Research Work
Exploring the frontiers of AI — computer vision, multimodal models, and reinforcement learning.
Multi-Stage Vehicle Detection and Tracking using YOLOv8 and SORT Algorithm
Serhan Emre Telatar
Formulated a multi-stage detection pipeline using YOLOv8 for vehicle identification and a custom-trained model for license plate localisation, achieving 95%+ accuracy. Integrated the SORT algorithm to assign unique vehicle IDs, reducing tracking errors by 30% across high-density video frames. Developed a post-processing module using Linear Interpolation (SciPy) to reconstruct missing coordinates, maintaining 100% data continuity in occluded scenarios.
Deep Q-Network Agent for Complex Game Environments with Euclidean Reward Shaping
Serhan Emre Telatar
Built a hybrid game environment establishing 60FPS real-time communication between a browser simulation and an RL agent via WebSockets. Created a Deep Q-Network (DQN) agent with PyTorch, implementing Experience Replay to stabilize training across a 48-dimensional continuous state space. Implemented a novel reward function using Euclidean distance calculations that enhanced exploration, decreasing average training time per agent by 35% and improving stability across diverse levels.
End-to-End ML Pipeline for Credit Default Prediction on 100K+ Loan Applications
Serhan Emre Telatar
Constructed an end-to-end ML pipeline using Python and SQL on 100K+ loan applications to predict credit defaults with an AUC of 0.88. Optimized Random Forest and Gradient Boosting models via grid search, achieving 82% accuracy and identifying 5 key risk factors. Conducted statistical analysis on 50+ features, improving early-warning detection by 25% and leading to a 15% reduction in bad-debt provision at BNP Paribas TEB.