this sector is run on

45–65%
Typical discrete-plant OEE, against a world-class 85%The 85% figure is Availability 90% × Performance 95% × Quality 99.9% — Seiichi Nakajima’s TPM benchmark for 1970s–80s Japanese automotive assembly. Vorne’s own published OEE dataset — a vendor’s, and worth reading as such — shows more plants below 45% than above 85%.
60–70%
Share of downtime minutes manual logging actually capturesStops under five minutes are skipped almost entirely, so manually reported OEE is commonly overstated by 10–25 percentage points; the drop you actually see at go-live is usually 10–20, because some of the gap is speed loss you recover immediately. Automated capture should clear 95%, with unattributed time under 5% of planned production time.
κ ≥ 0.75
Attribute agreement an inspection cell needs to enter a control planIATF 16949 §7.1.5.1 treats a vision system as a gauge. A pass/fail cell is an attribute gauge, so it is qualified by attribute agreement analysis per AIAG MSA-4 — kappa ≥ 0.75, effectiveness ≥ 90%, miss rate ≤ 2%, false-alarm rate ≤ 5% — not by a %GRR study. The <10% GRR bar applies where the cell measures a dimension; there the capability targets are Ppk ≥ 1.67 on the initial study and Cpk ≥ 1.33 ongoing.
95%+
Master-data accuracy an APS go-live actually needsThe practical bar on BOMs, routings, setup times, capacity definitions and inventory records. An excellent optimiser running on 85%-accurate routings produces a schedule the planner overrides by lunchtime.

What we build

Operator walking between rows of machinery on a factory floor

The 2026 plant reality we build against

Bulgaria adopted the euro on 1 January 2026, so every plant here has just recosted its ERP and rebased its standard costs — any 2025 cost-per-unit or energy baseline has to be restated before it can be compared with anything we measure, and we do that restatement first rather than quietly comparing across the changeover. The arithmetic that makes automating inspection and data capture pay here is labour: NSI put the average gross wage at EUR 1,475 in March 2026, up 12.9% year on year, with 32.4% of industrial enterprises naming labour shortage as a factor limiting activity in June 2026. And the discrete base we mostly address is the automotive-electronics cluster around Plovdiv (Trakia Economic Zone), Sofia and Botevgrad — plants shipping to German and French OEMs under IATF 16949, VDA 6.3 audits and Q-DAS data formats.

  • Any pre-2026 cost or energy baseline is restated into euro before it is used as a comparison
  • Inspection and capture cases are priced against your current wage bill, not a 2023 one
  • Sources: NSI wage and business-survey releases (March and June 2026), Bulgarian Investment Agency sector data, КЕВР for network and capacity tariffs
Two supervisors reading production numbers at a shop-floor terminal

Measurement two shifts can agree on

Before any dashboard, a written OEE charter aligned to ISO 22400: what counts as planned production time, whether a changeover is planned loss, the micro-stop threshold, and where the ideal cycle time came from. Capture comes off PLC tags, stack-light phases and CT clamps rather than a paper log, and unattributed time is published as a KPI in its own right.

  • Eight to twelve reason codes to start, maturing to 15–30 per asset area
  • “Other” held under 10% of logged minutes, or the Pareto cannot be trusted
  • TEEP alongside OEE when the real question is whether to buy another machine
Quality inspector measuring a metal component under inspection light

Inspection treated as a gauge, not a demo

Lighting, fixturing and an enclosure come before any model, because illumination variation causes more false rejects than any architecture choice. Unsupervised anomaly detection trains on good parts only, since no line arrives with a labelled defect library, and moves to supervised segmentation once the verification station has produced labels. Then MSA — attribute agreement for a pass/fail cell, gauge R&R for a measuring one — so it can enter an IATF 16949 control plan.

  • PatchCore or PaDiM via Anomalib; Cognex, HALCON or a Jetson Orin edge stack by cycle time and by who maintains it
  • Escape rate and false-reject rate budgeted and tuned as two separate numbers
  • Re-validation scheduled against the product-variant calendar, not treated as an incident
Maintenance engineer fitting a vibration sensor to a motor housing

Condition monitoring that ends in a coded work order

Criticality ranking and FMECA first, then physical indicators: ISO 20816-3 velocity zones and envelope bands computed from the real bearing part numbers, with motor current signature analysis where a sensor cannot be fitted safely. Machine learning is added only where no physical indicator exists. Alerts arrive in Maximo, SAP PM or Limble as notifications carrying an ISO 14224 failure mode — never as email to a shared inbox.

  • P-F interval checked against spare-part lead time before a single sensor is quoted
  • Failure mode, mechanism and cause kept as three fields, not collapsed into one
  • If the honest answer is to change the stocking policy instead, that is the recommendation
Automation engineer connected by laptop to an open control cabinet

A data layer your OT team and your works council both accept

OPC UA where a server exists, S7comm or Modbus TCP where it does not, normalised to MQTT Sparkplug B with birth certificates and report-by-exception. A modelling layer applies ISA-95 context so a metric means the same thing on every line and at every site. Network design follows IEC 62443 zones and conduits with a real OT DMZ and outbound-only flow, and operator-linked data is treated as personal data from day one.

  • The collector cannot stop your line — reads are passive and outbound-only, and if the gateway dies production continues while we lose data, not parts
  • Scan-time headroom confirmed before any subscription touches a production controller
  • The durable asset is the schema — use cases two and three cost a fraction of the first
  • Cameras and operator performance records get a lawful basis and consultation before the pilot

Worked examples

These are designs, and the ranges we would contract against, drawn from published sector data — not Palamed results. The four systems we have actually delivered are on case studies.

Automated OEE and downtime capture on brownfield lines

Problem

Shift reports are reconstructed from memory at the end of the shift. The plant reports 78% OEE while the constraint runs nearer 58%. Micro-stops and reduced-speed running are invisible, changeovers are classified as planned on one line and unplanned on the next, and two shifts describe the same event differently.

Approach

Tap what already exists before buying hardware: PLC tags over OPC UA where a server exists, S7comm, Modbus TCP or EtherNet/IP where it does not, stack-light phase inputs, and a CT clamp on the main feed for machines with no controller access at all. An edge gateway — Ignition with Cirrus Link, Kepware, or Node-RED — normalises everything to MQTT Sparkplug B. Definitions are frozen first in a written OEE charter aligned to ISO 22400: planned production time, changeover classification, a two-minute micro-stop threshold, and an ideal cycle time re-baselined from a timed capability run rather than the nameplate. Operators get eight to twelve reason codes on a shop-floor terminal, mandatory above threshold.

What we would target

The target we design against: downtime capture from roughly 60–70% of minutes to over 95%, with “Other” below 10%. Good count is the harder half and we scope it explicitly: scrap and rework are captured as separate quantities at the machine — never netted into one Quality figure — because rework hidden inside Quality is how an OEE number stays flattering. Where a brownfield line genuinely cannot produce good count at the asset, we say the Quality term is estimated and publish it as estimated rather than pretending otherwise. Reported OEE usually falls 10–20 points at go-live because the measurement finally became honest — we pre-sell that to the plant manager so it does not kill the project in week three. The top three Pareto losses on the constraint are then typically worth 4–8 points within two quarters.

Deep-learning visual inspection replacing a manual final check

Problem

Two inspectors per shift on 100% visual inspection, escapes in the hundreds to low thousands of PPM, and a rules-based AOI that over-rejects on glare and warpage badly enough that operators have learned to bypass it.

Approach

Fix the physics before the model: fixturing, telecentric, dome or darkfield lighting chosen for the defect class, strobing, and an enclosure that removes ambient variation. This is consistently the fastest and highest-return change. Then unsupervised anomaly detection — PatchCore or PaDiM via Anomalib — trained on good parts only, because no line arrives with a labelled defect library; migration to supervised segmentation follows once 100–300 labelled defects have accumulated from an operator verification station that adjudicates every AI reject. Deployment on Cognex VisionPro and ViDi, MVTec HALCON, or an OpenVINO and Jetson Orin edge stack, chosen on cycle time and on who will maintain it. Then MSA, because the cell is a gauge: attribute agreement analysis where the verdict is pass/fail, gauge R&R where the cell actually measures a dimension.

What we would target

The target we design against is an explicit budget — escape below 50 PPM at a false-reject rate under 1.5%, measured on your own line before and after. Against it, inspectors come off one or two shifts and customer 8Ds on the inspected defect class fall measurably. Every new variant and every lighting change triggers re-validation; that is scheduled work in the runbook, not a surprise.

Condition monitoring on critical rotating assets

Problem

Gearboxes, pumps and fans are changed reactively. Spare lead times run six to twelve weeks. The last unplanned gearbox failure took the line out for fourteen hours, and the maintenance manager can name the machine but not the date.

Approach

Criticality ranking by consequence and frequency first, then FMECA on the top twenty assets. Triaxial accelerometers and temperature on drive and non-drive ends where the asset justifies it; motor current signature analysis read at the MCC for assets that cannot be sensored safely. Alarms sit on ISO 20816-3 broadband velocity zones — classified by machine group and by rigid or flexible support, because that is what sets the zone boundaries — plus envelope bands computed from the actual bearing part numbers, so BPFO and BPFI energy is watched directly rather than inferred by a model. On gearboxes the 10–1000 Hz velocity band is the wrong measurand for the gear set: mesh frequency and sidebands are tracked in acceleration, and rising broadband velocity is the trigger to open a spectrum, not the diagnosis. Unsupervised anomaly detection is layered on only where no physical indicator exists. Every alert routes into Maximo, SAP PM or Limble as a notification carrying an ISO 14224 failure mode and a recommended action.

What we would target

The number that decides renewal is not accuracy — it is whether the P-F interval reliably exceeds the spare-part lead time on the assets that matter. Where it does, the target we design against is a known-bad asset converting from a fourteen-hour unplanned outage into a planned four-hour intervention on a Saturday, with the planned-to-reactive work ratio moving from roughly 40:60 toward 75:25 — measured against your own twelve months of CMMS history on the same functional locations.

Finite-capacity scheduling with sequence-dependent changeovers

Problem

The planner runs the shop from a spreadsheet. MRP proposes infinite capacity, so the plan is resequenced by hand every morning around colour, coating, tool-family or material-grade purge constraints. OTIF sits around 80% and changeover eats close to a fifth of available time.

Approach

Master data first, and we will sell you the remediation instead of the optimiser if the assessment says so: routings, work-centre calendars, and a sequence-dependent setup matrix built from historical changeover durations rather than a flat “standard setup 45 min”. Modelled in Opcenter APS, PlanetTogether or Asprova where the constraint set is conventional, and in OR-Tools CP-SAT where it genuinely is not — colour-run direction, tool-life windows and shared-fixture constraints usually are not. Real order and downtime status flows back from MES so the plan reflects the floor within the shift.

What we would target

Output is a dispatch list per work centre the operator can act on, plus a frozen horizon the planner is not allowed to churn inside. The target we design against: changeover minutes down 20–40% from sequencing alone. OTIF moves into the low 90s only to the extent your late orders are capacity-driven — so the first thing the assessment does is split twelve months of late orders into material-shortage, capacity and quality-hold causes. If most of yours are material, an APS will not fix them and we will say so before you buy one. The failure mode avoided here is an excellent optimiser running on 85%-accurate routings.

Energy per unit and billing-demand management under EED Article 11

Problem

One main meter, one monthly bill, energy booked as overhead. Nobody can say which SKU is expensive to make, what runs at the weekend, or which asset sets the billing demand peak — and the plant now also needs auditable evidence for the energy audit or ISO 50001 obligation.

Approach

Submeter the significant energy uses — compressors, chillers, ovens, injection machines, HVAC — with Janitza UMG, Schneider PM or Siemens SENTRON devices on Modbus TCP. Join the kW series to MES order and downtime context so kWh lands on the work order and produces specific energy consumption per unit and per SKU. Baseload analysis across non-production hours finds what nobody switches off. An ultrasonic survey to ISO 11011 quantifies each compressed-air leak in l/s (Nm³/h) and euros rather than listing them, and pressure-band optimisation removes artificial demand — every unnecessary bar costs roughly 6–7% more compressor energy.

What we would target

The target we design against: leak repair alone recovering 10–20% of compressor energy, off a leak load that is typically 20–30% of generated flow; staged start-up and interlocked idle shutdown cutting the capacity and network-charge component 12–20%, against your КЕВР tariff structure — we price it off your actual bill, not a generic demand charge — with no production impact. That last one holds only if the restart path is designed against ISO 14118 prevention of unexpected start-up and signed off by your EHS function before anything is commissioned; where an asset cannot be restarted automatically without a safety case, it stays on the manual list. The EnPI and baseline structure is built to feed the ISO 50001 and EED Article 11 evidence pack rather than to duplicate it.

Systems we work with

We integrate with what you already run. If a platform below is missing, tell us — the pattern usually transfers.

Controls, SCADA and industrial connectivity

  • Siemens WinCC and WinCC Unified
  • Rockwell FactoryTalk View SE
  • Ignition (Perspective and Vision)
  • AVEVA System Platform
  • COPA-DATA zenon
  • PTC Kepware KEPServerEX
  • Softing edgeConnector
  • HighByte Intelligence Hub
  • HiveMQ
  • EMQX
  • Node-RED
  • United Manufacturing Hub
  • OPC UA (IEC 62541)
  • MQTT Sparkplug B
  • PackML
  • MTConnect
  • umati
  • EUROMAP 77/84
  • PROFINET
  • EtherNet/IP with CIP
  • Modbus TCP/RTU
  • S7comm
  • EtherCAT
  • IO-Link

MES, ERP and planning

  • Siemens Opcenter Execution and Opcenter APS (Preactor)
  • SAP Digital Manufacturing Cloud
  • SAP PP/DS
  • S/4HANA
  • Dassault DELMIA Apriso and Quintiq
  • Rockwell Plex
  • Critical Manufacturing
  • Tulip
  • Microsoft Dynamics 365 Supply Chain Management, Infor CloudSuite Industrial, Epicor Kinetic, Odoo
  • PlanetTogether
  • Asprova
  • Optessa
  • Google OR-Tools CP-SAT
  • Gurobi

Maintenance, historians and metering

  • IBM Maximo Application Suite
  • SAP PM
  • Infor EAM
  • Limble
  • MaintainX
  • AVEVA PI System
  • Canary Labs
  • InfluxDB
  • TimescaleDB
  • ClickHouse
  • Siemens Senseye
  • Augury
  • SKF Enlight
  • Schaeffler OPTIME
  • ifm moneo
  • Samotics electrical signature analysis, Bently Nevada System 1, Emerson AMS
  • Janitza UMG, Schneider EcoStruxure Power Monitoring Expert, Siemens SENTRON PAC

Vision, metrology and SPC

  • Cognex In-Sight
  • VisionPro and ViDi
  • Keyence CV-X and IV4
  • MVTec HALCON and MERLIC
  • Basler pylon
  • Euresys Open eVision
  • Anomalib with PatchCore and PaDiM
  • Ultralytics YOLO
  • Intel OpenVINO
  • NVIDIA Jetson Orin and DeepStream
  • Minitab, InfinityQS ProFicient, Q-DAS qs-STAT (mandated by several German OEMs)
  • Hexagon PC-DMIS
  • Zeiss Calypso
  • Siemens Opcenter Quality

What we design against

30–50%
Fewer unplanned downtime hours on the monitored assetsMcKinsey’s published range (Manufacturing: Analytics Unleashes Productivity and Profitability, 2017), not a Palamed result. We scope it as a target measured against your own twelve months of CMMS history on the same functional locations, and only where the P-F interval beats the spare-part lead time.
20–40%
Changeover minutes recovered by sequencing aloneBaseline is your current changeover time as a share of available time — often 10–20% on a high-mix line. Measured from MES event data before and after on the same product mix, using a sequence-dependent setup matrix built from history rather than a flat standard setup. Sequencing avoids the expensive transitions; it does not shorten any individual changeover — that is SMED, which is your industrial engineers’ work rather than ours. The two are additive and we scope only the first.
<50 PPM
Escape budget, at a false-reject rate held under 1.5%Baseline is your current escape rate to the customer and the over-reject rate operators already bypass the AOI to avoid. Two numbers that move in opposite directions — anyone quoting a single accuracy figure for an inspection system has not run one on a live line.

Regulation and standards in scope

  • Machinery Regulation (EU) 2023/1230 — applies from 20 January 2027 and repeals Directive 2006/42/EC with no grace overlap. Annex I Part A lists safety components with fully or partially self-evolving machine-learning behaviour as high-risk machinery requiring notified-body conformity assessment; self-declaration is no longer available for them. Safety functions stay in the safety PLC under ISO 13849-1 and IEC 62061, and we document that boundary on day one. The clause that actually touches a retrofit is substantial modification: adding a rejecting vision cell, an interlocked shutdown or a sensor inside a guarded zone can make you the manufacturer of a modified machine, owing a new risk assessment and CE mark. We assess that against the Commission guidance before the scope is fixed, and where a passive, non-actuating read keeps the machine outside it, we say so in writing.
  • AI Act (Regulation (EU) 2024/1689) — core high-risk obligations from 2 August 2026, with AI embedded as a safety component in machinery following the longer 2027–2028 product-legislation timeline. The duties that actually bite in a plant are the deployer ones under Article 26: human oversight, input data relevance, and automatic log retention. An advisory system that leaves the operator in control is a different posture from closed-loop control, and the distinction belongs in the design file, not in an audit meeting.
  • NIS2 (Directive (EU) 2022/2555) with IEC 62443 as the technical route — manufacturers of electronics, electrical equipment, machinery, motor vehicles and medical devices are important entities where they are medium-sized or larger — roughly 50 staff or over EUR 10m turnover. Below that threshold the obligation usually still reaches you through your OEM customer’s supply-chain security requirements, which is the more common route into scope. Bulgaria transposes through amendments to the Закон за киберсигурност, against the Directive’s own transposition deadline of 17 October 2024; we check the text in force on the day the scope is written rather than quoting a status that keeps moving. Twenty-four-hour early warning, seventy-two-hour incident notification, supply-chain security, and personal accountability for the management body — so a supplier who shrugs at OT network design is now a liability rather than an inconvenience.
  • IATF 16949:2016 with the AIAG core tools — APQP, PPAP, control plan, AIAG-VDA FMEA, MSA and SPC. §7.1.5.1 requires calibration and MSA on every measurement system named in the control plan, which is exactly why a vision cell is scoped with an attribute agreement study — or a gauge R&R where it measures a dimension — rather than a “99.x% accuracy” slide. German OEMs additionally impose VDA 6.3 process audits and Q-DAS qs-STAT data formats.
  • Energy and product-data obligations — EED (EU) 2023/1791 Article 11 requires a certified ISO 50001 system by 11 October 2027 above 85 TJ per year and an EN 16247 energy audit between 10 and 85 TJ, transposed in Bulgaria through the Закон за енергийната ефективност and the АУЕР audit register. Alongside it, Ecodesign Regulation (EU) 2024/1781 Digital Product Passports, the Batteries Regulation (EU) 2023/1542 battery passport from 18 February 2027, and CBAM all pull the same per-lot energy and genealogy data out of the MES.

Measurement before modelling: what counts as a stop, and which asset is really the constraint.

What you are probably thinking

We ran a pilot two years ago and it never scaled.

That is the majority experience — the WEF and McKinsey “pilot purgatory” figure behind the Global Lighthouse Network puts it at over 70% of manufacturers investing in advanced analytics never leaving the pilot stage. Pilots die because they were wired point-to-point to one machine with no data model behind them, so use case two costs as much as use case one. We build the connectivity and semantic layer first, prove it on the constraint, and price the second and third use case in the same proposal so you can see the marginal cost falling before you commit.

Our machines are twenty-five years old and have no data.

Most brownfield lines give up usable data without anyone touching the PLC program. Stack-light phase taps, CT clamps on the main feed, proximity and photo-eye counters, IO-Link masters and existing serial ports get availability, cycle count and reason capture running. Where the controller is accessible, an external OPC UA or S7 read is added only after we confirm scan-time headroom and find out who owns the change process. And where the honest answer for a specific machine is that the data is not worth the retrofit, that goes in the assessment in writing.

OEE went up eleven points and I did not ship one more part.

Then the gain was on a non-constraint, which is what happens when OEE is rolled out plant-wide as a scoreboard. Any credible engagement identifies the bottleneck first and measures it hardest, because OEE gains anywhere else are by definition worthless. It is also why OEE must never be tied to shift performance reviews — that reliably produces recoded changeovers and unlogged micro-stops instead of throughput.

Operators will game it, and the works council will object.

Both are real and both are structural problems, not attitude problems. Measure the machine, not the individual. Keep OEE out of individual appraisal. Publish the reason-code definitions so the floor can see the system is not a stopwatch. Treat shop-floor video and operator-linked performance data as GDPR-relevant personal data with a documented lawful basis, a DPIA where monitoring is systematic, and consultation before the pilot rather than after. Systems that skip this get quietly sabotaged, and it shows up as a rising “Other” bucket.

Show me payback inside this fiscal year.

Reasonable for some of this and not for the rest. Automated downtime capture, compressed-air leak repair and demand-charge sequencing typically pay back inside a year and sometimes inside a quarter. Predictive maintenance on critical rotating equipment is more honestly a twelve-to-twenty-four-month case. Vision depends entirely on your current inspector cost and escape cost, so we price it against your numbers. Anyone promising a same-year return on an MES replacement or an APS rollout is not being straight with you.

When we are the wrong choice

  • Safety functions. We do not build, tune or advise anything inside a safety-rated function — that stays in the safety PLC under ISO 13849-1 and IEC 62061, and from January 2027 self-evolving ML in a safety component needs notified-body assessment. Our systems advise; your interlocks decide.
  • MES or ERP replacement programmes. Those are multi-year projects whose critical path is master data, interfaces and change management, not analytics. We work alongside that prime contractor and integrate with what they deliver, never instead of one.
  • Plants where nobody on site will own the data. Reason codes need an owner, AI rejects need an adjudicator, and retraining needs a named engineer. Without those three roles filled from your side, the system decays within a quarter and you will have bought a dashboard nobody trusts.

Questions we get asked

We already have an MES with an OEE dashboard. What would you add?

Often the dashboard is fine and the definitions underneath it are not. We start by asking which asset is the constraint, what percentage of logged minutes sits in “Other”, where the ideal cycle time came from and when it was last re-baselined. If those three answers are healthy, the dashboard is telling the truth and we will say so — that is a cheaper answer for you than a project.

We have no labelled defect images. Does that rule out a vision project?

No — it rules out the supervised proposal you were probably quoted. PatchCore and PaDiM train on defect-free parts only and reach production-relevant accuracy; PatchCore reports 99.6% image-level AUROC on the MVTec AD benchmark. Labels then accumulate from the operator verification station that adjudicates every AI reject, and supervised segmentation follows once there are enough of them to be worth the switch.

IT will not let anything touch the OT network. How does this work at all?

That is the correct position and it is the reference architecture anyway. Data leaves the control zone outbound-only through a broker or gateway in an OT DMZ, with no inbound sessions to Level 2 and no writeback to control unless it has been risk-assessed and explicitly authorised. Under NIS2 your management body is now personally accountable for that boundary, so it is not a detail we would argue with you about.

Our OEM customer is sending Digital Product Passport and CBAM questionnaires. Is that your scope?

Yes, and it is usually the same data problem as traceability. Serialisation and scanning at defined genealogy points let the MES bind serial or lot to components, process parameters, tooling, operator and inspection results, with identifiers to ISO/IEC 15459. One export layer then produces a per-part DPP record or a per-tonne embedded-emissions figure from the same underlying data, instead of a parallel spreadsheet exercise every quarter.

Our master data is a mess. Should we fix the ERP first?

For scheduling, yes — and that may be the whole first deliverable. An APS needs 95%-plus accuracy on BOMs, routings, setup times, capacities and inventory before the optimiser earns anything, so we run a short readiness assessment and, if it says remediate, we quote remediation rather than the optimiser. For downtime capture and vision, poor ERP data is largely irrelevant, so those can run in parallel.

What happens when your people leave? Can our own engineers maintain this?

The handover artefacts are named in the contract: tag dictionary and data model, the OEE charter with frozen definitions, the reason-code taxonomy with worked examples, retraining and re-validation runbooks, and named internal owners trained during the build rather than after it. On standard platforms — Ignition, Node-RED, Maximo, Opcenter — your own automation engineers maintain it. Where we wrote custom code, the repository and the CI pipeline are yours.

Keep reading

Business efficiency automationAI advisory and readinessHow we work: stages, durations, kill criteria
Warm light ribbons on a dark field

Book a 30-minute OEE and downtime review

Bring one month of downtime logs and your current reason-code list. On the call we will tell you what share of your minutes is sitting in “Other”, whether your ideal cycle time is a nameplate number, and whether the losses you are measuring are even on the constraint.

Book a callSee our projects →

You talk to the engineer who would do the work. If the honest answer for a given machine is that the data is not worth the retrofit, you will hear that on the call rather than in month four. Support is written into the contract, not improvised: a failed collector acknowledged within four working hours and restored the same working day, remote access only over the route your IT approves in advance — a jump host or a time-boxed account, never a standing tunnel — and one named engineer reachable outside shift hours.