Client problem: Traffic analysis might be a challenging problem to address. In particular, some of the widespread issues are overlapping vehicles trajectories, objects moving outside of the visible area of the cameras etc. Due to the said problems our customer had some inaccuracies in their solution such as double classification of the same object and difficulties in vehicles counting and average speed measurement.

Delivered solution: A complex AI-based solution efficiently analyzes and forecasts traffic (includes vehicles, pedestrians, other objects) and builds the optimal model of its maintenance, which increases the marginality of a business. Compared to the solution our customer had, we managed to increase detection accuracy by 35 percent and track accuracy by 30 percent.

Tech stack: Python, PyTorch, Yolo v5, IOU Tracker

Client problem: High operational costs for determining the condition of air power transmission lines and low accuracy of defects detection performed by traditional methods.

Delivered solution: In one flight, the drone examines dozens of kilometers of power lines. Photos are submitted for processing to several neural networks. Among thousands of photographs, artificial intelligence detects deviations from technical standards (insulator defect, line break). The geo-coordinate of an object with defects is transmitted to the dispatcher's console for planning a preventive maintenance schedule. As a result, preventive maintenance helps to reduce emergency crew hours and significantly reduce the cost of maintaining high voltage lines.

Tech stack: Python / PyTorch / Yolo5 / Detectron2 / 1С

Client problem: Low quality of the tracker of people, due to the lack of accuracy in determining the time and date of the event for the video analytics system.

Delivered solution: A neural network was selected and trained for recognizing the date and time on the frames of the video stream. A sufficient dataset was assembled with digits in various fonts, that was key success factor. As a result, the accuracy of the people tracker has improved. This approach can be used for systems requiring document recognition: passport, driver's license, etc.

Tech stack: Python / PyTorch / CRNN

Client problem: Lack of reliable and detailed information on client flows for disisions making in the business efficiency increasing.

Delivered solution: Analysis of the flow from IP cameras, detection of people, construction of their tracks using computer vision technologies. This allows to get statistics and calculate metrics, for example, conversion to target action, build dashboards and analytical reports.

Tech stack: Python / OpenCV / Yolo v3 / PyTorch / Airflow / Clickhouse

Client problem: Power grid companies annually formulate plans for clearing the routes of overhead lines from trees. The collection of data for this is carried out during pedestrian inspections and requires significant staff labor. Therefore, the organization has high operating costs for the execution of the process.

Delivered solution: Zeus AI analyzes space images using artificial intelligence and provides analytics to regulatory authorities and network organizations to make informed management decisions. The system is able to identify areas with trees, height of vegetation and tree species.

Tech stack: Python, PyTorch, TensorFlow, OpenCV, Airflow, PostgreSQL

Client problem: It is necessary to determine the category of the museum fund objects by the description or image. The State catalog of the Museum Fund of the Russian Federation is the main source of data.

Delivered solution: Optimal ML-models were selected and trained, an ensemble models was built. Image classification is performed using the EfficientNet-b3 neural network. The text is classified using TF-IDF and SVM.

Tech stack: Python / PyTorch / EfficientNet / scikit-learn

Client problem: Passengers injures on rail transport when getting on / off (jamming by doors, falling between platform and wagon) due to difficult weather conditions, blind spots, distractions cause which affects on train driver.

Delivered solution: Transfer the routine tasks of recognizing anomalous situations in the field of transport security to a "digital assistant" and thereby reduce the role of the human factor in decision-making through more efficient use of computer vision hardware and software. A concept of a "digital assistant" framework and a prototype for analyzing 3D point cloud from ToF cameras to identify dangerous situations have been developed.

Tech stack: Python / OpenCV / PyTorch

Client problem: Informing the company's IT-department about the construction readiness of facilities for timely planning works of starting-ups and adjustment.

Delivered solution: Classification of objects construction readiness on their photo by using computer vision:

• Drone flying around a long-distance construction site and obtaining objects photo
• Uploading photos of objects to the system and their automatic classification
• Reporting on construction readiness

Tech stack: Python / Yolo v5 / PyTorch / Cameratransform / 1С

Client problem: A business loses / doesn’t gain potential profit if:

• it has not enough servers, so customers who intended to get a service leave without one;
• it has so many servers, that expenses for them are not worth the number of customers they can serve.

That forces businesses to keep a fragile balance between "too much" and "too little" active servers, which is a challenging optimization problem.

Delivered solution: Given historical data, a tool called sQeduler predicts the expected number of customers, simulates how much profit they’ll bring in different cases and, based on that, predicts the number of active servers that'll yield the greatest profit.

Tech stack: Python, PyTorch, scikit-learn, statsmodels, pdmarima, SQLite, pandas, scipy

Client problem: First In First Out (FIFO) queueing system are believed to be built on a Poisson process. That randomizes the flow of customers in such a system to a certain extent, hence, complicates understanding of how the system would behave.

Delivered solution: A database-driven model named QueTEr simulates a FIFO queueing system with various parameters stepwise scheduling and managing numbers of events with regard for the Poisson process. After modelling, QueTEr calculates key metrics such as average queue before customers on their arrival, average waiting time, average service time etc. which provides managers information necessary for making decisions profitable for the quality of service.

Tech stack: Python, SQLite, pandas, scipy

Client problem: Different time series forecasting methods are designed for time series with different features (trend, seasonality, stationarity etc.), which means upon forecasting large numbers of time series with one method some of them will probably be forecasted incorrectly.

Delivered solution: A pre-trained classifier NN4TSF predicts the best method of time series forecasting (from a predefined pool) for each time series separately; using individually predicted by NN4TSF methods rather than a single method decreases MAPE so that NN4TSF outperforms all the methods from the pool.

Tech stack: Python, PyTorch, scikit-learn, statsmodels, pdmarima

Client problem: Implementation of a model for predicting average prices and consumption of certain groups of goods in 3 months horizon. The dataset of the Federal Tax Service contains prices and sales volumes for 32 grocery items, 4 types of fuels and consumer baskets (5,577,874 facts of observations for 821 days in 86 regions).

Delivered solution: We chose a wide range of models for research (regressive, neural networks, ARIMA, exponential models), built a model based on the training data, made predictions (horizon - 92 days) and evaluated the efficiency (MAPE and RMSE metrics), built top models.

Tech stack: Python / statsmodels / Keras / pmdarima / scikit-learn

Client problem: Clickbait is a deceptive manner to attract users to follow the linked attachment; the contents of the linked page does not usually correspond to the expectations of users and in some cases may be inappropriate, offensive or harmful.

Delivered solution: An ensemble model WebSemble Light extracts a fragment of the linked document that is supposed to satisfy users’ curiosity, which should withhold them from following the link.

Tech stack: Python, PyTorch, transformers

Client problem: Natural Language Processing (NLP) tools for the German language are fairly underdeveloped compared to English; in particular, there is no universal model for inflection (declination, conjugation) of German words.

Delivered solution: DERBI (a rule-based inflector) models inflection of German words with a finite-state-transducer-like patterns for regular rules and lexicons for exceptions. It is based on a powerful framework for NLP named spaCy and applicable regardless of part of the speech.

Tech stack: Python, spaCy

Client problem: Standard lead generation tools are ineffective for promoting ML-agency services on Telegram or through networking.

Delivered solution: Creation of a lead-generating chatbot in Telegram, which, on behalf of a company employee, conducts a dialogue with a client. Key bot features:

• bot character developed
• interaction script is executed with gamification elements
• the bot demonstrates the technical skills of the company through the opportunity to try ML-cases (identification of human emotions by photo, forecasting client business revenue, etc.)

Tech stack: Aimylogic