Umut ABALI

I build data pipelines, agentic AI systems, and full-stack applications. Interested in LLMOps, RAG, and production ML.

Projects

Agentic RAG System

Multi-agent retrieval system using LangGraph with supervisor and execution agents. Hybrid search combining BM25 and dense vectors with RRF fusion and cross-encoder reranking on Qdrant. Full-stack setup with Next.js, FastAPI SSE streaming, JWT auth, Redis/Postgres session memory, Docker Compose, and vLLM for Qwen3-8B-AWQ inference.

LangGraphQdrantFastAPINext.jsDocker

Code

Multimodal Agentic Chatbot

Fully local, GPU-accelerated chatbot powered by Gemma-4 E4B (llama.cpp/GGUF) with a voice-enabled Chainlit streaming UI. LangGraph ReAct loop with CRAG web-search fallback and MCP server integration for GitHub, filesystem, and calendar management. Multimodal PDF ingestion for text and visual data.

LangGraphllama.cppChainlitMCPGemma

Code

Turkish Gemma-4 Fine-tune

Fine-tuned Gemma on a custom dataset of 3,000 Turkish instructions using Unsloth and LoRA. Trained on a single A100 GPU, completing 2 epochs in under 45 minutes with ~18M trainable parameters. Published to Hugging Face in LoRA adapter, merged 16-bit, and GGUF formats.

PyTorchUnslothLoRAHugging Face

Code Hugging Face

Mini-CDN

Prototype CDN in Go featuring distributed edge nodes, round-robin load balancing, and origin fallback. Caching layer with TTL expiration, LRU eviction, and manual cache purging APIs. Integrated Grafana metrics to track cache hit/miss ratios and monitor real-time content delivery performance.

GoGrafanaLRU CacheLoad Balancing

Experience

Data Engineering Intern

Mapin Data · Istanbul, Turkey

Jan 2026 — Mar 2026

Built a data pipeline to extract and structure unstructured information from web pages, PDFs, and QR codes using FireCrawl, OpenCV, and Visual Language Models (VLMs).
Reduced token costs by 70% through OpenRouter optimization and custom text-chunking algorithms, keeping accuracy high.
Worked alongside the core team to clean and map parsed data into a production schema, making it ready for downstream AI tasks.

Public Areas Attendant

Princess Cruises · Alaska, USA

Jun 2025 — Oct 2025

Participated in the Work & Travel USA program and worked in a fast-paced environment on a cruise ship.
Improved English communication skills by talking with international guests daily and worked effectively within a multicultural team to maintain high service standards.

Software Engineer Intern

Menatek Technologies · Istanbul, Turkey

Mar 2025 — May 2025

Designed and built a full-stack inventory management system from scratch using PHP and MySQL to digitalize and centralize asset tracking.
Architected relational database schemas with optimized indexing to ensure strict data integrity.
Implemented secure authentication and Role-Based Access Control to safeguard sensitive operational actions based on user roles.

Education

B.Sc. Computer Engineering (GPA: 3.20)

Istanbul Medeniyet University · Istanbul, Turkey

2022 — 2026

Technical Skills

Programming Languages

Python, C/C++, Bash/Shell

AI & ML

PyTorch, OpenCV, LangChain, LangGraph, vLLM, llama.cpp

LLMOps & Agents

RAG, Agentic Systems, LLM Orchestration, Vector DB (Qdrant), MCP

Cloud & Backend

AWS, GCP, Docker, FastAPI, Linux

Languages

Turkish, English (Professional Working)