MACHINE LEARNING (ML)
Cost reduction
Improving competitiveness
NEURAL NETWORKS AND ARTIFICIAL INTELLIGENCE
Automation of production processes
Machine learning is a broad concept that includes the use of various mathematical models to solve different problems. The business card of ML is the ability of the model to improve the accuracy of its forecasts and solutions based on training data (datasets).
Machine learning and artificial intelligence find their application in almost all areas of business.
WHERE MACHINE LEARNING CAN BE APPLIED
Our team has implemented more than 100 machine-learning projects, from the cell colonies classification for the Human Stem Cell Institute to creating a 3D object classifier for the CERN charged particle detector.
It is also worth considering that artificial intelligence is not only used in production. Besides, it can be applied in many other widespread processes: recruiting, marketing, sales, purchasing, technical support, document management, and security.
production
medicine
agricultural industry
E-COMMERCE
WHAT TASKS DOES MACHINE LEARNING SOLVE?
WORKING WITH IMAGES AND VIDEOS
Classification, segmentation, detection, and counting of objects in photos and videos
- Detection of defects in production
- Classification of medical images
- Recognition of product types and counting their quantity
- Analysis of geological images
- Monitoring of compliance with safety regulations
- Recognition of human poses
- Content moderation
- Classification of medical images
- Recognition of product types and counting their quantity
- Analysis of geological images
- Monitoring of compliance with safety regulations
- Recognition of human poses
- Content moderation
Where it is used:
WORKING WITH AUDIO RECORDINGS AND TEXT
Recognition and synthesis of natural language, search for speech patterns and entities, determination of emotional connotation - Voice assistants.
- Chatbots
- Script control for operators
- Assistance systems and prompts
- Classification and processing of documents
- Data matching
- Text analysis and compilation
- Selecting relevant news or articles
- Script control for operators
- Assistance systems and prompts
- Classification and processing of documents
- Data matching
- Text analysis and compilation
- Selecting relevant news or articles
Where it is used:
DEVELOPING PREDICTIVE AND FORECASTING MODELS, DATA ANALYSIS
- Prediction of marginality, demand, commodity balances, production workload
- Simulation modeling
- Developing scoring models
- Process mining
- Simulation modeling
- Developing scoring models
- Process mining
Where it is used:
GENERATING TEXTS, IMAGES, AND VIDEOS
- Writing articles
- Chatbots
- Generation of images and posts on social networks
- Generation of advertising banners
- Dynamic rendering of advertising banners to the desired format
- Chatbots
- Generation of images and posts on social networks
- Generation of advertising banners
- Dynamic rendering of advertising banners to the desired format
Where it is used:
WE WILL CONSULT YOU ON ALL THE POSSIBILITIES OF MACHINE LEARNING TECHNOLOGY. WE WILL WORK OUT THE MOST PROSPECTIVE IMPLEMENTATION SCENARIOS
By clicking the «Submit» button, you agree with the policy of personal data processing.
LEAVE YOUR CONTACTS, AND WE WILL CALL YOU BACK FOR A CONSULTATION
OUR CASES
Implementation period
3 MONTHS
CLASSIFICATION OF SKIN LESIONS
Determine the main characteristics of the skin from a phone photo of the face
As part of implementing a mobile application for a network of clinics, it was necessary to develop a model that defines the main skin characteristics of the face. At the same time, the model had to process the photo in a fraction of a second.
01 Together with the customer, a dataset of more than 3000 images was prepared and marked up in our CVAT markup.
02 To detect skin lesions, the Fpn_vg16 model was taken as a basis in combination with the FeaturePyramid Network, which used Vgg16 predictions at each level. For the mobile version, the model has been converted to Tflite.
03 03Plenty of models has been developed to determine the indicators such as skin lesion type (acne or pimples), the percentage of skin lesions severity, skin type (dry, oily, combined).
02 To detect skin lesions, the Fpn_vg16 model was taken as a basis in combination with the FeaturePyramid Network, which used Vgg16 predictions at each level. For the mobile version, the model has been converted to Tflite.
03 03Plenty of models has been developed to determine the indicators such as skin lesion type (acne or pimples), the percentage of skin lesions severity, skin type (dry, oily, combined).
97
The accuracy of the lesion classification was achieved in 97% of cases.
%
3000
0,5
sec
The speed of photo processing on a mobile device.
Images were used to train a neural network.
in
Implementation period
3 MONTHS
CLASSIFICATION OF BRANDED BAG MODELS
Determine the brand and model of the bag from a photo taken at home.
For the customer (marketplace of branded bags, retail), it was necessary to develop a model, the input of which is a photo of a bag uploaded by the user, and the output is the identification of its brand and model to determine the recommended selling price.
01 A dataset of more than 45,000 images has been collected with the customer. To increase the volume, a part of the dataset has been synthetically generated by our specialists.
02 Models have been trained and configured to detect and classify bags.
03 The classifier has been implemented based on Resnet50, 400 epochs, 32 images per patch.
02 Models have been trained and configured to detect and classify bags.
03 The classifier has been implemented based on Resnet50, 400 epochs, 32 images per patch.
175
Most popular models of bags are determined by the neural network.
45000
95
%
Accuracy of classification.
Images were used to train the neural network.
Implementation period
4 MONTHS
CREATING DIALOG BOTS IN CRYPTOCURRENCY CHATS
Write a system of dialog bots simulating communication on the topic of cryptocurrencies and abstract themes to maintain the involvement of subscribers of the community in Telegram.
The task was to create a tool to help promote cryptocurrencies communities in Telegram. For this purpose, we have created a dialog bot management system. The difficulty was that the topic of cryptocurrencies is quite specific in terms of the abundance of industry terminology and slang, which complicates the training of bots to a level of communication indistinguishable from humans.
01 To train the dialog model, 25 relevant telegram channels were parsed, and more than 100,000 records dialogs were uploaded.
02 Dialogues were cleaned from stop words and undesirable topics (names of particular moments, unfriendly communication, etc).
03 The model based on GPT-3 Davinci was trained using the collected dialogues.
04 Several bot communication modes have been created. Bots initiate communication with users, respond to user replies, and communicate with each other. The bots are configured to work according to a schedule close to people's one (activity time, sleep).
05 A bot control panel has been created (creating bots, adding and removing bots from channels, selecting operating modes, activity statistics, and logging events).
02 Dialogues were cleaned from stop words and undesirable topics (names of particular moments, unfriendly communication, etc).
03 The model based on GPT-3 Davinci was trained using the collected dialogues.
04 Several bot communication modes have been created. Bots initiate communication with users, respond to user replies, and communicate with each other. The bots are configured to work according to a schedule close to people's one (activity time, sleep).
05 A bot control panel has been created (creating bots, adding and removing bots from channels, selecting operating modes, activity statistics, and logging events).
10
Chat activity increased by 10 times
100000
3
times
The growth rate of new subscribers has increased by 3 times
records were used to train bots about cryptocurrencies
by
by
times
AND WE WILL CALL YOU BACK FOR A CONSULTATION
шаг 1/2
LEAVE YOUR CONTACTS
