Identify
Understand exactly where counting broke down and why manual checks weren't catching it
The situation
Takenaka runs high-volume assembly and palletisation lines across multiple Japanese production floors, packing products in precise tier-by-tier configurations before shipping. Getting every layer counted correctly is what keeps the right quantities moving to the warehouse and, ultimately, to the customer.
On the floor, the team had already lived with:
- Manual tier-by-tier counting with no independent verification
- No image evidence to resolve shipment disputes
- Counts lost whenever an operator was reassigned mid-pallet
- A multilingual, gloved workforce that traditional keyboard logins didn't suit
- Errors that surfaced only after the pallet had already left the floor
As volume grew and shift patterns diversified, the cost of those errors — wrong shipments, inventory mismatches, after-the-fact quality investigations — kept accumulating. The vision behind fixing it properly, though, was still strong.
The challenge
The challenge wasn't just counting products faster — it was making it structurally impossible for a miscounted pallet to leave the floor, in an environment where operators wear gloves, work across shifts, and speak seven different languages between them.
Our approach
We were brought in to design a system that would work the way operators actually work on a factory floor — not the way a typical enterprise workflow assumes they do.
Our objective was to:
Design
Build a two-tablet camera and AI system that fits naturally into the existing pallet station
Gate
Make it structurally impossible for a mismatched pallet to leave the floor
Scale
Support multilingual, multi-shift operators with session handover and a full audit trail
What we did
1
Built a two-tablet system with a live camera feed
One tablet mounts at the pallet station as a dedicated camera device, streaming a live WebRTC video feed over the factory Wi-Fi. The operator's handheld tablet receives that feed in real time, so nobody has to hover their own device over each tier.
Built to survive real floor conditions:
- Live WebRTC feed streamed over the existing factory Wi-Fi
- Picture-in-picture view persists while operators check packaging guides or scan barcodes
- Count context is never lost mid-task
The result: a stable, well-positioned camera view the AI can count against, instead of a shaky handheld shot.
2
Trained a dedicated AI counting engine per product
Each product runs its own trained YOLOv8 model, built from real production-floor images — variable lighting, metal containers, partial occlusion, and natural placement variation included.
Two ways to count, one audit trail:
- Manual mode — operator photographs a tier, AI returns a count with bounding boxes
- Automated video mode — system samples the live feed and advances tiers on its own
- Both modes write to the same structured audit output
3
Added a mandatory barcode release gate
No pallet can be released until the operator scans the pallet barcode a second time and it matches the barcode scanned at the start of the session. If it doesn't match, release is blocked outright.
What the gate covers:
- Barcode scanned at session start and again at release
- A mismatch blocks release — no manual override without a logged reason
- Addresses the most common cause of wrong-pallet-wrong-location errors
4
Layered in session persistence, guides, and a full audit trail
If an operator is pulled away mid-count, the session saves automatically. Any operator, on any shift, can resume exactly where it was left, with every tier image and count intact.
Supporting details that mattered on the floor:
- In-app packaging guides — photos, PDFs, step sequences — right on the counting screen
- Seven-language support (English, Japanese, Hindi, Nepali, Portuguese, Tagalog, Vietnamese)
- Every count, scan, login, and pallet status change logged with timestamp, device serial, and operator identity
Quality teams now have a complete, immutable audit trail for every pallet that has moved through the floor.
How it works
From operator login to a released, fully audited pallet — every stage recorded, one hard gate before anything leaves the floor.
1
Operator login & barcode scan
Badge scan logs the operator in and locks the session to a pallet barcode.
iOS + barcode login2
Live camera feed (WebRTC)
A mounted tablet streams a live view of the pallet station to the operator's handheld device.
WebRTC · factory Wi-Fi3
AI tier count (YOLOv8)
A per-product model counts each tier and returns bounding boxes on the captured image.
YOLOv8 per product4
Barcode match & release gate
A second barcode scan must match the session start scan before the pallet can be released.
Hard block on mismatch5
Audit trail & supervisor dashboard
Every count, scan, and status change is logged and visible to supervisors in real time.
Full device-level log“
We used to find out about counting mistakes after the pallet had already left the floor. Now we catch them before they happen, every single time.
Takenaka
“
The app switching to my language automatically made a real difference. I don't have to guess at the instructions anymore — everything is clear before I start.
Takenaka (via translation)
~0
Shipment errors
After release gate adoption
100%
Pallet traceability
Every tier, operator, timestamp recorded
7
Languages supported
Including Japanese, Hindi, Tagalog
0
Sessions lost at handover
Any operator can resume a pallet mid-count
The outcome
What changed after rollout:
- AI counts every tier from a camera feed, with a mandatory gate before release
- Full tier-image archive with timestamps and operator IDs for every pallet
- Sessions persist across shifts — any operator can resume a count
- Packaging guides built into the app — no separate binder or supervisor walk
- Supervisor dashboard with pallet history, scan trails, and line status
- Device-level accountability let IT trace a compliance question to a specific tablet within minutes
Frequently asked questions (FAQs)
1) What is AI-powered pallet verification?
It's a system that uses computer vision to automatically count product tiers on a pallet as it's packed, replacing manual counting with a verified, image-backed count at every layer.
2) Why is manual tier-by-tier counting prone to error on a factory floor?
Products are visually similar, packed tightly across multiple layers, and counted under variable lighting by operators managing other tasks at the same time — fatigue and distraction both introduce errors that stay invisible until something is wrong downstream.
3) What is a mandatory barcode release gate and why does it matter?
It's a checkpoint that requires the pallet barcode to be scanned and matched a second time before release. If the scan doesn't match, the pallet is blocked from leaving — preventing wrong-pallet shipments before they happen rather than catching them afterward.
4) How does session persistence work across shift handovers?
If an operator is reassigned mid-count, the system saves the session automatically. Any operator on any shift can resume it exactly where it was left, with all tier images and counts intact.
5) Can an AI counting system support a multilingual workforce?
Yes. The interface can switch language automatically based on an operator's login profile, supporting multiple languages so instructions are clear without relying on a shared working language.
6) What is WebRTC live camera streaming and why use it for pallet counting?
WebRTC lets a dedicated camera tablet stream a live video feed to an operator's handheld device in real time, giving a stable, well-positioned view to count against instead of a handheld, shaky shot.
7) How does Atomic Loops build AI systems for factory-floor environments?
By designing around how operators actually work — fast logins, gloves-friendly interaction, multilingual support, and session continuity — rather than adapting a generic enterprise workflow to the floor.
8) Can this kind of system integrate with existing production line hardware?
Yes. It's designed to layer onto existing pallet stations and factory Wi-Fi without requiring a rebuild of line hardware or infrastructure.
Industry Services
Education
Focus areas - Education
© 2026 London School of DIGITAL BUSINESS.
All Rights Reserved