👋 Hi, I'm Tuba.
I research why Turkish LLMs fail — across the data, model and evaluation layers. I don't just train models; I question where data loses quality, under what conditions model behaviour changes, and whether the metrics in use actually reflect real performance. I diagnose the problem, compare technical options, build the case for my chosen approach, and turn the outcome into measurable systems.
I run end-to-end Turkish LLM development: synthetic data engineering, LoRA/DPO fine-tuning, benchmark design and on-prem deployment. I diagnose the problem, researching why a model fails across the data, architecture and evaluation layers.
I diagnose why a model fails: data quality, architecture choice, or evaluation system? End-to-end Turkish LLM workflows: synthetic data engineering, LoRA/DPO fine-tuning, metric-based evaluation and on-prem deployment. Active experience with Hugging Face, Unsloth, Qwen and Gemma families.
See projects →I analyze where raw data breaks down. I build JSONL training sets and quality control pipelines with Python and Pandas; design dashboards with GA4 and Looker Studio that make current state visible. Analytics infrastructure is a prerequisite for meaningful model development.
See my projects →I automate repetitive workflows with n8n and build accessible, fast websites and tools with HTML/CSS/JS and Next.js. This portfolio site included — I prefer to build the tools that produce the output myself.
See examples →
The scoring system was returning misleading results but the root cause was unclear. I audited the pipeline, identified 12 metric inconsistencies suppressing measured performance by roughly 15 points, and rebuilt it with standardized metrics across a 200+ question benchmark.
Local secure MVP · Synthetic test data
I questioned whether masking personal data is actually enough to protect it. Built a layered system: Turkish-specific PII detection with spaCy + tr_core_news_trf, reversible tokenization into an AES-256 encrypted Vault, human-in-the-loop approval workflow, and a same-origin proxy so sensitive data never reaches the LLM layer in plain text.
No real LLM connected. Synthetic data only. Local deployment.
Turkish LLMs were scoring low in domain tests but the reason was unclear. I designed a 200,000+ row synthetic data pipeline across 12 domains with automated quality validation reaching 94% format compliance, ran LoRA/DPO fine-tuning experiments, and contributed to on-prem deployment.
At Mad Cat Labs I work on Turkish LLM development. On the MMX project we built a 200,000+ row training set from synthetic and real data, ran LoRA and DPO fine-tuning experiments across model families like Qwen and Gemma, designed evaluation pipelines, and took part in on-prem deployment.
I spent many years in digital analytics and performance marketing. That's where I learned how data gets read, transformed and turned into decisions. Now I'm carrying that foundation into AI: building automations with n8n, evaluating model outputs, and building tools that smooth out the development process.
I love sharing what I learn. Everyday content on Instagram (@tubac3l1k), longer technical writing on Medium. Both channels are in Turkish, because the gap in Turkish-language AI resources is huge and I want to help close it.
I produce Turkish-language AI and tech content, translating complex topics into plain, accessible language.
AI and tech content. Everyday posts in Turkish.
Technical writing on LLM development, fine-tuning and data engineering.
Have a project, or just want to say hi? Don't hesitate to write.
tubacelikk11@gmail.com