open to ML / AI roles · Berlin

Irmak
Köseoğlu.

>

M.Sc. Big Data & AI student building RAG systems, deep learning models and the pipelines that ship them. I like AI that works in real products, not only in notebooks.

irmak@berlin: ~
irmak@berlin:~$
// 01 · about

Math brain,
engineer hands.

I studied Mathematics Engineering at ITU, and now I'm doing my M.Sc. in Big Data & Artificial Intelligence at SRH Berlin, focusing on Generative AI and Machine Learning.

I build things end to end: from RAG pipelines and fine-tuned vision models to the FastAPI services, Docker images and CI/CD that keep them running. Before Berlin, I worked as an analyst at KPMG Turkey, explaining tech results to both engineers and non-technical people.

And yes — I still love the math under the hood.

TR nativeEN fluentDE B1
0
ML projects shipped
0
plant diseases classified
0
ML pipelines benchmarked
0
news articles classified
// 02 · projects

Things I've built.

From retrieval pipelines to deployed vision models — each one tested, containerized and shipped.

rag_pdf_qa/● LIVE
RAG

PDF Question Answering API

retrieval-augmented generation pipeline
  • Ask questions about any PDF: chunking → embeddings → ChromaDB retrieval → LLM answer.
  • Packaged as a FastAPI service with automated tests and a CI/CD pipeline; running in production.
PythonFastAPILangChainChromaDBDockerGitHub Actions
llm_leakage_bench/research
96

LLM Data Leakage Detection

benchmarking & evaluation
  • Can an LLM spot data leakage in ML pipelines? Compared prompting strategies against a rule-based baseline on 96 pipelines.
  • Chain-of-thought won — and was the only method that caught a hard-to-find leakage type.
PythonOpenAI APIscikit-learn
plant_disease_cv/vision
38

Plant Disease Classification

CNN vs. transfer learning
  • ResNet50 with transfer learning classifies 38 plant diseases from leaf images — clearly beating a CNN trained from scratch.
  • Grad-CAM to check the model looks at the right part of the leaf before trusting it.
PyTorchResNet50Grad-CAMFastAPIDocker
plant_inference_api/mlops
API

Plant Disease Inference API

MLOps deployment
  • Turned the model into a live REST API returning top predictions with confidence scores.
  • CI/CD runs tests and builds the Docker image on every push — broken code never reaches deployment.
FastAPIDockerGitHub ActionspytestPyTorch
fake_news_clf/nlp
72k

Fake News Detection

text classification
  • Classifies 72,000 news articles as real or fake.
  • Feature engineering + GridSearchCV to pick the best-performing model.
Pythonscikit-learnNLP
// 03 · skills

$ ls ~/toolbox

What I reach for, grouped by where it lives in the stack.

drwx llm_genai/
LangChainRAGFAISSChromaDBPrompt Engineering
drwx machine_learning/
PyTorchscikit-learnTensorFlow / KerasHugging Face
drwx cloud_devops/
DockerKubernetesAWSAzureGitHub ActionsFastAPI
drwx languages/
PythonC++SQLR
drwx data_viz/
PandasNumPyMatplotlibJupyter
// 04 · journey

git log --graph

Istanbul → Berlin, math → machine learning.

Oct 2025 — Mar 2027 (expected)

M.Sc. Big Data & Artificial Intelligence

SRH University of Applied Sciences · Berlin

Focus on Generative AI, deep learning and deploying ML systems.

Oct 2024 — Jun 2025

Analyst, Digital Finance & Fintech

KPMG Turkey · Istanbul

Supported client projects on data and technology with cross-functional teams, reporting results to technical and non-technical stakeholders.

Sep 2019 — Jun 2024

B.Sc. Mathematics Engineering

Istanbul Technical University (ITU)

Where the love for optimization, probability and proofs began.

// 05 · writing

Notes & essays.

I write about the math behind AI.

Medium · Jun 2025 · 5 min read · game theory

The Minimax Theorem: Optimization Meets Game Theory

How the minimax theorem helps AI make safe decisions under uncertainty and competition — from saddle points and mixed strategies to poker bots and fraud detection.

↗
// 06 · play

Beat my AI agent.

A tiny search-based AI agent. You're X, the AI is O. Before every move it simulates every possible future of the game and picks the safest path. Spoiler: the best you can do is a draw.

your move.
you 0draws 0AI 0
positions evaluated: —
// 07 · contact

Let's build
something useful.

irmakkoseoglu01@gmail.com