Skills and Qualifications
Languages
- Python
Frameworks
- PyTorch
- NeMo
Libraries/APIs
- Keras
- Tensorflow
- Tensorflow-Serving
- sklearn
- Librosa
Other
- Transformers
- CNNs
- RNNs
Featured Projects
Projects
Dozor
Team Lead Data Scientist
This project was for a car-sharing company. The task was to develop an Android app that would utilize facial recognition to check whether the driver is the one who has been authorized by the app to drive. The client also wanted the app to be able to check whether people in the car are smoking. The facial recognition functionality was implemented via open libraries, while the smoking detection function had no prior libraries and so needed to be built completely from scratch. Anton conducted extensive research on how this function could be developed. Part of this process involved him gathering substantial labeled visual data of people smoking in their cars. After that was done, Anton and his team experimented with neural networks to develop the most accurate visual smoking detector system possible. The final system was able to work offline in the embedded systems that will be installed in the cars.
BusFactor
Team Lead Data Scientist
The main task of the project was to monitor public transport (buses, trams, etc.) drivers and record the cases of mobile phone usage while driving. The system was also required to check whether the camera in the driver's cabin is rotated correctly and monitor it to ensure that it is working all the time. Anton also organized the deployment of the final models to the client's machine.
Recognition of goods on the shelves
Team Lead Data Scientist
A solution was needed for merchandisers who compare the real arrangement of goods (realograms) with a plan (planogram). OCR and one-shot learning were used to search the product database using triplet loss network training. The generation of images in Blender was required, as was the development of tools marking real goods. Anton experimented with various detection and classification approaches, taught OCR models, developed image generation using Blender, and ported and optimized models for Android devices.
Hitfactor
Data Scientist
A project that was required to find athletic starting signals or shots on a soundtrack. Anton helped develop an architecture using convolutional and recurrent neural networks. He was also required to allow the system to be able to separate shots coming from other directions, which are not important. For this, Anton used clustering algorithms from a library.
