Manufacturing

AI on the plant floor: vision inspection, predictive maintenance and IoT that talk to your MES and ERP

Erbis builds computer vision, telemetry and planning systems that run on your lines and inside the systems you already operate. We have worked at the CAN bus level for a named vehicle manufacturer, modernized an ERP across 25 production sites, and modeled 304,441 shipments for a global equipment maker.

Erbis builds custom AI for manufacturers: computer vision quality control, predictive maintenance, industrial IoT and telemetry, production planning and scheduling, digital twins and simulation, and supplier and inbound logistics, integrated with your ERP, MES, SCADA and PLC data. Engineering since 2012, production AI since 2017, with named work for WH Bence and a global agricultural-equipment manufacturer.

Book a discovery call Take the AI readiness assessment
shipments modeled for a global agricultural-equipment manufacturer (ETA error 21.4 h)
304K
production sites running on one ERP Erbis modernized
25
system integrations completed
4,000+
AI accuracy in production, all industries
~95%
The problem

Why four in five manufacturers are not ready to scale AI

Manufacturers have moved first on quality. 77% use AI in some form, 47% use it in quality processes (up from 33% in 2025), defect detection is 44% of quality use cases, and more than 70% plan AI visual inspection within 18 months (Pulse of Quality in Manufacturing 2026). AI quality inspection is an $18.6B market in 2026. On the maintenance side, predictive maintenance programs report 30–50% less unplanned downtime and 25–40% lower maintenance cost (Deloitte).

Scaling is the gap. Deloitte's 2026 Manufacturing Outlook finds only about one in five manufacturers feel ready to scale AI, even as 80% plan to spend a fifth or more of improvement budgets on smart manufacturing and agentic adoption is expected to quadruple from 6% to 24%. The reasons are familiar from the line: too few labeled defect images, PLC and SCADA data that never reaches a data platform, models that cannot explain a reject to a quality engineer, and operators who have watched pilots come and go.

Erbis treats the plant as an integration problem first. We have worked at the CAN bus level, re-architecting the embeddable vehicle control console for WH Bence; modernized a decade-old POS into an ERP that runs production, stock and sales across 25 sites; and built the data pipelines and models behind a predictive ETA service for a global agricultural-equipment manufacturer. AI goes on top of that plumbing, where it can be measured in scrap rate, OEE and downtime hours.

What we build

Solutions we deliver

Catch defects with computer vision quality control

Vision inspection models trained on your parts, cameras and lighting, deployed at the station or on the edge, with reject reasons a quality engineer can read and a labeling loop that grows the dataset from production.

  • Defect detection and classification on existing or new camera installations
  • Edge deployment for line-speed inference, with cloud training and model versioning
  • Active-learning labeling loop that starts from a few hundred images, not a few hundred thousand
  • Explainable rejects: heatmaps and reason codes linked to the part record
  • Results written to MES and quality systems for traceability and SPC

Predict failures before they stop the line

Predictive maintenance from vibration, temperature, current and cycle data, tied to your CMMS so a prediction becomes a work order, not a chart.

  • Sensor ingestion from PLCs, SCADA historians and retrofit IoT devices
  • Anomaly detection and remaining-useful-life models per asset class
  • Work orders and parts reservations raised in the CMMS or ERP automatically
  • Thresholds and playbooks that maintenance leads control
  • Downtime and maintenance-cost tracking against the pre-AI baseline

Instrument machines and products with industrial IoT

Device integration, telemetry pipelines and control interfaces, from the CAN bus to the cloud. For WH Bence, Erbis re-architected an embeddable console that integrates third-party sensors over CAN and controls hydraulics, lighting, generators and communications from one dashboard.

  • CAN bus, Modbus, OPC UA and MQTT integration for machines and vehicles
  • Telemetry pipelines with edge buffering for unreliable connectivity
  • Control dashboards and operator interfaces built for field conditions
  • Installed-base monitoring for equipment makers, with service alerts
  • Security by design: device identity, signed updates and segmented networks

Plan and schedule production against real constraints

Planning and scheduling models that use actual cycle times, changeovers, labor and material availability, with AI agents that re-plan when a machine goes down or a supplier slips and propose the change to the planner.

  • Finite-capacity scheduling with changeover, tooling and labor constraints
  • Demand and material forecasting linked to MRP and purchasing
  • Re-planning agents that respond to downtime, quality holds and late inbound
  • Scenario comparison for the planner before a schedule is released
  • Schedule adherence, OEE and WIP tracked per line and shift

Integrate ERP, MES and the shop floor

Integration engineering between ERP, MES, quality, maintenance and warehouse systems, plus modernization of the legacy applications that run production today. Erbis rebuilt a decade-old POS into an ERP running production control, stock and sales across 25 sites in Germany.

  • ERP and MES integration (SAP, Oracle, Microsoft Dynamics 365 and custom systems)
  • Shop-floor data collection from PLCs and SCADA into a governed data platform
  • Legacy application modernization without stopping production
  • Master data cleanup for parts, BOMs and routings before AI touches them
  • Custom ERP modules where off-the-shelf systems lack the niche workflow

Simulate before you commit: digital twins and inbound logistics

Digital twins and simulation for lines, cells and supply networks, and inbound logistics that link your suppliers and carriers to production. Our supply chain practice handles the freight side: predictive ETA, carrier integration and inbound visibility.

  • Discrete-event simulation of lines and cells to test layout and staffing changes
  • Digital twins fed by live sensor data for what-if analysis
  • Inbound visibility with predictive ETA for supplier shipments
  • Supplier performance analytics on OTIF, quality and lead time
  • Carrier and 3PL integration through the Erbis supply chain practice
In production

How AI fits the systems you already run

A plant is a stack: PLCs and sensors at the bottom, SCADA and historians above them, MES and quality systems, ERP and planning on top. Erbis connects AI into that stack through the interfaces each layer already exposes, and keeps the operator in control of anything that changes a machine or a schedule.

01

Integration map

Field: PLCs, CAN bus, Modbus, OPC UA, MQTT and retrofit IoT devices. Plant: SCADA, historians, MES, quality (SPC, LIMS) and maintenance (CMMS). Enterprise: ERP (SAP, Oracle, Microsoft Dynamics 365 and custom), PLM, WMS and TMS. AI services read from historians and MES, and write back as work orders, quality results and schedule proposals through those systems' APIs.

02

Data governance and safety

Production data stays in your environment: AWS, Azure, Databricks or on-premises, with OT and IT networks kept separate. Erbis is ISO 27001 certified and applies device identity, signed firmware updates and segmented networks for connected equipment. Where safety-related control is involved, AI recommends and a person or a certified control system acts.

03

Evaluation and human-in-the-loop

Vision models are validated on held-out parts against your current inspection outcome; maintenance models against logged failures; schedules against what planners actually released. Quality engineers and maintenance leads set thresholds, review low-confidence outputs and feed corrections back into training. Nothing is scored on a demo dataset.

04

Run and support

Monitoring for model drift as products, lighting and tooling change, for pipeline health from edge to cloud, and for the operating metrics the project was funded on: scrap rate, unplanned downtime, OEE, schedule adherence. Erbis retrains, extends to new lines and sites, and hands over code, models and runbooks at every phase.

Proof

What we have built

All case studies

How it works

How we work with you

  1. 01

    Discovery and data audit

    Two to three weeks on site and in your systems. We map lines, sensors, historians and ERP/MES data, choose the metric (scrap rate, downtime hours, OEE, schedule adherence) and pick a first line or asset class with a baseline.

  2. 02

    Pilot on one line or asset class

    Cameras, sensors or data connections installed where needed, a model trained on your data, and working software in front of operators within 2 weeks. A production-grade pilot with a go/no-go result is typically ready in 4.

  3. 03

    Integration and validation

    Connect results to MES, CMMS, quality and ERP systems, run in shadow mode next to the current process, and validate against held-out parts, logged failures or released schedules until the model beats the baseline.

  4. 04

    Rollout and support

    Extend line by line and site by site with a template for new cameras, assets and SKUs. Erbis monitors, retrains and supports the system, and you own the code and models from day one.

Stack

Systems and technologies we work in

Industrial systems and protocols

  • OPC UA
  • MQTT
  • Modbus
  • CAN bus
  • PLC and SCADA integration
  • MES
  • CMMS
  • SPC and quality systems
  • SAP
  • Oracle
  • Microsoft Dynamics 365

Computer vision and ML

  • PyTorch
  • OpenCV
  • YOLO
  • Amazon SageMaker
  • XGBoost
  • Databricks
  • NVIDIA Jetson

IoT and edge

  • AWS IoT Core
  • AWS IoT Greengrass
  • Azure IoT Hub
  • Amazon Timestream
  • Grafana
  • Python
  • Java
  • React

AI and agents

  • Anthropic Claude
  • OpenAI
  • Google Gemini
  • Amazon Bedrock
  • LangChain
  • LangGraph
FAQ

Manufacturing: questions we get asked

Do we need a large labeled defect dataset before starting computer vision QC?

No. Most lines have far more good parts than defects, so we start from a few hundred images per defect class, use augmentation and active learning, and grow the dataset from production rejects that engineers confirm. The labeled-data bottleneck is the most common reason vision projects stall, so the labeling loop is built in the first sprint, not after.

Can you work with our PLCs, SCADA and MES, or do we need new hardware?

Usually no new hardware beyond cameras or retrofit sensors where none exist. Erbis integrates through OPC UA, Modbus, MQTT and CAN bus, reads historians and MES through their interfaces, and keeps OT and IT networks separate. For WH Bence we integrated third-party sensors over CAN into one control console; on other projects the data already sits in a historian and only needs a pipeline.

How do you measure whether predictive maintenance is working?

Against your own logged failures and maintenance history. We set the baseline for unplanned downtime hours and maintenance cost per asset class before the model runs, then track predicted-versus-actual failures, false alarms and the cost of the interventions the model triggered. Industry programs report 30–50% less unplanned downtime; we report your number, not the industry's.

Which manufacturers do you work with?

Discrete and process manufacturers, equipment and vehicle makers, and multi-site producers with an ERP or MES they need to extend. Named work includes WH Bence (vehicle control console over CAN bus), a German print manufacturer running the ERP we built across 25 sites, and a predictive ETA service whose end client is a global agricultural-equipment manufacturer.

Can you integrate with SAP or another ERP without disrupting production?

Yes. Integration runs through the ERP's published interfaces, is tested in a sandbox, and goes live in shadow mode next to the current process. Erbis has completed 4,000+ system integrations and modernized production-critical applications while they stayed in use, including a decade-old POS rebuilt into a 25-site ERP over a phased program.

How do you connect the plant to our supply chain?

Through the Erbis supply chain practice, our deepest industry. Inbound visibility with predictive ETA, carrier and 3PL integration (60+ carrier APIs for one TMS), supplier performance analytics and replenishment agents connect the plant schedule to what is actually arriving at the dock. See the supply chain industry page for the full scope.

Related
Supply Chain & LogisticsAgentic AI EngineeringData & AI PlatformsLegacy Modernization

Pick one line, one asset class, one number

A 45-minute discovery call to map your lines, systems and data and choose a first pilot with a baseline. Working software in front of operators within 2 weeks.

Book a discovery call Take the AI readiness assessment

AWS partner · Databricks partner · ISO 27001 · You own the code from day one