I'm a software engineering at Karabük University (graduating 2026) and founder of Visuality, an AI-focused initiative.
My current work sits at the intersection of large language models, vision-language architectures, and real-world deployment. I build things that run on consumer hardware as well as enterprise GPUs — from local GGUF quantizations to H100-trained fine-tunes.
- 🔭 Currently fine-tuning Qwen3.5-9B on localized Turkish instruction datasets
- 🧠 Researching Vision Language Models (VLMs) and multimodal architectures
- 🛠️ Built custom-clip-vit-b-coco, a Vision Transformer that scales from RTX 3050 Ti to A100/H100
- ⚙️ Advocate for workflow automation with n8n, agentic systems, and efficient local inference
A supervised fine-tune of Qwen3.5-9B for Turkish instruction following, reasoning, and NLG — trained on 500K samples from InstrucTurca.
Benchmark results (lm-evaluation-harness v0.4.2, 0-shot, H100):
| Task | Score |
|---|---|
| MMLU (English) | 0.7787 |
| HellaSwag | 0.7834 |
| ARC-Challenge | 0.5375 |
| Belebele (TR) | 0.8144 |
| Turkish MMLU | 0.6555 |
| XCOPA (TR) | 0.6780 |
No catastrophic forgetting — English reasoning is fully preserved after Turkish fine-tuning.
| Repo | Link |
|---|---|
| Full model (BF16) | MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM |
| Local inference (GGUF) | MuhammedKsee/Qwen3.5-9B-Instruct-Turca-TurkishLLM-GGUF |
AI / ML
Languages
Tools & Infra
Open to collaboration on LLM fine-tuning, VLM research, and local AI deployment.