Data scientist

STARK

About the Project

An AI-powered guidance and tracking system for fixed-wing UAVs, integrated with a ground control station. The system enables real-time tracking and autonomous guidance toward operator-designated targets, both stationary and moving. It processes telemetry and sensor data to maintain stable target lock and accurate trajectory correction for small-scale objectives (roughly 2x2x2 meters) at flight speeds of 50-60 m/s. The solution needed to handle varying target dynamics and mission conditions reliably.

• Defense
• UAVs/drones

Responsibilities

Developed the core guidance and control algorithms for target interception, handling both stationary and moving targets under real-time flight constraints. Built a custom simulation plugin for the UAV environment to validate guidance behavior before field testing. Integrated the guidance logic with telemetry and ground control data flows. Ran extensive testing across both simulated and real-world flight conditions. Analyzed flight log data iteratively to refine control algorithms for better accuracy and stability. The project ran 15 months from initial development through field validation.

OMON

About the Project

Building a Visual Inertial Navigation System (VINS) for copter drones operating in GPS-denied environments. The system uses camera imagery combined with IMU sensor data to calculate position without satellite signals. In simulation, the system achieves 4m error on 1km missions and 250m error on 20km missions. In real flight, it can complete a 5km mission with under 500m maximum error. The project also includes a map system with working keypoint precomputation, georeferencing, and feature matching capabilities.

• Defense
• UAVs/drones
• Navigation

Responsibilities

Developed and customized the VINS algorithm based on OpenVINS, adapting it for the specific hardware and mission requirements. Integrated the navigation system with ArduPilot to enable real-time autonomous flight. Set up and maintained the simulation environment using AirSim with ArduCopter for controlled testing. Deployed the system on Raspberry Pi hardware and validated accuracy through extensive flight testing. Led technical decisions and mentored team members. Achieved 2-5% positional error relative to total flight distance in real-world conditions.

OPTICAL NAVIGATION WITH SATELLITE IMAGERY

About the Project

Developing a satellite-based navigation algorithm for UAVs in GPS-unstable areas. The system captures images from the drone's camera and matches them against pre-loaded satellite maps to determine the aircraft's position. This enables autonomous return-to-base and position holding without relying on GPS. The matching algorithm needed to handle the significant visual differences between live aerial photos and satellite imagery taken at different altitudes, seasons, and lighting conditions.

• UAVs/drones
• Defense
• Analytics
• Navigation

Responsibilities

Built the full image georeferencing pipeline, from satellite map preprocessing to real-time matching against drone camera feeds. Created a cross-domain dataset of paired satellite and aerial images for training the matching model. Developed and trained the matching algorithm to handle resolution differences between 60m altitude drone photos and satellite basemaps. Set up the simulation environment with AirSim and ArduCopter for testing. Planned the system development roadmap from research phase through to hardware deployment.

SEETREE: TREE HEALTH INDEX

About the Project

A satellite-based system for monitoring tree health across large agricultural areas. The platform processes satellite imagery to recognize individual trees and determine their health index, flagging trees that show signs of disease, drought stress, or decline (yellowing, drying, etc.). The health index provides farm managers with an early warning system for tree problems before they become visible during routine field inspections.

• AgriTech
• Analytics

Responsibilities

Built regression models for predicting tree health scores from satellite-derived features. Worked with large geospatial datasets stored in H5 format, processing Planet satellite imagery for feature extraction. Handled the full pipeline from raw satellite data through to health index predictions using GeoPandas, NumPy, and Shapely for geospatial operations, with Rasterio and OpenCV for image processing. Deployed the service with Docker.

VARIABLE RATE SPRAYING

About the Project

Trees in the same field can vary significantly in size, meaning each one needs different amounts of water, fertilizer, and pesticide. This project aimed to determine the precise amount of material needed for irrigation and spraying in each zone of a farm, replacing blanket application with targeted, per-area dosing. The output feeds directly into variable-rate spraying equipment, reducing waste and chemical usage while maintaining crop health.

• AgriTech
• Analytics

Responsibilities

Researched and developed clustering models to group trees and field zones by size and treatment requirements. Built the analysis pipeline using GeoPandas and NumPy for geospatial data processing, with Scikit-learn for the clustering algorithms. Handled integration with Rasterio and OpenCV for processing the source imagery. Deployed the solution with Docker for production use.

DRONE ORTHOMOSAIC ALIGNMENT

About the Project

A patented approach to aligning drone orthomosaics captured at different times into a unified, georeferenced view. The technology enables temporal comparison of agricultural fields by accurately registering mosaics despite changes in vegetation, lighting, and camera positions between flights. This was a critical piece of SeeTree's data pipeline, and the solution automated 85% of what had previously been a manual alignment process.

• AgriTech
• Analytics

Responsibilities

Contributed to the research and development of the novel alignment algorithm. Built the orthomosaic construction pipeline using TensorFlow for learned feature matching and OpenCV for geometric transformations. Worked with Rasterio and GDAL for handling large geospatial raster datasets, and used Numba for performance-critical computation steps. Handled integration with the broader SeeTree data pipeline. The technology was ultimately patented.

SOAPY

About the Project

A platform that monitors handwashing quality using computer vision deployed at washing stations. The system scans hand movements in real time and provides immediate feedback on wash quality, tracking whether all required steps of proper hand hygiene are completed. The goal was to build a gesture recognition system capable of running on edge hardware (Raspberry Pi with Intel Neural Stick) to deliver real-time results without cloud dependency. The solution targets healthcare facilities and food processing environments where hand hygiene compliance is critical.

• Healthcare
• IoT

Responsibilities

Led the full project cycle over 10 months: gathered requirements, conducted research into gesture recognition approaches, developed the recognition models, and deployed to edge hardware. Built the pipeline using PyTorch for model training and OpenCV for real-time video processing. Optimized the models for Intel Neural Stick inference to meet real-time latency requirements on Raspberry Pi. Handled testing across different lighting conditions, hand sizes, and washing styles. Delivered regular progress reports to the client and iterated on model accuracy based on field testing feedback.

SMARTOMICA

About the Project

An NLP system for analyzing biomedical scientific literature. The platform provides AI-based tools for automatic text mining and information extraction from papers in PubMed and similar repositories. The core functionality is extracting structured relationships between biomedical entities (genes, proteins, diseases, compounds) from unstructured text, building a queryable knowledge base from published research. This helps researchers quickly survey relevant literature without manually reading thousands of papers.

• Healthcare
• Biotech
• Analytics
• Scientific Research

Responsibilities

Managed the data labeling pipeline for relation extraction training data. Built and deployed models for Named Entity Recognition (NER) and Relation Extraction using BERT and other transformer architectures. Worked with SpaCy for entity recognition, and TensorFlow for the relation classification models. Integrated NLTK for text preprocessing and built evaluation pipelines to measure extraction accuracy against manually annotated gold-standard datasets.

VIDEO ANALYTICS PLATFORM

About the Project

A web platform that provides real-time analytical data from connected video streams. The system is domain-agnostic, meaning it can be configured for different use cases like retail foot traffic, security monitoring, or industrial process oversight. The platform connects to existing camera infrastructure and runs detection and tracking models on the incoming feeds, surfacing analytics dashboards in real time.

• Retail
• Cyber-security
• Analytics

Responsibilities

Upgraded the tracking model to deep-sort for improved multi-object tracking accuracy. Implemented multi-camera tracking that maintains consistent object IDs across different camera views. Optimized both inference time and tracking performance to meet real-time requirements on live video streams. Worked with TensorFlow for the detection models and OpenCV for the video processing pipeline.

OIL AND GAS DNA CLASSIFICATION

About the Project

Exploratory data analysis and classification project working with DNA samples collected from oil basins. The goal was to determine whether machine learning algorithms could reliably classify "oil" vs "no oil" from high-dimensional biological data with a limited sample size. This is a niche problem at the intersection of bioinformatics and energy exploration, where traditional geological methods are expensive and time-consuming.

• Energy
• Analytics

Responsibilities

Cleaned and visualized the high-dimensional DNA dataset, identifying patterns and outliers. Explored multiple classification approaches to handle the challenge of limited samples with many features. Built a training and evaluation script for reproducible experiments. Tested XGBoost, Keras neural networks, and scikit-learn classifiers to find the best-performing approach for this data type.