March Code
Manufacturing

Computer vision system for managing paving tile production

A manufacturing execution system (MES) for production control and output tracking at a paving tile plant. Our own scanners on the lines photograph every pallet, a neural network counts tiles and defects, and planning and analytics update in real time.

Client: ProControlAIAutomationIoT
Computer vision system for managing paving tile production

200K+

pallet scans processed by AI

100%

output tracking with no manual input

14 ms

AI tile detection per scan

4 lines

running our own scanners

01

Challenge

A typical day at a plant without ProControl

End of shift. The foreman pulls out a notebook and writes: “Line 1: 847 pallets, 12 defective.” On the next shift, another foreman writes down his own numbers. They don't match. Nobody knows who's right, because both counted by eye.

This isn't the exception, it's the norm. At the vast majority of paving tile plants, records are kept by hand: in notebooks, on scraps of paper, in Excel at best. Data shows up hours or days late. The owner sees the state of production once a week at best, and even then as a summary someone put together by hand.

What it costs the business

~15%
lost to unrecorded defects
3 days
before defects are discovered
0%
real-time data

Management sees reports once a week. By then, the defective tiles are already in the warehouse, maybe even with the customer. Returns, complaints, a damaged reputation, all because the problem was found too late.

Planning is done by eye: some items are overproduced, others run short. The warehouse is packed with dead stock while popular items are out. Shifts argue constantly: “That's not our defect, the previous shift left it.” Without objective data, there's no way to settle it.

The operator fills in forms by hand and spends 15–20 minutes on it at the end of every shift. The data is subjective: some round off, some forget, some pad the numbers. A quality control system? There simply isn't one, just a visual check and the hope that the foreman spots the defect.

The core problem: 99% of paving tile plants are working blind. Data arrives hours or days late, and decisions are made on gut feeling. This isn't an engineering problem. It's a management dead end you can't get out of without going digital.

02

Solution

Our own hardware: scanners for the line

We designed and built our own scanners: an edge device based on a Raspberry Pi with an Orbbec 3D camera (RGB plus a depth map) in a rugged industrial enclosure. It handles the dust, vibration and temperature swings of the shop floor. Installing it on a line takes a few hours and doesn't require stopping production. A turnkey hardware-plus-software solution.

The journey of one pallet

Scan

The scanner photographs the pallet, and the conveyor doesn't wait (fire-and-forget)

AI: mold

The neural network identifies the tile type and matches the product from the catalog

AI: count

The detector counts every tile: 14 ms per scan

Order

The scan is added to the line's order; a product that doesn't belong triggers an auto-pause and an alert to the operator

Reports

Output, defects and line utilization update instantly

Right on target, not “a bit extra just in case”

The system knows the target volume adjusted for defects and counts only good pallets. On the tablet, the operator sees “1,050 needed, 870+ produced” and stops exactly on target. No overproduction, no overstocked warehouse.

Smart planning
  • A Kanban board of orders by line, with drag-and-drop
  • Live progress straight from the scans
  • Machine station view: a tablet at the line with the target and the queue
Integrations
  • Orders from Bitrix24 (a CRM popular in Eastern Europe) via smart processes and webhooks
  • Catalog from the MDM system, pallet photos in S3
  • Statuses and actual output go back to the CRM
AI that keeps learning
  • A classifier plus a detector for each mold
  • Operator corrections build up into training datasets
  • New products without rewriting code

Bottom line: every pallet counted, every order tracked, every line measured. From a physical tile on the conveyor to a row in the management report, fully automatically.

03

The system from the inside

Real screens of a working system, not mockups. Click to take a closer look

Scan details: the neural network recognized the mold and counted 75 tiles in 14 ms, with types, area and layout
Scan feed: 200,000+ pallets with filters by line, product, AI status and review
Planning: a Kanban board of orders by production line with live progress from the scans
Production summary: output, defects and utilization of four lines in real time
Machine station view: the operator's tablet at the line with the current order, progress and queue
04

Results

200K+

pallet scans processed by AI

Every pallet on four lines is photographed by our own scanner and counted by a neural network

100%

output tracking with no manual input

Data appears the moment a pallet leaves the line: no more notebooks or Excel

14 ms

AI tile detection per scan

The neural network counts every tile on the pallet and identifies types and area faster than the screen refreshes

4 lines

running our own scanners

Edge devices with 3D cameras, built by us, run on the shop floor around the clock: defects are visible the moment a pallet leaves the line, not 3 days later in the warehouse

05

In-depth breakdown

Value at every level

For the owner

Transparency, predictability, control. Decisions based on data, not intuition. You see the real picture of the plant right now, from your phone, not a week later in Excel. How much was produced, how much was written off, how loaded the lines are: it's all in one place.

The owner gets a tool for asking the right questions. Why is the defect rate on line 3 higher than on line 1? Why does the second shift consistently produce 12% less? Before, nobody could ask these questions, because the data simply didn't exist.

For the production manager

No more arguments between shifts: all the numbers are in the system. The real load and performance of every line are visible. Defects are recorded immediately, not at the end of the month. Planning stops being guesswork, because the system shows real capacity and bottlenecks.

The production manager can compare line efficiency, track trends and spot anomalies in real time. If defects suddenly spike on a line, it shows up right away, not three days later when the defective batch has already shipped to the customer.

For the operator

No spreadsheets, notepads or manual counting. The system counts the output itself. A clear shift report with no padded numbers, and no accusations from the next shift. The operator does the actual job, not paperwork.

Why it works

ProControl solves the fundamental problem of production accounting: the gap between what really happens on the line and the data in the system. When information is collected automatically, in real time, with no human involved, it's objective. You can base decisions on it.

The system scales from one line to dozens. Each new camera is one more data source that joins the shared system in hours, not weeks. The architecture is designed to grow with the plant.

ProControl isn't just a tile counter. It's a digital layer over the production line that grows with your plant. Today, it's tracking and control. Tomorrow, predictive analytics and AI quality control.

Project technologies

Next.js 15React 19TypeScripttRPCPrismaPostgreSQLPythonFastAPIYOLO / Computer VisionRaspberry PiOrbbec 3D camerasBitrix24 APIDockerGitLab CI/CDS3

Want a similar result?

Pilot from $3,900, prototype free

Prices are indicative and not a binding offer.

We build an agent prototype for one scenario using your examples, so you can judge answer quality before signing.

Step 1 of 2 · Task