BIOCAD: PPE compliance monitoring at a pharmaceutical plant.
- CV, Real-Time Detection
- Technologies: detectron, ultralytics, Roboflow
- YOLO and FasterRCNN models were evaluated
- The work resulted in a published paper — see achievements
HSE University, a first-year course project.
- CV: Math Recognition, OCR, Detection; NLP: RAG
- Technologies: ultralytics, detectron, Roboflow, PyTorch, transformers, FAISS
- A synthetic dataset was built for detection; YOLO and FasterRCNN were trained
- A ViT was trained for math recognition, a pretrained model used for text
A solution for a competition on detecting language model hallucinations.
BIOCAD: instance segmentation when labelled data is scarce.
- CV: Instance Segmentation (Few-Shot / One-Shot / Small Data)
- Technologies: transformers, detectron, PyTorch, OpenCV
- Compared approaches: Supervised Learning, Zero-Shot, Few-Shot, Transfer Learning
The Alabuga hackathon.
- NLP: Sentiment Analysis on Named Entity Recognition
- Technologies: natasha, transformers
- Named entities recognised with a pretrained model from natasha
- Two BERT models fine-tuned with masking entity: neutral vs non-neutral and positive vs negative
- catboost tested as a baseline: text_features, word2vec, TF-IDF
Content moderation for the Pulse social network
T-Bank: multilabel text classification.
- NLP: Multilabel Text Classification
- Technologies: keras, catboost, sklearn, fasttext
- Tested CNN, RNN and catboost with fasttext embeddings (catboost + TF-IDF as a baseline)
- Focal Loss tried against class imbalance
Gender balance in company management
The DANO olympiad: an analytical study.
- Technologies: pandas, numpy, scipy, sklearn, matplotlib, seaborn
- Cleaned the dataset of outliers and missing values
- Hypotheses tested with robust bootstrap methods: permutation test and Bootstrap Confidence Interval
Leaders of Digital Transformation and Itelma: a web service.
- Time Series
- Technologies: catboost, numpy, pandas, sklearn, scipy
- Preprocessing pipeline: anomaly detection, interpolation of gaps, denoising
- System for detecting key events and predicting complications (catboost, RandomForest)