Hi, I'm Doruk Özgen.

Research Paper · Vision-Language Models

A review of Google's PaliGemma 2 paper that asks whether model size or image resolution pays off more. Covers the SigLIP and Gemma 2 architecture and the staged training, then compares both ways of scaling over 42 benchmarks: going from 224 to 448 px lifts scores by 8.4% on average at 4.6× the compute, going from 3B to 10B parameters by 4.5% at 3.7×.

PaliGemma 2, Multimodal AI, Transformers, VLM

AI Agent Command Center

A cross-platform product featuring a Rust/Tauri desktop client, web platform, and management dashboard. Built for multi-agent orchestration, interactive session streaming, telemetry monitoring, and fine-grained permission controls.

Rust, Tauri, TypeScript, Multi-Agent, Desktop

Neural Networks & Language Models from Scratch

Implemented neural network fundamentals from scratch: automatic differentiation, backpropagation, MLPs, character-level language models, and WaveNet. Developed and trained GPT-style Transformer models using PyTorch, exploring attention mechanisms, weight initialization, and GPU-accelerated training.

PyTorch, Transformers, Deep Learning, LLMs, From Scratch