With 25.000 employees and 2.500 branches Nova Poshta is the biggest private logistics company in Ukraine. The main offering is express delivery of parcels which is extremely important for internet shoppers. As salaries rise and competition is getting harder, staff allocation needs to be improved to prevent over-staffing on the one hand and queues on the other hand. Nova Poshta only started the journey of using machine learning, artificial intelligence etc. to optimize efficiency. Using machine learning we predict the number of parcels/customers for each branch and each hour of the day. The subsequent software component uses this information and creates a work schedule for each branch. It is a classic problem of artificial intelligence (constraint satisfaction problem) and we use the typical algorithms to solve them. I have built the first version of the software by myself and then used the results to convince top management that investing more in this direction is reasonable. After that I was allowed to hire a small team. Now we enhance the product as a team of 6 people and are in the middle of roll-out to production.
Machine Learning powered Language learning tool (https://brainextension.net/) This is my own "side project". I love learning foreign languages and this tool helps me to maintain and expand my vocabulary. It measures and predicts the learning and forgetting pattern for each user. By using this information it optimizes the learning efficiency. It also leverages the translation API of www.deepl.com. I am actively working on this project and will soon start digital marketing in order to make it available for the public. A the code can be seen here: GitHub
GliaLab's goal was to create a SAAS platform that allows radiologists to upload a mammogram and obtain a second opinion regarding potentially maign tissue. I designed and developed the prototype as a web application using Python/Flask running on Nginx. In the backend the platform starts up on demand additional AWS/EC2 instances to run prediction jobs on powerful GPUs and shuts them down when idle. I also developed the entire pipeline of image preprocessing with OpenCV (for example normalize contrast, apply sliding window algorithms, histogram normalization) and machine learning algorithms in Python/Keras/Tensorflow running on AWS. I worked a lot with algorithms like AlexNet, VGG16, ResNet. The platform is a multi-threaded microservices architecture. Modern X-Ray devices create as an output a special file format called DICOM. The platform can read and write DICOM images. GliaLab planned to do an ICO and had very high profile supporters. But parrently they ran into regulatory issues relating to SEC regulations in the USA. Unfortunately the project was then put on hold.
The company so far identified a customer by the mobile phone number. But from a marketing standpoint the phone number does not represent the economic beneficiary because people can have several phones, and buy products on behalf of other people or companies. In order to consolidate the real customer base it was necessary to review the past purchasing history and as well as the geographic location of the purchase. This required the processing of tables with more than a billion records and applying similarity of names calculation using algorithms like Levenshtein, Metaphone, and Soundex. The project was done in Java and PostgreSQL
The company is an investment company searching for investment targets. In order to take decisions, data from various sources need to be combined and visualized. The result needs to be put into a CRM system. This was so far done all manually and the software is now collecting and aggregating the data automatically. This leads to more speed and less manual work. In order words, it increases operational efficiency. Technology stack: Python, Vue.js, PostgreSql