cv

General Information

Full Name: Serhat Tadik

Languages: English, Turkish

Frameworks & Tools: Python , PyTorch , PyG, Tensorflow, Streamlit, SQL , AWS , MATLAB, Julia, , C, C++, GIS

Publications: View Publications

Education

  • 2023-Present
    PhD in ECE Dept.
    Georgia Institute of Technology, USA
    • Developed a method to quantify and visualize what is known about spectrum occupancy in a region from spectrum monitoring data recorded at a few fixed locations. Extrapolated the average observed signal strengths at the monitors throughout the region through likelihood estimation of transmitter location(s) and a simple log-distance path loss model. Introduced new georeferenced spectrum occupancy visualizations that combine estimates of occupancy power with duty cycle, and of signal variation with confidence level offering insights into planning for future allocations, interference, and broadcast coverage analysis. Analyzed and interpreted the spectrum consumption trends over different times of day and seasons.
    • Took part in the implementation of an ambient scatter communications system. Worked on a complex- valued neural network demodulation technique in Python using PyTorch that outperforms conventional de- modulation techniques such as matched filtering.
    • Improving existing electromagnetic wave propagation models by applying regularized regression-based error correction to the models considering channel and propagation characteristics in Python (PyTorch) and MATLAB. Includes feature design & engineering and a multilayer regression model. The proposed model enables a 58% to 87% reduction in loss difference (error) variance.
    • Designing digital spectrum twins for tracking historical and current usage of radio spectrum and predicting future patterns and usage. Designing aggregation rules to combine received signal strength, variance, duty cycle, and confidence values for various radio users in a region of interest using georeferenced maps. Utilized GIS, OSM, Python, and MATLAB.
  • 2016-2022
    BsC in Electrical Engineering & Physics
    Bogazici University, Turkey
    • Submitted undergraduate thesis on "A Comparison Between GNN Architectures and Implementation on Brain Connectomes". Used Python/PyTorch and Julia for node classification on a citation network and graph classification on brain connectomes structured as graphs. Showed that graph isomorphism networks (GIN) outperform graph convolutional networks (GCN) in the graph classification task thanks to their isomorphic neighbor aggregation and graph readout operations.

Experience

  • 2024-Present
    Research and Development Intern
    Cognosos Inc., USA
    • Designed and developed an advanced data analytics dashboard leveraging Streamlit in Python, significantly enhancing data analysis efficiency through automation. This comprehensive tool provided deep insights into key factors, including outlier detection, accuracy scoring, RSSI distribution, and tag/zone diagnostic visualizations, enabling informed decision-making and operational improvements.
    • Working on indoor zone classification using various machine learning techniques including ANN, CNN and pre-ML techniques such as XGBoost. Analyzed the sources of misclassification, developed a hierarchical model combining an ANN with a One-vs-One classifier that helped improve accuracy by 1% overall. Brought to attention that the models in use not only provide a single prediction, rather multiple predictions with associated probabilities. Illustrated that the models' first two predictions provide an accuracy greater than 99% that has the potential to improve customer satisfaction. Trained and deployed the aforementioned models developed through PyTorch in AWS using Sagemaker and Lambda.
    • Suggested and implemented various voting algorithms including meta-model based voting, majority voting, and summation of probability estimates that combine multiple models' predictions for a more stable and accurate prediction. Showed that the accuracy can be improved by at least 0.5 % through voting.
    • Showed that combining samples at the training and inference stages improve the accuracy by up to 1.6% by eliminating noise and nulls in the time series distribution of signal strength.
    • Developed a radio-mapping diagnostics tool that helps detect outliers in the radio-mapping process. This, in turn, helps repeating radio-mapping of certain zones when necessary increasing reliability of the data used for training.
    • Responsible for many data collection campaigns that help analyze what kind of tools provide better accuracy and determine the optimal transmitter density to achieve the maximum accuracy.
  • 2021-2022
    Machine Learning Engineer
    Caretta Software, Turkey
    • The work included feature design & engineering, customer segmentation, and churn prediction analysis. Experimented with major clustering algorithms including KMeans, Gaussian mixtures, DBSCAN, and agglomerative clustering. Successfully clustered customers with small intra-cluster variance and large inter-cluster variance. Built insights from these clusters that helped analyze customer shopping behavior.
    • Utilized PyTorch for training an ANN classification model for churn prediction as well as SQL to manipulate the data. The promotions/nofitications specifically targeted for the predicted customers that had the potential to churn then resulted in the re-engagement of 26% of those customers.
    • Worked on customer product recommendation using graph neural networks. Experimented with major GNN algorithms including graph isomorphism network (GIN), graph convolutional network (GCN), GraphSAGE, and graph attention network (GAT). Using a GAT model that leverages customer-customer and product-product relations extracted from matrix decomposition along with the known transactions resulted in customer transaction prediction with an accuracy of 90%. This work has been accepted for publication in IEEE Signal Processing and Communications Applications Conference and is available online.

Open Source Projects

Honors and Awards

  • 2022
    • Graduated with high honors from the Electrical Engineering and Physics Departments at Boğaziçi University.
  • 2016
    • Ranked 36th among 2,000,000 candidates in the university entrance exam, LYS, in Turkey.

Professional Focus Areas

  • Digital Signal Processing / Machine Learning
    • Deep learning architectures and techniques for general prediction problems.
    • Wireless signal processing and analysis
  • Spectrum Sharing and Allocation
    • For a given bounded geographic region, frequency band, and time interval, optimal allocation of spectrum resources that maximizes spectrum usage and total utility, and minimizes the overall cost among various types of spectrum users.