Using machine learning, I can help you to solve a wide range of problems related to transforming data into useful information by automating decisions and predicting outcomes, giving you a competitive edge. Whether you represent a company or just working on a standalone project, I can implement, update or maintain (ongoing or ad-hoc) statistical algorithms, neural network models or ML pipelines for classification, regression, clustering, dimensionality reduction, anomaly detection and other (mixed) problems.
The following list represents some of my highlighted machine learning projects (excluding open-source contributions and other proprietary or small projects).
Prompt engineering, MLOps
Automatic prompt engineering, testing and load balancing for AI models in production and R&D
— Automatic prompt engineering, testing and load balancing for AI models in production and R&DContent generation, digital marketing
Personalized, LLM-based email marketing automation with user segmentation and A/B testing analytics
— Personalized, LLM-based email marketing automation with user segmentation and A/B testing analyticsComputer vision, dashboards
Medical image/video processing: optical flow (Farneback, Lucas-Kanade, DL), CBIR (SIFT/ORB/CNN embeddings), object detection and Dash interactivity
— Medical image/video processing: optical flow (Farneback, Lucas-Kanade, DL), CBIR (SIFT/ORB/CNN embeddings), object detection and Dash interactivityMarket analysis, recommender systems
E-commerce pricing optimization and dynamic user profiling
PoC tool for demand forecasting, regression and ensemble‐based dynamic pricing optimization; ML-driven user profiling with personalized recommendation systems for real‐time sales and engagement optimization. It can predict future sales based on historical data, competitor pricing and seasonality using time series models, dynamically compute real-time optimal pricing. Profiling is done by clustering users based on behavior tracking (clicks, time on page, purchase history). Users are segmented for personalized recommendations using collaborative filtering techniques.
— E-commerce pricing optimization and dynamic user profilingText analysis, image analysis
A Telegram bot with NLP and vision transformer that detects and moderate messages in real-time, logging actions and auto-removing/warning users. Designed for Telegram groups.
— A Telegram bot with NLP and vision transformer that detects and moderate messages in real-time, logging actions and auto-removing/warning users. Designed for Telegram groups.Anomaly detection, risk analysis
PoC financial fraud detection & credit risk scoring system that uses anomaly detection, ensemble methods, deep NNs, gradient boosting, custom metric evaluation (ROC AUC & Kolmogorov-Smirnov) and explainability techniques on synthetic financial data to simulate real-world risk analytics
— PoC financial fraud detection & credit risk scoring system that uses anomaly detection, ensemble methods, deep NNs, gradient boosting, custom metric evaluation (ROC AUC & Kolmogorov-Smirnov) and explainability techniques on synthetic financial data to simulate real-world risk analyticsContent generation, RAG
PoC chatbot app that combines LLMs with RAG techniques. It leverages semantic text embeddings, vector search using ElasticSearch and LLM integration (via OpenAI API or open-source alternatives) to produce responses for customer service and automated content generation.
— PoC chatbot app that combines LLMs with RAG techniques. It leverages semantic text embeddings, vector search using ElasticSearch and LLM integration (via OpenAI API or open-source alternatives) to produce responses for customer service and automated content generation.Digital marketing, web scraping
Digital marketing analytics solution that scrapes websites for SEO factors and predicts advertisement CTR
— Digital marketing analytics solution that scrapes websites for SEO factors and predicts advertisement CTRRecommender systems
A simple pipeline that collects, preprocesses and labels raw user-item interaction data to build a hybrid recommender system using collaborative filtering (SVD, ALS) and learning-to-rank (XGBoost ranking) methods, evaluated with NDCG and MAP metrics
— A simple pipeline that collects, preprocesses and labels raw user-item interaction data to build a hybrid recommender system using collaborative filtering (SVD, ALS) and learning-to-rank (XGBoost ranking) methods, evaluated with NDCG and MAP metricsTime series, predictive analytics
Time series forecasting system leveraging ARIMA/SARIMA, Prophet, LSTM networks and ensemble methods along with automated feature engineering (lag features, rolling statistics, date-time components) for real-time predictive analytics on simulated streaming data
— Time series forecasting system leveraging ARIMA/SARIMA, Prophet, LSTM networks and ensemble methods along with automated feature engineering (lag features, rolling statistics, date-time components) for real-time predictive analytics on simulated streaming dataTopic modeling
NLP pipeline for text classification and topic modeling using LDA, spaCy, NLTK, BERT model & FastAPI interface
— NLP pipeline for text classification and topic modeling using LDA, spaCy, NLTK, BERT model & FastAPI interfaceCustomer analytics
End-to-end analytics suite that integrates churn/CLV prediction, customer segmentation and lead scoring into a solution with data pipelines and an interactive dashboard
— End-to-end analytics suite that integrates churn/CLV prediction, customer segmentation and lead scoring into a solution with data pipelines and an interactive dashboardOSINT, NLP, web agents, web scraping
LLM-based OSINT tool designed to perform deep web searches by orchestrating multiple web agents and a knowledge agent that uses state-of-the-art machine learning methods. The tool crawls the entire web to gather massive amounts of publicly available data, then leverages advanced LLM techniques to perform natural language tasks on that information.
— LLM-based OSINT tool designed to perform deep web searches by orchestrating multiple web agents and a knowledge agent that uses state-of-the-art machine learning methods. The tool crawls the entire web to gather massive amounts of publicly available data, then leverages advanced LLM techniques to perform natural language tasks on that information.Deep learning
A TensorFlow implementation and demonstration of Kolmogorov-Arnold Network proposed in April 2024 (arXiv:2404.19756)
— A TensorFlow implementation and demonstration of Kolmogorov-Arnold Network proposed in April 2024 (arXiv:2404.19756)Audio analysis
Mixture of experts architecture for speech-to-text and language identification, built in PyTorch
The system combines transformers, RNNs and CNNs. It supports various audio and video formats, and can work with multiple languages and dialects. The project was originally developed as just a pet project on speech recognition to gain some PyTorch skills. The model was trained on private data comprising thousands of hours of speech.
— Mixture of experts architecture for speech-to-text and language identification, built in PyTorchTime series, anomaly detection
Anomaly detection algorithm for time series based on the dynamic threshold generation model
— Anomaly detection algorithm for time series based on the dynamic threshold generation modelComputer vision
A complete road traffic analysis system for roundabouts: real-time processing from multiple cameras, integration with Kafka for messaging, InfluxDB for time series storage and Grafana for interactive dashboards. Processes RTSP streams or local .mp4 files, computes per-road vehicle count and intensity (vehicles per minute), supports parallel processing.
— A complete road traffic analysis system for roundabouts: real-time processing from multiple cameras, integration with Kafka for messaging, InfluxDB for time series storage and Grafana for interactive dashboards. Processes RTSP streams or local .mp4 files, computes per-road vehicle count and intensity (vehicles per minute), supports parallel processing.Question answering, big data analysis
LLM-based Q&A on preloaded docs, raw data, Wikipedia articles and scraped web pages with knowledge graphs, analytics, charts and Streamlit interface
QASATIK is an app dedicated to helping interrogate large volumes of documents, data files and web pages. Built with Streamlit, it supports file uploads, online article scraping and querying using configurable language models (OpenAI, LangChain and LlamaIndex). In addition, it provides interactive knowledge graph visualizations, analytics and charting utilities to help explore and understand the data.
— LLM-based Q&A on preloaded docs, raw data, Wikipedia articles and scraped web pages with knowledge graphs, analytics, charts and Streamlit interfaceAudio analysis
Dockerized benchmark model and API for classifying music by genre based on TensorFlow and Essentia
— Dockerized benchmark model and API for classifying music by genre based on TensorFlow and EssentiaRecommender systems
An implementation of a full recommender system pipeline using PyTorch, Elasticsearch, Redis, Flask, Feast and Triton: from offline data processing and model training to online serving and prediction.
— An implementation of a full recommender system pipeline using PyTorch, Elasticsearch, Redis, Flask, Feast and Triton: from offline data processing and model training to online serving and prediction.Computer vision
My thesis project: the lightweight and fast scene text detection based on EAST architecture and MobileNet layers
— My thesis project: the lightweight and fast scene text detection based on EAST architecture and MobileNet layers
