Industrial computer vision
Cameras, models, and integration delivered by one team for most complex industrial scenarios
Where computer vision helps
| Quality control | Workplace safety |
|---|---|
Defects and deviationsMicro-scratches, streaks, short shots, deformation, a missing element—also on transparent and reflective parts. The station stops, diverts, or logs the piece. |
PPE verificationHelmets and vests checked in the CCTV image, with no biometrics involved. Audible and light alerts follow, up to blocking the entry gate until the worker is equipped. |
Classification and countingProduct type and piece count read at the station in real time. Release is blocked on a short or over pick; the WMS is confirmed only at the exact count. |
Restricted zones and man downEntry under a suspended load or into a forklift path, and a worker's fall. The operator is alerted within seconds, on three levels, and the event is stored for review. |
Measurement and OCRDimensions in hundredths of a millimeter, plus numbers and markings read off the part. The tolerance decision follows from the value, and the reading lands in the record. |
Event logNear-miss events captured with an image and metadata, building a data set for EHS audits, insurers, and safety training. |
How computer vision works in a plant
Step 1
Industrial cameras
GigE / RTSP · fixed stations and line-mounted units · lighting matched to the scene
Step 2
Edge AI (GPU)
local inference on an edge server · video does not leave the plant
Step 3
Vision models
detection and classification · segmentation · industrial OCR · trained on your data
Step 4
Events and integrations
real-time alerts · evidence recording · integration with MES, CMMS, and Nexen
What our team delivers
- Camera, optics, and lighting selected for the scene, plus the design of the station.
- Edge server supplied, installed, and configured on your network.
- Integration over API, alerting at the station, and the model kept under SLA.
- Recording, labeling, and training on your products until the agreed accuracy holds.
What you provide
What you provide: access to the line, a sample of footage from the scene, and the systems we integrate with. One contract, one point of accountability for hardware, models, and software.
Technology partner
Sensors and machine vision components—cameras, optics, and lighting—selected for the scene rather than ordered from a catalog.
Privacy and bandwidth stay under control. The platform receives events and metadata, not raw video. Deployment is on-premise where the data requires it, and images are anonymized where GDPR applies.
Examples from our project portfoliot
Quality inspection of car headlamps
Challenge: Visual assessment of transparent lenses is subjective and tiring, concentration drops within a shift, and light reflections on clear plastic hide defects that cost multiples to fix later in vehicle assembly.
Solution: A deep-learning vision system that identifies micro-scratches, milky streaks, and short shots, and works through the reflections on transparent lenses where human judgment is least reliable.
Value: Defects stopped at the source, with identical accuracy across the whole shift and no defects passed to later assembly stages.
Vial seal verification measured in hundredths of a millimeter
Challenge: A pass or fail verdict on cap seating was not enough; the process needed the actual gap under the cap, measured repeatably on the line.
Solution: RT-DETR detection combined with a ResNet18 regression head, on a pilot station with two cameras per vial path, incident ring lighting, the separator used as a backlight, and a 150 mm working distance.
Value: 0.04 mm mean absolute error on the gap, a 66% precision gain over the baseline version, and full cap detection on the stream.
Incoming delivery control at unloading stations
Challenge: Raw material cost depends on honest deliveries. Attempts to inflate weight by adding water were hard to catch in the moment, manual checks could not keep up with the flow of trucks, and there was no hard evidence for renegotiation.
Solution: An edge AI system analyzing camera streams locally on a GPU server in real time, flagging water added during unloading and storing a buffered clip as evidence. Later extended with full unloading logs, plate recognition, and integration with the yard system and NVR.
Value: Evidence for payment renegotiation with suppliers and a statistical record per supplier, with no extra load on the network.
Rail wagon identification and shipping documents
Challenge: Identifying wagons, weighing them, and filling in shipping documents by hand was slow across hundreds of wagons, and left no auditable link between wagon, mass, load, and document.
Solution: Multi-camera acquisition detects a passing train and reads wagon numbers in real time with a confidence score and a correction path. Rail scale integration assigns mass to the right wagon, empty or loaded, and documents are generated from templates and matched to the train list.
Value: One automated chain from wagon to number to mass to document, buffered through connectivity gaps, deployed on-premise with images anonymized.
Vision-based component counting at picking
Challenge: Breaking down bulk packaging is a bottleneck in high-volume assembly, manual counting of connectors, tapes, and cables slips past WMS terminals, and one missing part worth a fraction of a cent stops the flow line.
Solution: A vision system on the Nexen platform turns the picking station into a verification point: counting runs in real time, the screen turns green at the exact quantity, and Nexen confirms completeness to the WMS over API. An extra item or a short close raises an alarm.
Value: A zero-defect standard on warehouse release, no production stoppages from material shortages, and a two-stage rollout that limits investment risk.
Safety around vertical transport on site
Challenge: Zones under a suspended load carry the highest risk of serious accidents, organizational measures depend on people staying alert in noise and poor visibility, and near-miss events go unrecorded despite growing regulatory and insurance requirements.
Solution: A two-module edge system pairing an external LiDAR and camera sensor with an in-cab module carrying an AI accelerator. It detects a worker in the danger zone and warns the operator on three levels with an LED and buzzer, logging each event with an image and metadata. Offline-first, with no LTE needed and no biometrics.
Value: Operator decision support across five machine types on one hardware variant, mounted in under fifteen minutes, with near-miss data for safety audits.
How to start
Start with one station and one defect. The audit tells you whether the image carries the information at all—before anything is bought.
1. Workshop and audit
Scope. The stations you have in mind: what has to be detected, how it is checked today, scene conditions, and the systems the result has to reach.
Result. A feasibility verdict per case, a target architecture, and a pilot scope with accuracy criteria and costing.
2. Pilot station
Scope. One station, your parts, your line—footage recorded and labeled on site, models trained on your products, events wired into one system.
Result. Measured accuracy against the agreed criteria, and a business case built on your numbers.
3. Decision based on data
Result. A go / no-go decision with scope, budget, and timeline.
4. Deployment and scaling
Scope. Rollout in stages, without stopping production. Further stations run on the edge infrastructure and model pipeline already in place.
Result. A production system with working integrations, trained operators, and models kept under SLA.
Start with an audit or a pilot
Workshop and audit
Find out what your data is worth before you spend anything on collecting it.
Pilot or PoC
One line, your machines, your numbers. If they don’t work, you stop there.
Dive deep into the world of AI-powered software development
AI can rewrite your legacy system. The hard part is proving it works.
AI can make legacy rewrites faster, but safe modernization depends on rigorous validation, differential testing, reconciliation, and controlled cutover.
AI agent development is changing the economics of custom software
AI agent development lowers software delivery costs by automating bounded tasks, accelerating modernization and making more custom projects viable at scale
Vibe coding: a fast track from idea to application, or a risky shortcut?
Vibe coding: find out what it is, where it works best, and what risks it brings. Learn when AI accelerates software development.
