Running in production
See every defect.Count every unit.DetectionAutomationIntegration
Film 10 minutes of your line — we deploy a trained detector in 2–5 days.
What we build
Six problems one camera solves.
Each system is trained on your footage, validated on your line, and deployed where the work happens. The question is which one fits your line first.
01
Counting on the line
Counts every unit passing a checkpoint by type or colour class, 24/7 — replaces the manual shift spreadsheet.
02
Colour change detection
Catches colour deviations in real time before the problem moves downstream or becomes a customer complaint.
03
Presence / absence of an object
Confirms every component, label, or fastener is present on every unit — 100% frame coverage, not spot sampling.
04
Fastener & assembly verification
Verifies bolts, clips, and connectors are correctly seated on every unit, not every 50th — catches assembly errors at the line.
05
Defect & surface anomaly detection
Flags cracks, scratches, and contamination at line speed with consistent detection threshold across the whole shift — no fatigue, no drift.
06
Process & throughput monitoring
Measures cycle time, dwell, and bottlenecks directly from existing camera feeds — no new sensors on the floor.
What we deliver
From first footage to live on your floor — in under a week.
From footage to production
Pilot turnaround — footage to deployed detector
A typical system integrator quotes 4–6 months for a vision project — we hand you a working detector in the first week, trained on your footage, running on your hardware.
Of footage — the only thing we ask from you to start
Ten minutes of recording from any camera — your phone, an IP camera, or a GoPro — is your entire commitment at this stage.
Every unit checked — no sampling, no end-of-shift gaps
Every unit checked in under 50ms — zero sampling, every unit, every shift, no exceptions.
Data stays on your hardware
The model runs on a local server inside your network — footage, counts, and anomalies stay on-premise, no firewall ports needed.
How it works
Your job: film 10 minutes. Our job: everything else.
01
Film
Film 3–10 minutes of your process with any camera — same angle and lighting as production. No special hardware required.
02
Label
We extract frames and label every object your line needs to track — automatically, in hours. No labeling work on your end.
03
Train
We train a detection model on your objects, your lighting, and your defect patterns — you get accuracy benchmarks before we deploy anything.
04
Deploy
Results push to your ERP, dashboard, or alerting system in real time via API — no cloud, no new servers, no downtime.
Industries
Where we have real track record or a concrete application.
The same detection and tracking pipeline adapted to your domain — trained on your data, tuned to your objects and tolerances.
01
Agriculture & Food
Produce counting and colour-class grading on conveyors, yield delta calculation, contaminant detection. Live reference: tulip counting with ERP integration running in production.
02
Manufacturing
Assembly verification and defect detection at line speed — bolts, clips, connectors, surface anomalies — on every unit, not every 50th.
03
Logistics & Warehousing
Parcel and pallet counting, damage flagging at dock intake, throughput monitoring at dock doors.
04
Construction & Safety
PPE compliance monitoring, restricted-zone entry detection, equipment utilisation tracking.
05
Retail & Inventory
Shelf stock monitoring, queue length estimation, people counting, and product placement verification.
06
Pharmaceuticals
Fill-level and seal inspection, label and lot code verification, particulate detection in packaging lines.
Technology
A practical stack that gets to production quickly.
We use proven open-source tools — no black boxes, no vendor lock-in. The same pipeline that worked on our first pilot scales to your facility.
DETECTIONYOLOv8n · Ultralytics
TRACKINGByteTrack (built-in)
ANNOTATIONRoboflow · SAM3
DEPLOYMENTDocker · on-premise
Custom-trained models
Models trained exclusively on your footage. Your objects, your environment, your tolerances — not a generic pre-trained classifier.
AI-assisted annotation
Auto-labeling cuts annotation time from weeks to hours. You review edge cases if you want to. Most customers don't need to.
Stable object tracking
Each object gets a unique ID across frames — one product, one count, no duplicates, even on fast-moving conveyors.
Connects to your systems
REST and WebSocket outputs. Counts and events flow to your ERP, MES, dashboard, or alerting system in real time. We have integrated with 1C — we can connect to whatever your operation already runs.
Contact
Tell us what's slipping past your line today.
Describe what you want to count, catch, or verify — and what happens today when it's missed. We'll come back within one business day with a pilot scope and an honest read on whether vision is the right tool. If it isn't, we'll say so. No sales deck. No retainer.
tolyakravtsov@kraytex.com+48 500 613 730 · Warsaw