This thesis presents CBC-Distilled, an efficient multimodal CNN–Transformer model for remote sensing semantic segmentation under missing-modality conditions. The model combines RGB, near-infrared (NIR), and shortwave infrared (SWIR) information and uses feature distillation, cross-band correlation, and random modality masking to improve robustness when one or more modalities are unavailable. The teacher and auxiliary branches are used only during training and removed during inference to reduce computational cost. Experiments show that the proposed model maintains competitive segmentation performance while requiring fewer parameters and lower computational cost than the baseline.

Date: 03.09.2026 / 13:30 Place: A-212

Continuous production environments (refineries, power grids) generate vast multivariate sensor data, yet remain stuck in fragmented “Narrow AI” while foundation models transformed language and vision. This thesis introduces Prometheus, a compact 1.6M-parameter bidirectional Transformer encoder that reads a plant’s sensors as a language and learns its coupled physics through self-supervised “Four-Teacher” geometric masking. On a crude distillation unit, Prometheus wins all 38 channels on every metric, surpasses zero-shot Time-Series Foundation Models up to ~300× larger, reconstructs entirely missing sensors, and beats a deployed industrial soft sensor, evidence that domain specialization, not generalized scale, is the path toward Industrial Foundation Models.

Date: 25.06.2026 / 14:30 Place: A-212

This thesis presents a conversational analytic system that enables users to query complex data environments through natural language instead of writing SQL. The system combines Data Mesh and Data Fabric principles on a lakehouse architecture and uses LLM-based agents to discover datasets, enrich metadata, and infer relationships between tables. The goal is to reduce the manual effort required for data discovery and schema exploration, while keeping the underlying metadata transparent, reusable, and reproducible. The approach is evaluated through a multi-domain benchmark designed to measure how relationship metadata affects query correctness and reproducibility.

Date: 15.06.2026 /15:00 Place: A-212

This thesis presents a conversational analytic system that enables users to query complex data environments through natural language instead of writing SQL. The system combines Data Mesh and Data Fabric principles on a lakehouse architecture and uses LLM-based agents to discover datasets, enrich metadata, and infer relationships between tables. The goal is to reduce the manual effort required for data discovery and schema exploration, while keeping the underlying metadata transparent, reusable, and reproducible. The approach is evaluated through a multi-domain benchmark designed to measure how relationship metadata affects query correctness and reproducibility.

Date: 18.06.2026 /15:00 Place: A-212

This study identifies and investigates the critical barriers to adopting Agile methodologies within Turkey’s defense industry. Utilizing an extensive literature review and expert consultations, the research identifies key challenges in transitioning to flexible and collaborative methodologies. By applying Interpretive Structural Modeling (ISM) and MICMAC approaches to questionnaire data, the study maps the causal relationships and interdependencies among these barriers, establishing a layered, hierarchical framework. Ultimately, this research provides defense industry managers with a systematic understanding and a comprehensive framework to enhance awareness, mitigate adoption impediments, and lay a solid groundwork for successful Agile transformation.

Date: 17.06.2026 Place: A-212

This study takes a quantitative look at how protein-coding genes replicate across different cell lines using SigProfilerTopography, where we score earlier-replicating genes higher. We originally set out to build a model that could predict replication timing directly, but when that didn't pan out as expected, we shifted our focus to mapping out the actual timing differences between cell lines.

Date: 23.06.2026 / 14:30 Place: A-212

This thesis proposes a self-supervised framework for detecting Primary (P) waves in strong motion accelerograms, an essential task for Earthquake Early Warning systems. Using Variational Autoencoders trained exclusively on P-wave segments sourced from 648 recordings of the Turkish National Strong Motion Network, the model detects P-waves through reconstruction-error-based detection without requiring labeled data. A systematic search across 492 configurations of four VAE architectures reveals that attention mechanisms achieve the best detection performance. The final Attention-VAE model achieves an AUC of 0.998, surpassing supervised baselines and demonstrating potential for real-time deployment.

Date: 12.06.2026 / 14:00 Place: A-212

This thesis develops a regime-aware day-ahead electricity consumption forecasting framework for Türkiye. It aims to predict the next day’s 24-hour national consumption profile using historical load, weather, and calendar variables from 2019–2025. Special attention is given to weekends, public holidays, Ramadan, and Eid periods, since these days create different demand patterns. Several statistical, machine learning, deep learning, hybrid, and meta-learning models are compared. The best result is achieved by a Ridge and LightGBM-based meta-learner, reaching 1.63% MAPE in 2025.

Date: 04.06.2026 / 10:00 Place: A-212

This thesis proposes an attention-enhanced Variational Autoencoder for generating station-specific strong ground motion records. The model encodes three-component PEER NGA-West2 seismic acceleration waveforms as six-channel STFT spectrograms and learns compact latent representations through a convolutional encoder with an attention-based bottleneck. A station-aware latent sampling strategy produces site-specific synthetic recordings from limited per-station data. A structured evaluation framework is introduced in the scope of the thesis. This framework combines time-domain metrics and pseudo-spectral acceleration analysis through intensity-shape binning. Generated records are benchmarked with the evaluation framework against original recordings and SCEC Broadband Platform simulations across Southern California stations.

Date: 30.04.2026 / 14:30 Place: A-212

This thesis presents a systematic evaluation of Static Application Security Testing (SAST) tools. Related studies mostly use synthetic codebases and per-vulnerability evaluation methods. In this study, both synthetic benchmarks and real-world intentionally vulnerable applications are tested against tools, along with per-issue evaluation. Conducted experiments measure various metrics and explain these results with reasons behind them. In addition to quantitative results, qualitative features and internal mechanisms of tools are examined to further explain results and observed performance differences. The results demonstrate the difference between evaluation models and tool effectivenes. Overall, thesis offers practical insights for SAST tool research and selection.

Date: 14.04.2026 / 11:00 Place: Cisco Lab