Hi, I’m Arif Dağ

An AI-powered interactive software development platform using multi-agent LLM architecture with LangGraph orchestration. Features specialized agents (Planner, Test Generator, Coder, Executor, Reviewer) that collaborate through a self-correcting cyclic workflow. Supports Python, C#, and Java with automated unit test generation, UI testing, and intelligent code explanation. Built with RAG architecture and Docker sandboxing for secure code execution. (SAYZEK Academic Thesis Project)

A state-of-the-art deepfake detection system achieving 97% AUC by fusing remote Photoplethysmography (rPPG) physiological signals with EfficientNet-B2 visual features. Detects subtle heartbeat patterns disrupted in manipulated videos. Trained on Celeb-DF v2 dataset with a Flask web interface for real-time video analysis.

A Multi-Agent Reinforcement Learning (MARL) framework for autonomous UAV dogfighting with opponent intention modeling. Implements PPO-based training in gym-pybullet-drones simulator, featuring 1v1 and 3v3 combat scenarios with emergent tactical coordination. The intention prediction module enables proactive maneuvering against adversarial agents. (Research Internship at Fırat University)

Developed a novel Graph Neural Network framework for crowdsourcing platforms that maintains reliable truth inference even under coordinated adversarial attacks. Features heterogeneous graph modeling of worker-task relationships, attention-based malicious worker detection, and adversarial training for robustness against Collusion and Sybil attacks. (Research Internship at IOTECH)

A rigorous comparative study benchmarking U-Net++ (semantic segmentation, 96.12% Dice) against YOLOv8-Seg (instance segmentation, 92.77% mAP@50) on the BCCD dataset. Includes multi-seed evaluation, watershed-based instance derivation, and comprehensive analysis of speed vs. accuracy trade-offs in medical imaging applications.

A production-ready music recognition engine featuring asynchronous audio fingerprinting with SpectralMatch algorithm. Architecture includes FastAPI backend, Celery workers for background processing, dual-database system (PostgreSQL + MongoDB), and a Flutter mobile app. Handles partial/noisy audio clips with high accuracy.

Designed and implemented a novel blockchain-based storage architecture for metaverse digital assets on Hyperledger Fabric. Features smart contracts for asset ownership, provenance tracking, and secure peer-to-peer transactions with CouchDB state database for complex queries. (TÜBİTAK Research Project)

Capstone project for Yapay Zeka ve Teknoloji Akademisi, a narrative-driven co-op puzzle mystery developed by a team of 5 in one month using Agile methodology with Trello. Two players investigate a murder across different time periods: one explores the past while the other navigates the present, with actions in one timeline affecting the other. Features asymmetric puzzles requiring real-time communication.

A real-time multiplayer party game featuring a unique build-then-race gameplay loop. Players collaboratively place interactive items (cannons, bounce pads, hazards) to construct chaotic obstacle courses, then race to the finish. Built with Unity Netcode for GameObjects and Unity Gaming Services for seamless lobby matchmaking.

A multi-agent reinforcement learning environment where AI teams learn cooperative soccer strategies including passing, shooting, intercepting, and positioning. Agents develop emergent team behaviors through self-play training using Unity ML-Agents toolkit with custom reward shaping for teamwork.

An infinite procedural terrain system built with Unity featuring multi-layered Perlin noise generation, dynamic LOD mesh optimization, custom texture blending shaders based on height/slope, and GPU-instanced foliage placement for thousands of objects with minimal performance impact.

A deep learning pipeline for multi-label emotion classification from audio. Extracts Mel-spectrograms using Librosa, then processes through CNN architectures to predict emotional valence across multiple dimensions. Enables applications in mood-based music recommendation and affective computing.

A high-performance visual simulation creating mesmerizing, interactive geometric patterns using Unity's particle system and DOTS (Data-Oriented Technology Stack). Explores compute shader optimizations for rendering thousands of synchronized particles with complex mathematical transformations in real-time.

A polished top-down shooter featuring responsive twin-stick controls, wave-based enemy spawning with escalating difficulty, multiple weapon types, and satisfying visual/audio feedback. Built focusing on tight gameplay feel and enemy AI behavior patterns.

A puzzle-platformer developed in just 48 hours by a team of 4 for the Yapay Zeka ve Teknoloji Akademisi Game Jam. Players navigate parallel dimensions to solve interconnected puzzles. Demonstrates rapid prototyping, effective team collaboration under pressure, and delivering a polished experience within extreme time constraints.