Data scientist

SATELLITE NAVIGATION

About the Project

Developing a GPS-denied navigation system for UAVs using satellite imagery matching. The system takes photos from the drone's onboard camera and finds matching terrain patterns in pre-loaded satellite maps to determine the drone's position. This enables autonomous navigation in areas where GPS signals are jammed, spoofed, or simply unavailable. The matching pipeline needed to handle significant differences between real-time aerial photos and satellite basemaps captured at different times and conditions.

• UAVs/drones
• Defense

Responsibilities

Built the core image matching pipeline using SuperPoint and SuperGlue for robust feature detection and matching across different image domains. Worked with Planet API satellite imagery for basemap preparation and Rasterio/GeoPandas for geospatial data processing. Set up the simulation environment using Unreal Engine with Cesium terrain data and ArduPilot for controlled testing. Handled deployment considerations for edge hardware and maintained the system through iterative improvements based on flight test results.

OMT

About the Project

A set of AI services for smart city management built on open data from Barcelona and Singapore. The system includes modules for tabular classification and clustering, time series forecasting, simulations, and automated expert decision generation. A key component analyzes traffic congestion and air pollution patterns, generating recommendations for traffic restrictions in response to accidents and environmental conditions. The project demonstrated how ML can support urban decision-making at city scale.

• IoT
• Smart Cities
• Transportation
• Analytics

Responsibilities

Built multiple ML services from research through deployment: classification and clustering models for urban data analysis, time series forecasting for traffic and pollution trends, and a simulation environment for testing policy scenarios. Used PyCaret for rapid model prototyping and scikit-learn for production models. Handled geospatial analysis of traffic and pollution data using GeoPandas. Built the reinforcement learning module (using OpenAI Gym) for the decision generation service that recommends traffic interventions based on real-time conditions.

RL WEB

About the Project

An internal R&D project exploring AI-driven website interaction automation for web testers. The system identifies UI elements on a webpage and translates natural language user requests into specific mouse and keyboard actions. Two approaches were tested: an intent classification and slot filling pipeline (using BERT and DIET) to map user text to interface elements, and an end-to-end approach that directly converts text instructions into mouse actions. Built with LayoutLM for understanding page structure.

• AI
• Machine Learning

Responsibilities

Researched and implemented both the intent classification pipeline and the end-to-end approach. Built the NLU component using BERT and DIET architectures for intent recognition and slot extraction from user commands. Worked with LayoutLM for understanding webpage layout structure and mapping between text descriptions and visual elements. Developed the action execution layer that translates classified intents into actual browser interactions. Used PyTorch for all model development.

CORVUS

About the Project

Computer vision system for processing drone imagery captured inside greenhouses. Drones fly through the greenhouse photographing plants, but GPS coordinates are unreliable indoors. The core challenge was developing an algorithm to find common visual features across overlapping images and align them into a unified projection using linear algebra, despite having noisy or missing position data. The project also explored OpenDroneMap as a potential solution for stitching drone photos into orthomosaics.

• AgriTech
• Analytics

Responsibilities

Spent 6 months researching and developing the image alignment algorithm. Built the feature matching pipeline to identify common points between neighboring drone images. Implemented the geometric alignment system using linear algebra transformations to create consistent orthographic projections from misaligned source imagery. Evaluated and experimented with OpenDroneMap for comparison against the custom approach. Handled all geospatial processing using QGIS, Rasterio, and GeoPandas for the final output generation.

Q&A BOT

About the Project

An internal R&D project to build a question-answering bot for the company website. The system parses the corporate site to build a text corpus, then uses a two-stage retrieval pipeline to answer customer questions. The first model searches for the most relevant text fragments from the site content, and the second model generates natural language answers based on those retrieved passages. This retrieval-augmented approach ensures answers stay grounded in actual company information rather than hallucinating.

• AI
• Machine Learning

Responsibilities

Developed the full pipeline from web scraping through answer generation. Built the retrieval stage using PySerini for efficient passage search over the parsed text corpus. Implemented the answer generation model using PyTorch, fine-tuning it to produce accurate responses from retrieved context. Handled text preprocessing with Sentencepiece tokenization. Built the layer for integration with the website frontend. Managed the iterative improvement cycle based on answer quality evaluation.

IRRIGATION LINE BREAK DETECTION

About the Project

Detecting irrigation system failures using satellite radar data from ICEEYE SAR sensors. The project combined SAR data preprocessing and georeferencing with statistical analysis of tree health patterns during autumn and winter seasons. By correlating tree health anomalies with irrigation infrastructure, the system could identify locations where irrigation lines had broken or were underperforming, enabling targeted maintenance instead of manual field inspection across large agricultural areas.

• AgriTech
• Analytics

Responsibilities

Handled the full data pipeline from raw SAR data preprocessing through to predictive model output. Managed ICEEYE SAR data georeferencing using SNAP and QGIS. Conducted statistical analysis of seasonal tree health patterns to establish baseline expectations. Built the predictive model using scikit-learn to detect irrigation anomalies based on deviations from expected health patterns. Prepared geospatial analysis outputs using Rasterio, GeoPandas, and Shapely. Contributed to the published results of the research.

BEARING FAULT DETECTION

About the Project

A collaborative research project with KhPI (Kharkiv Polytechnic Institute) focused on predicting bearing failures from sensor data. The team assembled a physical experimental stand with bearings, collected vibration and performance data, and built predictive models to identify faults before they cause equipment failure. The results were published in a scientific article, contributing to the field of predictive maintenance.

• Manufacturing
• Scientific Research
• Analytics

Responsibilities

Participated in assembling the experimental data collection setup. Managed data collection, cleaning, and analysis of bearing sensor readings. Built the predictive model using scikit-learn, testing various classification approaches to distinguish healthy bearings from faulty ones based on vibration signatures. Conducted statistical analysis using SciPy and prepared visualizations with Matplotlib and Seaborn. Prepared data and results for scientific publication.