Client’s quote
“Thanks to First Bridge’s comprehensive approach, we were able to implement loading management automation that reduced our goods loading time by an impressive 15%.”


Client is a wholesale purchaser of various types of goods with subsequent sales to distributors in multiple regions.


The company has proprietary software to track the stock and supply with connected 3rd party tools to control the supply chain. All goods got delivered to the central warehouse, from which they were distributed among regional warehouses. 

Company wanted to go through the process of digital transformation and by using cutting edge technologies to optimize its core process associated with storage
and transportation of goods.

“During the pandemic, we’ve encountered a huge bottleneck in process of loading the cargo in trucks. The existing process didn’t work well with new regulations. “

Upper level plan

Together with the client, First Bridge R&D team has identified the core processes that will benefit the most from the technological transformation.

Development of specialized loT tracking


First Stage - IoT Element

The first stage was the introduction of an IoT element in the form of a protective QR code, the data from which is recorded in the blockchain database. Throughout the entire supply chain, all stakeholders could use a dedicated mobile application, scan this code and receive all data from the blockchain about a specific product - date, production time, place, etc.

We’ve developed a multi-component system that uses a mathematical algorithm to check whether the medium is original or reprinted. This solution consists of the following elements:

  • Graphic Security Element
  • QR code
  • Administrative Software
  • Quality Control System
  • Control Terminal
  • Register

“We were already using a lot of automation in our process, but we didn’t make a good use of all data that we were accumulating over the years. This was an opportunity for us to make a more robust decisions and we need a tool to work with all this data. “

Second Stage - AI.
The next step was the implementation of a neural network for the following tasks:
  • Recommendations for the supply of products to various regions
  • Procurement planning based on historical data, taking into account seasonality, region, and external factors, and choosing the best delivery method
  • Planning the placement of goods in containers/trucks during transportation between warehouses

Supply Planning

To schedule deliveries, we analyzed historical data from the company's reports over the past three years. We conducted a quarterly analysis of this data and obtained average indicators and statistics for each period. We further divided these periods into smaller ones, such as up to a week, to enable more precise analysis.

Procurement planning and selection of the optimal
delivery method

Using the categorized deliveries and considering the seasonality of shipments, we designed a neural network. This network recommends a procurement plan for a specified period and the optimal delivery method. Its purpose is to ensure adequate stock levels in the warehouses while avoiding excess inventory, resulting in significant savings on warehouse rental costs.

This has reduced warehouse rental costs by ~20%.

The First Bridge team experimented with several types of neural networks to determine the most effective option for data analysis. These included direct propagation networks, recurrent neural networks, and Kohonnen map-based neural networks. After careful consideration, the team selected multilayer neural networks of direct propagation due to their versatile and adaptable structures. The figure below illustrates the general architecture of a feedforward neural network.

Planning the placement of goods in containers and trucks during transportation between warehouses

After choosing the type of neural network architecture, the first step was to determine the key parameters and develop a system for bringing and coding all the data provided by the customer to a single format.
The next step was experimental selection of a specific architecture of neural networks for a system of building a procurement plan for a certain period which delivery method should be used to ensure a sufficient volume of goods in warehouses. The dataset was divided into two components:

Based on the test results, an architecture was chosen that most accurately displays the data provided by the customer. As a result, we were able to build a relationship between the increase in demand for certain groups of goods with seasonal fluctuations and world events affecting the respective regions.

Planning the placement of goods in containers and trucks during transportation between warehouses

To optimize the placement of goods in containers and trucks, we collected data on their dimensions and different-format product packaging. AI was applied to get the optimal configuration of the placement of products on pallets and the most efficient use of the available space. We used deep reinforcement learning to train the AI to load containers and trucks better.
Furthermore, we used a convolutional neural network to evaluate the Q-function, which describes the best action for each state. A code with encrypted dimensions, weight, and category of goods is fed to the input of the neural network, and the neural network also receives the current state of the container or truck load through the camera. The neural network determines where the box can be placed and how to rotate it.

Diagram representing the work of the truck/container loading neural network

A trained neural network, implemented as a software and hardware complex, shows the storekeeper on a graphical interface where, how, and what box to put.

By grouping actions together, we have allowed the network to focus on learning the optimal configuration of goods loading rather than how to move packages. Moreover, we implemented transfer learning from the heuristic model, which allowed us to leverage previously learned optimal container/truck loading configurations and significantly improve performance. And by using prioritization, we were able to reduce some of the volatility inherent in learning to evaluate individual actions.

The use of a neural network for loading management has reduced the loading time of containers/trucks by 15% and increased the volume load of one shipment by ~3%.

Third Stage - Blockchain

For data storage we used private blockchain technology with zero transaction fees. All blockchain nodes were on private distributed servers to avoid data spoofing. Furthermore, all data was backed up to a central database to increase the system's overall reliability. This was achieved by forming transaction queues when one of the databases was unavailable and then recording all the data after the database was restored.

Simplified diagram of the interaction between the central database and the blockchain

Final Results

  • 20%
    Reduced rental costs
  • 15%
    Reduced the loading time
    of containers & trucks
  • 3%
    Increased the load
    of one shipment