Applied AI Engineer | 13+ years across full-stack software and ML
I take LLM systems from prototype to production: multi-agent workflows, RAG, Document AI, fine-tuning, and the evals that prove they work. My edge is diagnosis: knowing when an LLM is the right tool, and saying so plainly when it isn't.
- 🔬 Fine-tuning legal reasoning models (Qwen + LoRA/QLoRA)
- 🏗️ Building agentic systems that turn messy PDFs and scans into structured data
- 🧪 Designing eval harnesses and benchmarks for LLM agents
- 📹 Explaining AI on YouTube and LinkedIn
Agents & RAG · LangGraph · LangChain · LlamaIndex · AutoGen · Qdrant · Pinecone · Weaviate · Document AI/OCR
Training & Eval · PyTorch · Hugging Face · Unsloth · PEFT/LoRA/QLoRA · LLM-as-judge · Benchmarking
Serving & Cloud · vLLM · Kubernetes · Docker · Airflow · MLflow · GCP · AWS · Azure
Backend & Web · FastAPI · Node.js · NestJS · React · Next.js · PostgreSQL
Languages · Python · TypeScript · Go · C#
Multi-Agent Workflow Automation
LangGraph and AutoGen agents with planning, tool use and self-refinement that automated data QA and report generation, saving 22 hours a week and cutting manual errors by 37%.
Production RAG
Hybrid lexical + dense retrieval with corrective RAG (LangChain, LlamaIndex). Cut hallucinations by 35% and lifted exact-match accuracy by 12 points on an internal gold set.
Legal AI Fine-tuning Pipeline · models on Hugging Face
Custom training framework for legal reasoning. LoRA + QLoRA for an 85% memory reduction and a 40% gain in retrieval precision.
LLM Training Data & Serving
SFT and RLHF data pipelines covering 2.5M+ prompt–response pairs across 12+ domains, with automated QC that cut training errors by ~40%. Fine-tuned models served with vLLM and FastAPI on Kubernetes at ~180 ms p95 and 300+ RPS.
LinkedIn · Hugging Face · YouTube · rzeraat.tur@gmail.com
Open to senior AI engineering roles, consulting through Elyoni, and interesting collaborations.




