The graduation project topics offered by the faculty members of the Software Engineering Department are listed below:

NOTE: All our students are required to meet with a faculty member and select their graduation project in their field of interest until September 17, 2026.

Assoc. Prof. Ali NİZAM

• Detailed text analysis and map visualization of historical travelogues (Assigned): Historical travelogues locations can then be geocoded and visualized on interactive maps to reveal historical travel routes and geographic patterns.
• Creating secure SQL using Agentic AI (Assigned): Developing an agentic AI system that can translate natural-language queries into SQL while automatically checking the generated queries for security risks such as SQL injection, unauthorized data access, and unsafe operations. The system can use database schemas, validation rules, and execution feedback to iteratively improve and safely execute queries.
• Explainable AI on medical AI: Developing an agentic AI system that can translate natural-language queries into SQL while automatically checking the generated queries for security risks such as SQL injection, unauthorized data access, and unsafe operations. The system can use database schemas, validation rules, and execution feedback to iteratively improve and safely execute queries.

Assoc. Prof. Burcu Bektaş DEMİRCİ:

• Medical Image Analysis Using Mean Curvature Flow (Assigned): In this project, students will apply mean curvature flow, a geometric partial differential equation, to segment and denoise medical images by evolving contours toward anatomical boundaries.
• Applications of PINN (Physics-Informed Neural Networks) to Engineering Problems (Assigned): This project introduces Physics-Informed Neural Networks, which solve differential equations by building physical laws into the network's training.

 Asst. Prof. Melike Nur YEĞİN:

• Effects of Pruning and Quantization Techniques on Turkish Sentiment Analysis: This project investigates model compression techniques to reduce the memory footprint and computational cost of transformer models (such as BERTurk or LLMs) applied to Turkish sentiment analysis. By systematically evaluating how aggressive parameter removal and bit-precision reduction (e.g., INT8/INT4) affect classification accuracy, latency, and throughput across Turkish benchmark datasets, the study aims to identify optimal compression trade-offs for deploying high-performing Turkish sentiment models on edge devices and resourceconstrained environments.
• Evaluating Turkish Idiom & Proverb Understanding of Large Language Models: This project assesses how effectively state-of-the-art Large Language Models (LLMs) comprehend non-compositional and culturally rooted figurative language, specifically focusing on Turkish idioms (deyimler) and proverbs (atasözleri). By curating or leveraging high-quality evaluation datasets (derived from sources like the TDK dictionary), this study benchmarks open-source and proprietary models (such as GPT-4, Llama, and Turkish-adapted models) across tasks like idiom sense selection, contextual disambiguation (literal vs. figurative usage), and sentence paraphrasing.
• Evaluating Text-to-Image Generation Performance with ViTs: This project focuses on assessing how effectively Vision Transformers (ViTs) assess and measure the quality, relevance, and fidelity of images synthesized from textual prompts. Rather than relying solely on traditional evaluation metrics like Inception Score (IS) or Fréchet Inception Distance (FID)—which typically depend on older Convolutional Neural Networks (CNNs)—this project leverages ViTs to capture global spatial context and fine-grained visual relationships by breaking generated images into interconnected patches.
• Analysing Prompt-to-Prompt Image Editing with Cross Attention Control: This project explores Prompt-to-Prompt Image Editing, a framework that allows precise, localized modifications of AI-generated images simply by editing the text prompt while maintaining the original image's structure and layout. By injecting and overriding the cross-attention maps—which map specific words in the prompt to spatial regions in the image—during the reverse diffusion process, the project analyzes how controlling word-to-pixel bindings enables seamless object replacement, style transfer, and attribute scaling without altering unmentioned elements.
• Comparative Analysis of ViTs and CNNs in Image Classification: This project delivers a side-by-side performance evaluation of Vision Transformers (ViTs) and Convolutional Neural Networks (CNNs) across standard image classification tasks to identify their relative architectural advantages and constraints. By systematically benchmarking models (e.g., ResNet/EfficientNet vs. ViT/Swin Transformer) across varying dataset sizes, spatial resolutions, and levels of computational resource limits, the study analyzes key trade-offs—including data efficiency, training convergence speed, robustness against image perturbations/noise, and inference latency.

Asst. Prof. Zeki KUŞ:

• Automatic counting application: A system that automatically counts objects in a given image without distinguishing their class and marks them on the image. Students are expected to develop a mobile application as shown below.
  Example: https://countthings.com/en/
• Text-Based Clothing Search System: Students are expected to develop a clothing search system that uses only text-based search. Students are expected to tag clothing items with specific attributes. E.g.: “Category: Dress, Color: Red, Fabric: Cotton, Description: A comfortable, V-neck dress that is ideal for summer and doesn’t make you sweat.” These tags will be converted into feature vectors, and searches will then be performed on VectorDB based on similarity. In the second phase of the project, LLMs are expected to be incorporated to handle more ambiguous searches. Local LLMs can be used for tagging.
• Music Recommendation System: Students will collect music files in MP3 or WAV format from various categories. These files will be converted into embedding vectors using different libraries. A similarity-based search system will be designed using these vectors. At the end of the project, students are expected to create a web page capable of providing music recommendations.

Asst. Prof. Burçin DANACI:

•  Mobile Application for Physics 1 Laboratory: Within the scope of this project, a mobile application will be developed to be used in mechanics experiments, enabling students to measure and calculate physical quantities such as velocity, acceleration, and period.
• Optimization of Machine Learning Processes Using Tensor Networks: Tensor network algorithms, originally developed for modeling high-dimensional quantum systems, have recently been adopted in the field of machine learning. These methods enable the representation of high-dimensional data in a lower-dimensional space through local tensors and facilitate the optimization of machine learning processes.
• Numerical Simulation of Quantum Walks: Quantum walks, which offer advantages such as faster spreading and superposition compared to classical random walks, are particularly effective in search algorithms. In this project, the evolution of discrete-time quantum walks on different graph structures will be analyzed and simulated using existing quantum computing platforms (e.g., Qiskit).

Asst. Prof. Abdulkerim Mohammed YIBRE:

• Federated learning for health problems: One of the critical issues in the AI model is preserving privacy. This becomes serious when it comes to health data. Federated learning technique helps to train models from data located in different locations (it can be hospitals) without sharing private data. This kind of approach is vital in situations where there are strict rules in data sharing, for example, KVKK. Students can simulate federated learning using  TensorFlow Federated, PySyft, NVIDIA FLARE (NVFlare) and other tools.
• Optimization techniques in machine learning: Machine learning algorithms involve a number of hyperparameters, whose values have an impact on the performance of algorithms. Therefore, selecting or setting those values manually would be difficult to achieve the desired accurate results unless optimization algorithms are used together. Optimization techniques, such as heuristic/metaheuristic ones, are used as the best solution for a given problem. Students can apply machine learning and optimization algorithms for different types of problems in software engineering such as software effort estimation, project planning, log analysis and anomaly detection.
• Explainable AI: It is a relatively new approach in machine learning and AI in general. Traditional or 'black box' models have no capability of explaining why the model has reached a decision. Opposed to, black box AI models, Explainable AI is a technique in AI that helps to build trust, transparency and interpretability in the decision or predictions of AI models.

Asst. Prof. Cumali TÜRKMENOĞLU:

• LLM-assisted metaheuristic / many-objective algorithm design (A newly emerging niche in evolutionary computation):  "Can an LLM design a better NSGA-III variant for 5-objective vehicle routing?" — pick a many-objective problem, use an LLM to evolve operators, compare against standard methods, write up the results.
• GraphRAG and SLM-based NER (NAme Entity Recognizer): NER is the bottleneck of GraphRAG. The quality of the graph depends entirely on how well you extract entities. Bad NER = useless graph. Doing this with SLMs instead of LLMs makes the whole pipeline cheap enough for real deployment.
• XAI-augmented forecasting with foundation models (Time-series foundation models (TimesFM, Chronos, Moirai, Lag-Llama)): "Are zero-shot forecasts from Chronos trustworthy? A SHAP-based study on agricultural price prediction." You take Chronos, run it on a public crop-price dataset, apply SHAP, analyze which input regions drive predictions, and discuss whether the explanations are stable, consistent, and useful for decision-makers.