/ai
AI & Technology
I care as much about showing why a model's result can be trusted as about building the model. This page brings together my AI work, notes and writing.
Areas of work
Deep learning & computer vision
Medical image classification, transfer learning and segmentation (U-Net).
Natural language & RAG
Multilingual embeddings for Turkish text, article-level chunking and source-grounded answers.
AI security
Robustness of ML-based web application firewalls against LLM-generated attacks.
Automation & agents
Workflow automation with n8n; connecting AI to internal business processes.
AI projects
- 2026NeuroVision AIMRI-based brain tumour classification with five architecturesComputer Vision
- 2026Turkish Law RAGA question-answering system over Turkish legislation using article-level chunking.NLP · RAG
- 2026ML-WAF RobustnessLLM-generated attacks and adversarial trainingSecurity
- 2026Segmentation (planned)BRISC2025 · U-Net + EfficientNetB0 encoderComputer Vision
Methodology notes
Deterministic data preparation
All experiments share the same seed (SEED=42) and the same cleaned split, so comparisons stay fair.
Measuring leakage
Duplicate cleaning with SHA-256 + pHash, GroupShuffleSplit and leakage assertions in code.
Anatomy-aware augmentation
No horizontal flips on brain MRI; the symmetry assumption can destroy information.
Mistakes are findings too
The EarlyStopping callback carrying state was documented as a methodology note that affected results.
Technologies used
- Languages
- Python · Dart · C#
- Models
- ResNet50 · DenseNet121 · EfficientNetB0 · YOLOv8 · U-Net
- NLP
- sentence-transformers · paraphrase-multilingual-MiniLM-L12-v2
- Data
- ChromaDB · pHash · SHA-256
- Infrastructure
- Kaggle (Tesla P100 / T4) · Firebase
- Automation
- n8n