Teaching assistant for the course: "Deep Neural Networks".

Working as a teach lead of the Machine Learning team at Sigmoidal (over 10 people). Designing ML/DL solutions for our clients, supervising multiple projects, including:- Extensive content categorization and clustering solution deployed in 4 languages.- Deep Learning NLP model for content ranking currently deployed in 6 languages with extensive usage of Bayesian treatment via Variational Dropout.- Developing a new algorithm for uncertainty assessment which needs only a single forward pass through a deep neural network (as a part of a research project).- Multiple objects tracking system using deep convolutional neural network and Kalman Filters smoothing deployed on a portable device.- Keyword detection mechanism using neural network and Hidden Markov Model post-processing deployed on a portable device.- Sales prediction on a limited data using Bayesian treatment.Mentoring research projects including:- building algorithm for efficient representation generation in a few-shot learning approach,- building a model for NLP model decision explanation based on variational prediction and policy gradients.

Developing deep learning solutions for computer vision and NLP. Mostly concentrated on image and document classification as well as object and activity detection.Projects I have worked on:- object/object - part/activity detection with tracking: we have built a deep-learning based model for vehicle actions analysis which was deployed on Jetson TX2 - NVIDIA mobile device. We used a YOLOv2 and Microsoft ResNet 101 as feature extractors for a frame flow and late fusion technique in order to provide the final decision. Technologies used: Jetson TX2, Yolo v2, Microsoft ResNet 101, Faster-RCNN (all implemented in Keras / TensorFlow),- article ranking/article classification - we have built a comprehensive article detector and ranker which was used in order to find an article from a fixed domain across multiple feeds with articles from different countries (and with different languages). Due to highly biased data, we used a novel - physics-inspired techniques which we've called diffusion. Thanks to it we were able to augment existing biased flags in such manner that our models achieved the state of the art results. Our system includes also uncertainty estimation (which shows which decisions our model is not so certain) and NER. Technologies used: Convolutional Neural Networks for text, word embeddings, Recurrent neural networks, Variational Dropout, diffusion, Named Entity Recognition, Spacy.

I worked as a researcher / research mentor for TensorCell research group on multiple projects including:- Prediction of results of cellular automaton using neural networks,- Variational Gradient optimization methods for traffic optimization,- Uncertainty estimation,- Active Learning,- Image segmentation.

Developing a deep learning solution for a financial application including a mixture of Recurrent - Convolutional - Static network which blended the information from different data sources in order to solve a classification task.Developing a deep learning solution to solve the pixelwise image segmentation task of an orthophotography data.Developing a mathematical framework for measuring uncertainty of a deep learning solution in a business domain.

Developing a Deep Learning solution for one of company's NLP BI services which augmented risk estimation for automatic translation. Before I was a part of a team that was responsible for the development and maintanance of huge data pipelines for BI. Technologies used: Keras, Scikit-Learn, Azure Data Warehouse, Azure Data Factory, Powershell, SSIS, SQL Server.

preparing and developing mathematical models for measuring behavioral synchrony

Introduction to Mathematics, Linear Algebra

conducting statistical analysis of flight safety data