projects

mostly the thesis these days:

active

refusal geometry in llms

msc thesis at sabancı: how refusal and harmfulness live in the internal geometry of llms. ask a model something harmful and it refuses. that refusal can be shown as a direction in activation space, and jailbreaks work by pushing the model off it. i am mapping what those attacks actually do to the representations. no public repo yet, we will see where it goes.

Mechanistic Interpretability
Refusal Directions
LLM Safety

earlier work

Neural Text Summarization: Comparative Analysis of Transformer Architectures

Comparative study of BART and T5 for automated news summarization, with sentiment-aware evaluation and ROUGE-based benchmarking.

Transformers
BART
T5
NLTK
Sentiment Analysis
ROUGE Evaluation

source

Multi-Label Classification System for Financial Recommendations

Multi-label recommendation system for İş Bankası, built with one-vs-all XGBoost, feature engineering, and ensembling.

XGBoost
Multi-Label Classification
Feature Engineering
Financial Analytics
Kaggle

source

Machine Learning Algorithm Implementations

A collection of practical machine learning implementations: boosting, ensemble techniques, and feature engineering workflows on real-world datasets.

XGBoost
Ensemble Methods
Feature Engineering
Data Science
Algorithm Implementation

source