Method of assessing harmfull websites.
Optimal control calculations.
Detection if illegally connected segemnts.
Implementation of predictive control.
A system capable of detecting anomalies in data produced by web search engine.
First arrival picks from explosions were detected, improvements for geophysical meas., simulations, improved safety, decreasing costs.
Accounting management control system.
Method for generation of optimal sales territories based on geographic and other types of data.
Solution for real time identification for technical issues in ecommerce.
A tool for sharp traders identification allowing more efficient risk management.
A tool for realistic simulation of the cryptocurrency ecosystem.
The current method of assessing website harmfulness is an ad-hoc combination of several metrics. Our job is to evaluate features which are useful for malware detection and develop a machine learning-based approach that would simplify and refine the malicious sites detection.
The development of the final algorithm is still underway. Approaches tested so far include logistic regression, random forest classifiers, boosted oblivious classificatio trees and histogram-based gradient boosting classification trees.
Calculation of optimal control of the substation for the next day (tariff switching and boiler blocking) using consumption prediction and photovoltaic production and subsequent comparison of its advantages with the currently used mass remote control method.
We have created the new optimisation algorithm, which is based on weather forecasting, consumption prediction, PV production and the expected behaviour of consumers when switching tariffs in different situations.
Comparison of the use of mass remote control method and LODIS was analysed on real measured data. The key indicator was the “Losses” – the sum of squares of daily balance runs.
Using the method of Losses comparison by searching for similar days proved significantly lower Losses rate in favor of our LODIS method, which was subsequently implemented.
Identification of illegally connected network segments and places with illegal power consumption.
30 transformer stations
(1200 consumption points)
Determining suspicious network segments
With our method we have achieved 150% accuracy increase in determining the illegal connections, proven by physical inspections of the network (compared to the previously used method).
Implement predictive control of distributed systems using predictive analytics directly within already existing Smart Meters.
Similar approach can be taken to any measuring device collecting data. Instead of sending all the data out, it is possible to do the computations within the device and send only the results.
Our BORT and BOCT algorithms use significantly less CPU power than other conventional AI algorithms, so the prediction models can run directly on Smart Meters.
Our Algorithms are capable of adapting to each individual end-point and are self-improving over time.
We have working systems in place implemented on several Smart Meters and continue to test with others.
Web search engine produces and stores large amount of metadata providing detailed information about processed search requests. This information could be used to estimate the current state and load of the search engine itself and other components involved.
After analyzing the data logs, we have identified important variables and designed statistical aggregations that highlight potential anomalies.
We have implemented algorithms to remove seasonality and noise from the aggregated data and developed a set of detectors to find anomalies in the processed data series.
Finally, we have tuned the detectors’ parameters to optimize the accuracy of the detection.
We’ve developed the anomaly detector that will allow to detect anomalies in real time to alert the operators of possible problems.
Customer examines the composition of rock in front of tunnels. This is done by detecting sound waves reflected from different structures in the rock. Sound waves are generated by explosives or a hydraulic hammer. The task was to create an algorithm for detecting the sound wave from a signal obtained by geophones.
An algorithm that uses changes in the energy of the signal over time and cross-correlation between signals from multiple shots was implemented in Python.
Algorithm for detecting first arrivals of sound waves was created. Proposed algorithm works well for signals generated by explosions, but its success on hammer data is limited. Approaches to create a more complex algorithm more suitable for hammer data was suggested.
Develop a dynamic analytics system for Jaspar s.r.o., an accounting firm specializing in comprehensive accounting and tax record services, aimed at enhancing accounting quality management through detailed monitoring of multiple indicators.
A data warehouse was created to consolidate and optimize data from POHODA accounting systems and Dynamics 365 CRM, enabling efficient data analysis. A Power BI dashboard displayed critical metrics with color-coded indicators to identify accounting status, integrated with client and employee data for tailored access control. The system updates nightly to ensure up-to-date information.
Dashboard-like interface in Power BI, that visualises quality of accounting over multiple clients and indicators.
Customer is a company specializing in web-based mapping. Their main focus is on visualizing geographic patterns, plotting sales territories and other geographical data.
Our task is to design and implement a method for partitioning geographic maps into individual territories in a way that would minimize certain metrics. The solution should supersede the current third party method. It is also expected to provide some additional features and improved functionality over the existing method developed by WeMapSales.
After conducting exploratory data analysis and problem research, we devised a solution based on methods of iterative optimisation. Our approach was based on a type of iterative local search algorithm as well as implementations of various other graph optimisation techniques.
We developed an algorithm for dividing areas within mainland USA into territories capable of optimizing a multitude of metrics such as population variation. Currently work is being done on implementation of additional features and generalization to a wider range of geographic data and metrics.
Develop a real-time anomaly detection framework primarily focused on identifying technical issues experienced by customers during the shopping process. The response time needs to be as rapid as possible due to the high volume of customers and revenues at stake.
Framework for anomaly detection based on Google Analytics data was created.
Identify sharp traders based on their behavior. The easiest way to identify traders possessing some kind of edge over the market is to examine their long-term profits. However, it is challenging to determine what qualifies as 'long enough' and furthermore, it is advantageous to identify sharp traders as soon as possible. The edge could stem from either the technical side or the skill of a client. Examples of sharp trader activity are not readily available since they are typically unique, and the classification of a trader can evolve over time.
Various types of accounts with an edge over the market were discovered. This enables Purple to better manage risks associated with individual clients. The time spent on classification by experts is significantly reduced.
Develop a digital twin able to mimic behavior of the whole ecosystem including interactions of individual entities. Use the solution to run simulation with different parameters and incentives for individual entities of an entire ecosystem.
Digital twin of a XIXOIO ecosystem was created and used for various scenarios. The solution allows to avoid potential problems when parameters of the ecosystem like fees are about to change.
