Projects - 2D/3D Human Generator Analysis VaE and GAN algorithms are investigated. Dalle, Stable diffusion and Midjourney performances are researched with several prompts. - Escape Game PC game has been developed in Unity. This is a single character 2D escape game implemented on C#. The character tries to escape monsters and ghosts of different powers and movement characteristics. - Animal Detection and Recognition System Animal detection and recognition system is developed. Animals and numerators are recognized by Yolo that is trained on a custom dataset. OpenCV is utilized for numerator region detection. Holistic - nested edge detection algorithm is used that works upon pretrained Caffe. PaddleOCR and easyOCR are applied for the number recognition. For perspective correction, several morphological operations have been applied. ESRGAN has been run to improve the image resolutions for higher OCR accuracy. - Speech Recognition System Nemo and WhisperX libraries have been used for ASR, diarization and transcription of videos. - Moreover, the identified speakers (people) are classified in terms of gender/age groups. Face Recognition System Eigenface, Local Binary Pattern (LBP) and Local Derivative Pattern (LDP) are applied, and their identification accuracies are compared. LBP has higher accuracy than LDP. We further apply the multi presentation. However, it does not further improve accuracy. Moreover, DeepFace library is used for face recognition as a SOTA technique. - Multi-Biometric Recognition Analysis Face and fingerprint recognition modalities, multi-modal, multi-presentation versions are implemented and compared in terms of equal error rate (EER). Face + Fingerprint with and without z-score normalization using product rule are not reducing EER. Multi-presentation methods outperform both individual and multi-modal methods. - Social media AR Project Social media filter project is developed in python where eye landmarks are detected, and a glass is placed on the face. - Animation Studio Website This portfolio website is optimized over WordPress with portfolio builder extension Semplice GPT Calculator Website This application is implemented in C# with MS SQL. The course list has been fetched from the database. A student can calculate GPT for her expected grades in the semester. - Restaurant Website Restaurant menu manager is implemented in PHP with MySQL. The chefs create a daily menu with meals and corresponding ingredients. The chef enters the portion for each meal and the system calculates the calories per portion. Customers will be able to monitor the menu, meals, and calories of each meal portion. - Career Website Job posting management is implemented in PHP with MySQL. Firms can create job postings with offers and expected qualifications. Job seekers can apply for the jobs and see their application status. Search and filtering mechanisms on job postings were built. -Satellite Integrated Content-Centric Cognitive Radio Networks, PhD Thesis Project We analyze content-centric wireless networks from energy and service capacity perspectives. First, we provide a compound spectrum sharing heterogeneous network model enriched with satellite networking, cognitive radio, and device-to-device paradigms. We propose a new content model and caching methods based on our new content model. We enhance our caching with cooperation, optimization. We analyze all caching verticals with respect to different system dynamics and compare our proposals to alternatives - Audio Watermarking via Beat Detection, BS Graduation Project In our application, we embed watermark in music file that can only be detected by authorized users. The message is spread in spectrum. The decoder reads the file asynchronously and tries to extract the hidden embedded message. For synchronization, beat detection is used. The audio watermarking is investigated in terms of the signal quality, bit error rate and robustness. This is implemented in MATLAB -Popularity-Aware Content Modeling and Caching A chunk popularity model is developed, and a ladder popularity prioritized cache replacement algorithm is proposed and investigated. It is shown that in-network caching improves energy efficiency.Moreover, our popularity prioritized caching algorithm is beneficial in terms of energy usage for HQ content dissemination scenarios - Energy-Aware Routing for Content Delivery in Software Defined Networks An application is developed that routes layered content considering the energy levels of route switches. It is implemented on Floodlight. Regarding the importance of base layer, two least energy consuming routes are selected. For the enhancement layer only the least energy consuming route is selected. -Handwrite Differentiation (Hand Sep) We differentiate between cursive and normal handwriting. We collect the mean magnitudefrom the gyroscope sensor of mobile devices. With this data set by the Wilcoxon test, we manage to achieve the differentiation. Moreover, we apply ANOVA test. - Sarsa Learning in a Grid World In this project a robot moves in a grid with the goal to reach a deterministic reward at a given point. Sarsa learning is implemented, Q (s, a) values on the grid and the optimal action for each state have been displayed. -SMS Spam Filter with Classifiers Multilayer Perceptron (MLP), Support Vector Machine (SVM), Recurrent Neural Network (RNN) are implemented to detect spam. MLP and RNN perform better than SVM in terms of accuracy, recall and precision. Moreover, we observe that MLP has better performance than RNN because RNN is more beneficial for problems that have temporal change. - Simple Calorie Calculation System (CALS) CALS is proposed to assess calorie expenditures. Accelerometer data by phones are used. Windowing, feature calculation and LDA are applied. Then, ML methods kNN, decision tree, MLP and SVM are implemented. - Sport Sense An android application is implemented to detect energy consumption of user during running, walking, or jogging. This app collects GPS, speed, acceleration data and sends them to the server periodically. We focus on the optimal interval for energy consumption of the device battery. With short period, there is a large power consumption. However, with long period, even if the consumption decreases, the data becomes old and irrelevant. - Importance Sampler and Die Markov Chain Analyzer First, the area of the unit p-norm ball is computed by importance sampling with a chosen proposal distribution. Then, for a given Markov Chain of a 6-sided die, initially the transition matrix A is constructed. The stationary distribution of A is computed by computing the eigenvalues and eigenvectors. The detailed balance condition is checked, the chain is analyzed for approximately reaching the stationary distribution and its convergence time is investigated.

2022 April-: Machine Learning Engineer at Electrosalus (in Boğaziçi University Teknopark), Full-time • Feature extraction over lung acoustic signals. • Machine learning (SVM, kNN, naive Bayes, decision tree types) and deep learning algorithms for the classification of lung acoustic. • Regression for air flow detection • Computer vision edge detection and segmentation algorithms over spectrogram images for the air flow detection. • Recurrent neural network (RNN) types (GRU, LSTM, bi-directional versions) for the phase detection over air flow spectrogram images. • Convolutional neural network (CNN) for the noise detection over air flow spectrogram images. • Feature extraction over heart signals DWT, MFCC • Decision tree classifiers (Random Forest, XGBoosting etc.) applied over heart sound signals. • Wavelet denoising and Hilbert envelope for S1 and S2 peak detection algorithms • S3 and S4 peak detection algorithm with multi-phase evaluation • Noise cancellation algorithm with multi-layer encoder decoder structure and bi-directional LSTM • Signals welch analysis • Transformer, Time Series Transformer algorithms for signal denoising • They are all implemented as a web service in python using data science libraries NumPy, matplotlib, SciPy, machine learning libraries scikit-learn, TensorFlow, Keras and PyTorch and computer vision library OpenCV. Also, Nvidia Cuda has been used for accelerated GPU training. • Back-end web development with Node.js and non-relational database management with MongoDB • Utilizing AWS S3, container-based technology Docker • Version control with git, GitLab, project management environment Jira 2023 May-2023 November: Data Science Specialist at Turna.com, Part-time, Project based • Smart Unconditional Ticket Cancellation Service has been studied. • Feature extraction and selection • Dimension reduction analysis (PCA, t-SNE) • Random forest, gradient boosting (XGB) and NN algorithms have been studied for service selection and service usage classifiers. • Imbalance data analysis (up/down sampling, stratified k-fold cross validation) • Investigation in terms of precision, recall due to imbalanced data. • Random forest and XGB regression for service price prediction • Investigation of time series analysis with NBEATS, NHITS, PatchTST, TIMESNET • MongoDB for database, scikit-learn and TensorFlow for algorithms 2023 March-2023 August: Senior Artificial Intelligence Engineer at Mindlee, Part-time • Question-Answering Model, Transformer-based AI chatbot (NLP), Generative diffusion model from text to image, Face detection and sentiment analysis over text, image • Back-end web development with Flask, • Utilizing version control system git, Notion for project tracking 2016-2021: PHD Graduate Research Student at BOĞAZİÇİ UNIVERSITY COMPUTER ENGINEERING • Scholar in a project on content-centric 5G networks, multimedia data management • Markov chain channel model for satellite integrated D2D enabled network. • Data service and multimedia caching algorithms • Cooperative techniques and dynamic programming approaches • Optimization for improved energy efficiency and acceptable QoS of a network • ML projects: o Spam filter where word-based tokenization and bag-of-words text modelling Natural Language Processing (NLP) techniques are utilized. Machine learning (MLP, SVM and RNN) algorithms are implemented and their accuracies are compared. o Robot movement algorithm with reinforcement technique Sarsa learning. o Calorie calculator with utilized windowing, feature calculation and linear discriminant analysis (LDA) for the preprocessing. The machine learning methods kNN, decision tree, multilayer perceptron (MLP) and support vector machine (SVM) were applied. • MATLAB, Java, python, and R 2015-2019: TEACHING ASSISTANT at BOĞAZİÇİ UNIVERSITY COMPUTER ENGINEERING Introduction to Computers, Introduction to Computing (JAVA), Principles of Programming Languages Introduction to Satellite Space Network (3 times), Satellite Networks 2011-2012: STUDENT ASSISTANT at BOĞAZİÇİ UNIVERSITY COMPUTER ENGINEERING