We create and implement solutions powered by Artificial Intelligence, enabling sustainable competitive advantages in the digital era
Artificial Intelligence & Machine Learning
We create and implement solutions powered by Artificial Intelligence, enabling sustainable competitive advantages in the digital era
Solutions
ML-models for financial credit risk management
We develop and implement models for assessing credit risks in the financial sector (PD, LGD, early warning models, macroeconomic models for IFRS9)
Accurate forecasts and planning optimization
We develop services for forecasting multidimensional time series (sales volumes, warehouse loading, information systems load) using ML models
Detection of the hidden threats
We create services for detecting anomalies based on ML models
ML-driven HR management
We develop and implement models for the HR domain: forecasting employee termination and automating the search for employees based on ML-models
Customer retention, churn handling and cross-sell recommendations
We develop ML-models of customer churn, NBA/NBO models, AI-based recommendation systems
NLP for improved operational efficiency
We develop natural language processing (NLP) services for processing customers' requests, news information, automated parsing and verification of text documents
Solutions
ML-models for financial credit risk management
We develop and implement models for assessing credit risks in the financial sector (PD, LGD, early warning models, macroeconomic models for IFRS9)
Accurate forecasts and planning optimization
We develop services for forecasting multidimensional time series (sales volumes, warehouse loading, information systems load) using ML models
Detection of the hidden threats
We create services for detecting anomalies based on ML models
ML-driven HR management
We develop and implement models for the HR domain: forecasting employee termination and automating the search for employees based on ML-models
Customer retention, churn handling and cross-sell recommendations
We develop ML-models of customer churn, NBA/NBO models, AI-based recommendation systems
NLP for improved operational efficiency
We develop natural language processing (NLP) services for processing customers' requests, news information, automated parsing and verification of text documents
Case Studies
DEVELOPMENT OF PD, LGD MODELS AND MACRO ADJUSTMENT FOR BANKS
Customer
A number of banks in Europe, Africa and Middle East countries
Task
To comply with the requirements of regional and European regulators in terms of the IFRS9 financial reporting standard, the banks should develop models for assessing the borrowers' credit risk: default probability models, loss given default models and macroeconomic adjustment models.
Outcomes
A developed stack of models estimating observed default rates (ODR) using a vintage approach, a number of implemented algorithms for fitting parametric curves to ODR (including Weibull-Pareto, Cox PH, Gen Gamma, RF, etc.). LGD models, automated backtesting procedures, and macroeconomic adjustment models (regression models and SVMs) have been developed. A series of successful external audits of Big-4 companies (EY, PWC) has been conducted.
WAREHOUSE STOCK VOLUMES FORECASTING
Customer
Large logistics operator. 200+ regional branches, 400 000 sq.meters of warehouses, 10 000+ employees, 3,5M+ clients
Task
Implement a model for forecasting the warehouse stock volume in all departments of the logistics company to increase the accuracy of transport resources planning.
Outcomes
Enfint has developed a set of time series models using various approaches (STL + SARIMA, ARIMA+Fourier, TBATS, Prophet, LSTM). The SARIMA model has shown the best results; median weighted percentage error is ~28%. The accuracy of the developed planning algorithm has allowed the planning accuracy to be increased by more than 5% compared to the previously used algorithms.
EMPLOYEE CHURN FORECASTING MODEL
Customer
Large international IT company, 1200+ employess in 3 countries, 100+ clients
Task
To reduce the turnover of front office employees (developers, testers, system analysts), we need to identify a set of key factors affecting a significant increase in the probability of employee termination.
Outcomes
A developed model using AutoML approaches (the best result was obtained using the Distributed Random Forest algorithm), AUC = 0.64 and recall of terminated employees is approximately 0.68. This model helps identify employees who are in the "risk zone" based on more than 50 factors which, in turn, enables vertical managers and HRs to specifically intensify actions to retain employees and take timely measures.
FORECASTING OF CUSTOMER DEMAND
Customer
Large distribution company, 30+ years on the market, 42 warehouses, 600 suppliers, 4000+ employees
Task
To develop a set of algorithms based on the basis of machine learning tools, which will be used to automatically forecast the demand for the purchase of goods in the company's warehouse system.
Outcomes
We have developed a model which uses ML-algorithms of time series forecasting. With this model, it is possible to forecast the demand for particular products or groups of product categories on the horizon of 1-2 months
For different product categories (depending on the demand characteristics), the model uses different algorithms (ARIMA, Exponential smoothing, Prophet, LSTM) and predictor sets.
New models improved existing forecast accuracy for 5-10% depending on the region/group of goods
GRAIN YIELD FORECASTING MODEL
Customer
A large manufacturer of plant protection products
Task
To develop a service for scenario modeling of spring and winter wheat gross yield forecast, taking into account the dependence on weather conditions
Outcomes
Developed a set of models based on wheat gross yield time series decomposition into three components: forecast of area under crop, yield trend and yield deviations from trend.
In order to forecast the area under crop, Holt-Winters model was used, the forecast of the yield trend was estimated using linear regression from time, the deviations from the trend were projected using the LightGBM algorithm.
The accuracy of the forecast for all the regions exceeds 80%
Developed a cloud service for scenario modelling based on the created ML-models.
QUEUE CONTROL USING COMPUTER VISION
Customer
Large retail company
Task
Develop of an automated service for controlling queues in stores to increase customer loyalty and reduce the level of lost customers
Outcomes
The speed of response to the emergence of a queue has increased by 3 times
A model was developed using Yolo and its own algorithm for determining the queue
The algorithm for determining the queue "in the future" has a number of advantages over the standard use of mounting cameras above the cash registers:
Ability to work in rooms with low ceilings
Determination of groups of buyers (families) to eliminate false positives
Case Studies
DEVELOPMENT OF PD, LGD MODELS AND MACRO ADJUSTMENT FOR BANKS
Customer
A number of banks in Europe, Africa and Middle East countries
Task
To comply with the requirements of regional and European regulators in terms of the IFRS9 financial reporting standard, the banks should develop models for assessing the borrowers' credit risk: default probability models, loss given default models and macroeconomic adjustment models.
Outcomes
A developed stack of models estimating observed default rates (ODR) using a vintage approach, a number of implemented algorithms for fitting parametric curves to ODR (including Weibull-Pareto, Cox PH, Gen Gamma, RF, etc.). LGD models, automated backtesting procedures, and macroeconomic adjustment models (regression models and SVMs) have been developed. A series of successful external audits of Big-4 companies (EY, PWC) has been conducted.
WAREHOUSE STOCK VOLUMES FORECASTING
Customer
Large logistics operator. 200+ regional branches, 400 000 sq.meters of warehouses, 10 000+ employees, 3,5M+ clients
Task
Implement a model for forecasting the warehouse stock volume in all departments of the logistics company to increase the accuracy of transport resources planning.
Outcomes
Neoflex has developed a set of time series models using various approaches (STL + SARIMA, ARIMA+Fourier, TBATS, Prophet, LSTM). The SARIMA model has shown the best results; median weighted percentage error is ~28%. The accuracy of the developed planning algorithm has allowed the planning accuracy to be increased by more than 5% compared to the previously used algorithms.
EMPLOYEE CHURN FORECASTING MODEL
Customer
Large international IT company, 1200+ employess in 3 countries, 100+ clients
Task
To reduce the turnover of front office employees (developers, testers, system analysts), we need to identify a set of key factors affecting a significant increase in the probability of employee termination.
Outcomes
A developed model using AutoML approaches (the best result was obtained using the Distributed Random Forest algorithm), AUC = 0.64 and recall of terminated employees is approximately 0.68. This model helps identify employees who are in the "risk zone" based on more than 50 factors which, in turn, enables vertical managers and HRs to specifically intensify actions to retain employees and take timely measures.
FORECASTING OF CUSTOMER DEMAND
Customer
Large distribution company, 30+ years on the market, 42 warehouses, 600 suppliers, 4000+ employees
Task
To develop a set of algorithms based on the basis of machine learning tools, which will be used to automatically forecast the demand for the purchase of goods in the company's warehouse system.
Outcomes
We have developed a model which uses ML-algorithms of time series forecasting. With this model, it is possible to forecast the demand for particular products or groups of product categories on the horizon of 1-2 months
For different product categories (depending on the demand characteristics), the model uses different algorithms (ARIMA, Exponential smoothing, Prophet, LSTM) and predictor sets.
New models improved existing forecast accuracy for 5-10% depending on the region/group of goods
GRAIN YIELD FORECASTING MODEL
Customer
A large manufacturer of plant protection products
Task
To develop a service for scenario modeling of spring and winter wheat gross yield forecast, taking into account the dependence on weather conditions
Outcomes
Developed a set of models based on wheat gross yield time series decomposition into three components: forecast of area under crop, yield trend and yield deviations from trend.
In order to forecast the area under crop, Holt-Winters model was used, the forecast of the yield trend was estimated using linear regression from time, the deviations from the trend were projected using the LightGBM algorithm.
The accuracy of the forecast for all the regions exceeds 80%
Developed a cloud service for scenario modelling based on the created ML-models.
QUEUE CONTROL USING COMPUTER VISION
Customer
Large retail company
Task
Develop of an automated service for controlling queues in stores to increase customer loyalty and reduce the level of lost customers
Outcomes
The speed of response to the emergence of a queue has increased by 3 times
A model was developed using Yolo and its own algorithm for determining the queue
The algorithm for determining the queue "in the future" has a number of advantages over the standard use of mounting cameras above the cash registers:
Ability to work in rooms with low ceilings
Determination of groups of buyers (families) to eliminate false positives