AI Research → Edge AI → Industrial AI → Generative AI → Embedded Systems → HPC
I build research-driven AI and intelligent systems with an emphasis on measurable performance, robustness, explainability, deployment and engineering evidence.
My work spans three connected areas:
- AI Research: deepfake forensics, explainable AI, domain generalization and reliability-aware decision systems
- Intelligent Systems: edge AI, real-time computer vision, embedded systems and parallel computing
- Industrial AI: visual anomaly detection, generative design, model lifecycle governance and production-oriented ML workflows
I prefer projects that move beyond a notebook demo: problem definition → data contract → model design → evaluation → failure analysis → deployment → documentation.
| Track | Flagship Project | What it demonstrates |
|---|---|---|
| AI Research | DeepfakeULTRA | Multi-evidence deepfake forensics, XAI, domain shift and external evaluation |
| Edge AI | SİNAPTİC5G | Real-time perception, tracking, governed data pipelines and model-lock evidence |
| Industrial AI | WeaveVision | One-class anomaly detection, calibrated decisions, drift lifecycle and CI |
| Generative AI | Halı AI Carpet Design | Controlled SDXL + LoRA generation, provenance, retrieval and analytical validation |
| Embedded Systems | TinyOLED Desktop | Raspberry Pi, SSD1306, custom framebuffer architecture and browser simulation |
| HPC | MPI Parallel Matrix Multiplication | Distributed C/Python MPI benchmarking, scaling and efficiency analysis |
| Project | Engineering Focus | Evidence / Highlights |
|---|---|---|
| DeepfakeULTRA | Reliability-aware deepfake forensics with multi-evidence fusion, frequency analysis and XAI | 0.9820 internal ROC-AUC, 0.7405 mean external ROC-AUC across 5 datasets, explicit failure disclosure |
| SİNAPTİC5G | Real-time Edge AI and 5G road-safety perception | 15,487-image governed dataset pipeline, 9 canonical labels, model-lock/FTR acceptance evidence |
| WeaveVision | One-class visual anomaly detection for textile quality control | PatchCore / EfficientAD, PASS · REVIEW · FAIL · ABSTAIN, drift lifecycle, live CI |
| Halı AI Carpet Design | Controlled generative design with SDXL + LoRA | Provenance-aware workflow, retrieval, CIELAB / ΔE analysis, documented 52-test pilot snapshot, live CI |
| TinyOLED Desktop | Raspberry Pi + SSD1306 embedded desktop | Custom framebuffer engine, cooperative scheduler, 57+ apps and live browser simulator |
| MPI Parallel Matrix Multiplication | C / Python distributed computing benchmarks | MPI scaling experiments across P=1,2,4,8,16; Python reached 11.33× speedup at P=16 in the documented N=512 benchmark |
AI / ML
Python · PyTorch · Computer Vision · OpenCV · Transformers · XAI · Anomaly Detection · Generative AI
Systems / Deployment
CUDA · FastAPI · Streamlit · Gradio · Docker · C · MPI · Raspberry Pi
Engineering Practice
Benchmarking · Testing · Data Governance · Failure Analysis · Model Evaluation · Reproducibility · Technical Documentation
The contribution graph is useful as a consistency signal; the portfolio is optimized for auditable engineering evidence rather than commit volume alone.
- Reliable and explainable AI under domain shift and unseen conditions
- Edge AI and real-time perception systems
- Industrial computer vision with calibrated / abstaining decisions
- Generative AI systems with provenance, evaluation and controllability
- Research ideas that can be converted into testable engineering artifacts
Problem
↓
Assumptions & Data Contract
↓
Baseline
↓
Architecture / Algorithm
↓
Evaluation & Stress Tests
↓
Failure Analysis
↓
Deployment / Interface
↓
Evidence & Documentation
The goal is not only to make a model work, but to make the claim auditable.
- Flagship research: DeepfakeULTRA
- Real-time / Edge AI: SİNAPTİC5G
- Industrial vision: WeaveVision
- Generative design: Halı AI Carpet Design
- Embedded systems: TinyOLED Desktop
- Parallel computing: MPI Parallel Matrix Multiplication