EgeÖzçelik

A comprehensive multi-agent AI platform carrying 20+ specialized pipelines for film & animation production. The architecture spans task-oriented agent crews under an LLM planner/executor, per-keyframe Vision QC mechanisms, and a model-agnostic dispatch layer running across 8+ generative engines.

Two AI-based travel & exploration apps, built and shipped: an AI-orchestrated, location-grounded discovery pipeline (schema-constrained LLM in edge functions) plus HealthKit-synced, Kalman-filtered route analytics.

End-to-end free-flow access & positioning system: a high-precision (±10 cm, 4–5 fixes/s) design on DS-TWR Time-of-Flight and PDoA azimuth. Within it: C firmware for the anchor hardware, native iOS tag integration on user devices, a PySide6 monitoring suite for operations, and a zero-hardware simulator that streamlines testing.

NPU-accelerated real-time recognition (Edge AI): an INT8 SSD/MobileNetV2 detector + FaceNet embedder (VX delegate) running fully on local hardware, cloud-free. Optimizing the models for the NPU yields ~20× the speed of standard CPU-based processing.

Stacked generalization & model optimization: a neural meta-head over a CatBoost base lifts F1 by 2.5 points (to 89.8%) and reaches 96% ROC-AUC. A high-performing ML model approved by an academic jury.

Comprehensive model evaluation under a single controlled protocol: leakage-protected feature engineering and a systematic analysis of 8 classifiers. RandomizedSearchCV tuning took CatBoost to 80.6% accuracy and 0.886 ROC-AUC, a clear performance lift.

Generator + monitor desktop pair speaking standards-correct 29-bit J1939 over SocketCAN (20 PGNs / ~30 SPNs); real industry traffic, with the two systems talking across Linux & Windows platforms.

Deep-learning image colorization: autonomous colorization of grayscale images as a/b-chrominance regression in CIE Lab. Autoencoder, U-Net and ResNet50-UNet architectures were built and compared, scored on PSNR / SSIM and validated with pairwise t-tests.