I'm a Data & AI Engineer based in Trabzon, Turkey. I build production-grade LLM systems, RAG pipelines, and data infrastructure — with a focus on observability, security, and systems that are defensible at every layer.

• 🔭 Current Focus: LLM integration, RAG pipelines, on-premise AI infrastructure with DLP layers
• 🛠️ Core Stack: Python, FastAPI, Apache Airflow, dbt, Snowflake, PostgreSQL, Docker
• 🤖 AI/LLM: Anthropic API, LangChain, Qdrant, sentence-transformers, RAG, prompt engineering
• 📊 Observability: Grafana, Zabbix, OpenTelemetry — applied to both data pipelines and LLM systems
• 🎓 Education: M.Sc. Entrepreneurship & Innovation Management, Karadeniz Technical University (ongoing)

Production-grade Turkish RAG pipeline with automated quality scoring

• End-to-end pipeline: PDF/DOCX → chunker → Qdrant → LangChain → Claude → RAGAS → PostgreSQL → Grafana
• Every query automatically scored on Faithfulness and Answer Relevancy via RAGAS (reference-free, no ground truth needed)
• Benchmarked two chunking strategies on identical documents: fixed (faithfulness 0.33, ~3.8s) vs semantic (0.42, ~10s) — quantified trade-off instead of assuming
• Eval results persisted in PostgreSQL for longitudinal analysis: drift detection, model version comparison, Grafana alerts on quality degradation
• Qdrant chosen over Chroma for production-grade filtering, payload indexing, and horizontal scaling

Hybrid classification system: Rule Engine → TF-IDF → LLM fallback

• Three-layer architecture: rule engine handles known TCODEs at 100% confidence with zero API cost, TF-IDF covers familiar patterns offline, Claude Haiku fallback handles only ambiguous tickets — minimizing both latency and API spend
• Prompt engineered for deterministic JSON output with temperature=0.1
• Covers 10 SAP modules (FI/CO, MM, SD, HR, PP, PM, QM, Basis, Authorization, E-Solutions)
• Built from real experience managing 250+ SAP BW/4HANA process chains at enterprise scale

100% on-premise LLM usage with DLP layer

• Intercepts and masks sensitive data (PII, credit card info) before it leaves the local network
• KVKK/GDPR compliant, runs in isolated Docker environments
• Designed for enterprises that need LLM capabilities without cloud data exposure

End-to-end telemetry platform with decoupled microservice architecture

• Four-service Docker Compose stack: FastAPI ingestion → PostgreSQL (raw + analytics layers) → Python ETL worker → Next.js dashboard
• Decoupled ingestion from processing so API latency stays low under load while ETL scales independently (horizontal scaling)
• Idempotent ETL via DELETE → INSERT pattern — same time window can be reprocessed 100x with identical results; safe against retries and partial failures
• Solved service startup race conditions with Docker healthcheck + depends_on + in-service retry logic for self-healing resilience
• Resolved cross-container CORS/networking by separating server-side vs client-side request paths

AI / LLM

Data Engineering

Infrastructure & Observability

• 📜 Snowflake Data Engineering — Snowflake (2026)
• 📜 Apache Airflow 3 Fundamentals — Astronomer (2026)
• 📜 dbt Fundamentals — dbt Labs (2026)
• 📜 IBM Data Engineering Professional (v2) — IBM (2024)
• 📜 PostgreSQL for Everybody Specialization — University of Michigan