End-to-end language model development for Turkish digital marketing. Built a 200,000+ row synthetic training set spanning 12 domains, designed a JSONL pipeline, and added automated quality validation reaching 94% format compliance.
Compared four Qwen variants from 2B to 27B using LoRA and DPO. Diagnosed a token-budget gap at 27B that shaped the production layer decision. Wrote up the findings: 2,000+ views.
Audited the scoring pipeline, found 12+ metric inconsistencies that were suppressing measured performance by roughly 15 points, and redesigned it with standard metrics across 200+ questions.
Local secure MVP · Synthetic data
In AI chat systems, personal data users type into messages — national ID numbers, IBANs, phone numbers, addresses, emails and order references — can flow uncontrolled to model, log or browser layers. The security boundary is not just masking visible fields; it also includes where token mappings are stored, what the frontend can access, and whether high-risk actions require explicit human confirmation.
I built a local AI chat MVP with Turkish PII detection using
tr_core_news_trf and structural rules, two protection modes
(PLACEHOLDER and TOKENIZED), an encrypted PostgreSQL Token Vault, and a
second-confirmation mechanism for order cancellation and email simulations.
I verified the system not just through successful API responses but by
inspecting the DOM, console, storage layers, and browser network traffic
separately. No sensitive data leakage was observed on tested surfaces.
No real LLM, ERP or email service connected. Synthetic test data. No full KVKK compliance claimed. Local deployment only.
Contributed to setting up a secure, in-house LLM platform with Gemma and Qwen. Model serving on Ollama and Open WebUI, deployment experience on DGX Spark infrastructure.
Built an open-source framework that evaluates fine-tuned LLMs for Turkish digital marketing with multi-provider support. While benchmarking models against a 100-question test set, I also surfaced design flaws in the scoring system itself.
Audited the AI-powered strategy and media-mix modules of a digital marketing platform against a real test scenario; identified points where the model over-assumed and where KPI/budget checks were missing, and delivered recommendations that clarified product scope.
Designed the end-to-end architecture for a RAG assistant providing secure access to internal documents: chunking, embedding, permission-based access control, PII masking, audit logging and a model router, all as part of one comprehensive system.
Worked with 162,000+ rows of bank customer complaint data on text cleaning, fraud-detection rules, PII masking and sentiment/intent analysis; built classification with TF-IDF and classic ML models (Logistic Regression, Naive Bayes).
Built a 1,000,000-transaction dataset in SQLite and automated internal-control scenarios with SQL: data quality checks, fraud scanning, and turning manual review processes into automation.
Built a fully automated video production system made of 4 sub-workflows, from viral content research (Apify) to script planning (Claude) to video generation (Veo 3 / Seedance).
I manage weekly content production for Medium, LinkedIn and X with an automation system that requires human approval at every stage — a workflow that balances speed with control.
Built a prototype that helps technical founders quickly test an idea: AI-generated theses and 30/60/90-day roadmaps, built with Next.js and TypeScript.
Demo environment · Kaggle data · Canvas in draft
Processed food-delivery order data on Kaggle with Python and Pandas, exported it in a Braze-compatible CSV format and designed two lifecycle journeys: a first-week activation flow for new users and a win-back flow for inactive customers. Each segment has its own target audience, conversion event and Action Path logic.
Wrote an academic paper framing the Qwen fine-tuning process through real-options theory; used decision theory to frame which model size to invest in under a capital constraint. Currently under peer review.