ML at the railway station

The implementation of a monitoring system for railway carriage and rolling stock based on data from a wheelset sensor, video cameras and GPS data from locomotives at the enterprise allowed to reduce the total turnover time by 5%. Data from external sources was collected in a database management system with temporary indicators. As a result a logical algorithm based on Machine Learning showed the position of the railway carriage on freight fronts and rolling stock in real time. The result and expected effect of the online railway carriage monitoring system is estimated at 5% of the total turnover:

CV at the railway station

To assess the performance of the wheelset sensor and to collect addition information on rolling stock, we implemented a computer vision system based on video cameras from lighting towers. As a result the system shows a high level of accuracy in various weather conditions.

As part of the project to increase the throughput capacity of the slag production warehouses of the metallurgical plant on front-end loaders and at the exit from the checkpoint, a hardware-software complex has been developed that detects and classifies the type of fraction that the loader loads. To develop the complex, we used a computer vision and neural network technologies, as well as an experimental design solution for working in difficult conditions.

The complex increases the throughput of the warehouse and thereby increases the operational efficiency of the sales department. By automating the routine of arranging the access control system and loading, it turns out to halve the implementation time of existing business processes. This increases the maximum throughput capacity of the bulk materials warehouse. This makes it possible to reduce the tariff rate of the carrier, sell more products in the high season and maintain turnover when tightening weight control of freight vehicles in 2020.

Retail ML

Analytics based on the results of the largest federal promo for the «Magnit» chain of stores, in which more than 700 thousand people from 67 regions of Russia took part. We analyzed more than 4 million checks using computer vision and got the following results.
Creation of a recommendation system for Caterpillar, which increased call-to-lead conversion by 40%, based on regression models, classification and clustering of the existing customer base and analysis of previous sales.

Retail CV

Retail of various formats is increasingly implementing traffic data inside and outside the store into their Data Lake. In practice, we implemented a data collection system for the "Adamas" jewelry chain and "Slata" grocery retail. At the entrance to the store there were cameras aimed at entering the outlet, the task of the system was to determine the uniqueness of the user (visit), identify the employee, and also collect gender signs of traffic. Additionally, on the same cameras, a warning system was implemented for security personnel if a person from the blacklist appeared at the outlet.
A / B was conducted with the marketing department, testing new hypotheses to understand the best customer experience. We selected two stores in the "Slata" chain with the same audience based on demographic data, as well as unique visits. In this store we select the same type of zone (for example: grocery). And by the example of this zone, we begin to conduct testing: in the A-test, the goods are put on a shelf in the middle, in the B-test on the top shelf. The key factor is the measurement of the effectiveness of the A-test relative to the B-test under the same conditions, which is the key factor for conducting A / B testing.

Development of a dashboard and ETL system for converting data from social networks and search engines. The goal is to understand and mention brands related to the respective user clusters. This BI system completely replaces panel polls and gives businesses insights into Market Desire and Brand Awareness. The system gives a clear understanding of the distribution of information load, both among a certain segment of users (For example: "Creative Class"), and generally in the market among competitors.

Automation of inventory and search for stocks using an interactive chat bot using Natural Language Processing (NLP) with a nomenclature of more than 1,000 positions.

Due to the large nomenclature and the lack of a single warehouse base, as well as the difference in the nomenclature names of suppliers, the names of the nomenclature for managers and logisticians in the warehouse were different (for example, the manager has a “cap”, the logistician has a “baseball cap”). The solution was implemented as a chatbot with a strong server part with morphological processing of the language and NLP.
The chatbot selected from thousands of nomenclature items the positions corresponding to the search phrase. For example, when you request "red jacket" (a typo in the word jacket), to the base in which there is a position: red coat, green windbreaker, red windbreaker, red pants, etc. the bot would show a "red windbreaker" and a "red cloak." Including the bot noted arrival / departure, made the necessary documents and sent reports by e-mail. As a result of the introduction of the bot, the number of logisticians decreased from 4 to 2 people, the search time for positions decreased by an average of 2 times, and about 40 people / hours per month were saved on filling out documents.

Drones are equipped with various technologies, such as infrared cameras, GPS and lasers (this applies specifically to military models). Drones can be controlled by a remote system, sometimes also called a ground-based cockpit. That is, we can say that the drone consists of 2 parts: the drone itself and its control system. For orientation in space, drones use communication with a ground cabin or with GPS satellites. But there are cases when the drone needs to navigate in space in the radio silence mode. In this regard, the only way of orientation in space for the drone is to compare the area where at the moment it is located with the maps that are loaded into it.