Dmytro Klepachevskyi
ML/AI engineer with 4+ years of experience working with machine learning, computer vision, and deep learning. Passionate about pushing the boundaries of AI capabilities with a focus on real-world applications.
Currently pursuing MSc in Artificial Intelligence, with 6 research publications and experience in machine learning, statistical data analysis, biomedical signal processing, NLP, and deep learning.
Projects
-
3D semantic segmentation with point cloudsTrained and fine-tuned various 3D semantic segmentation models using large point cloud dataset (100K+ points per sample). Achieved 90% test accuracy (increase of 50%+ over the baseline solution) on 3 classes on a challenging medical domain point clouds. Applied hyperparameter tuning techniques to find optimal configurations.
-
Abstractive and Extractive Text Summarization for Lithuanian and English languagesEngineered a pipeline for abstractive and extractive summarization models for Lithuanian language. Fine-tuned various classic and LLM approaches, which included models mT5, mBART, BertSum, and LLM models Gemma, Llama 3.1, Mistral, Qwen. Managed to increase the Rouge-2 F1 score from initial scores of 10.1% up to 40.5% which depicts a high bigram matching, and also achieved BertScore semantic similarity of 92%. Used different fine-tuning optimization techniques.
-
BERT-based models fine-tuning for Lithuanian languageFine-tuned RoBERTa for Named Entity Recognition, achieving 99.2% token accuracy across 8 entity classes.
-
Malnutrition risk predictionAnalyzed clinical data with imputation strategies, improving accuracy from 60% to 92% and reducing features by 97%.
-
Parkinson’s disease automated detectionProcessed 116-channel signals, achieving 87.5% accuracy and improving prior solutions by 15%.
Publications
-
Synthetic 4D human datasets for robust pose estimation.
-
Statistical analysis and predictive modeling for malnutrition.
-
MEG data analysis for biomarkers in Parkinson’s.
-
Classification models for neurodegenerative subtypes.
-
Machine learning on MEG data for Parkinson’s biomarkers.
-
Forecasting framework using pattern recognition.
Education
-
MSc in Artificial IntelligenceResearch in object re-identification, 3D tracking, and deep learning for 3D reconstruction.
-
BSc in Computer ScienceThesis on Parkinson’s diagnosis using ML, graded 100%.
-
BSc in Applied System Analysis and Data ScienceThesis on reasoning strategies with LLMs, graded 90%.
Learning Wiki/Resources
A curated collection of resources for AI, machine learning, and computer vision learning.
-
Comprehensive introduction to deep learning concepts and neural networks.
-
Access to the latest research papers in AI and computer science.
-
Hands-on tutorials for building neural networks with PyTorch.
-
Machine learning tutorials and examples using Scikit-Learn.
News
Recent updates and announcements.
-
New Publication on arXivReleased "Avatar4D: Synthesizing Domain-Specific 4D Humans" preprint.
-
Conference PresentationPresented research on Parkinson’s biomarkers at LNA 2024.
-
MSc Program StartBegan MSc in Artificial Intelligence with focus on 3D reconstruction.
Contacts
GitHub
github.com/dklpp