Paper · PDF

A written treatment of neural networks — the artificial neuron and its weights, edges and activation thresholds, how layers compose into a network, and how that machinery is put to work training convolutional models. Accompanied by a notebook that fine-tunes CNNs on local image datasets, so every idea in the text has running code behind it.

Senior capstone · Technical report

Semantic segmentation of satellite imagery to identify land suitable for construction, and an algorithmic planner that turns the resulting mask into a city plan. The study compares U-Net against DeepLabV3+ across ResNet, VGG16, Xception and EfficientNet backbones. The data pipeline is the contribution as much as the models: PASTIS panoptic agricultural annotations were aggregated into a binary buildable / non-buildable mask, converted from multi-channel .tif to .npy and then .png, and resized to 160×160 so every backbone's pooling stages line up. Output is rendered as an explorable 3D web visualization rather than a flat mask.

Computer architecture · Report

One algorithm — converting decimal BCD values of up to eight digits into binary — implemented and measured three ways: portable C, IA32 assembly, and x86_64 assembly. The report walks through the conversion of a worked example (532.13022) at each level and compares what the three implementations cost. Written in Turkish.

Research notes · Code

Implementation notes for tracking moving objects across video frames with a Kalman filter, together with the dataset assembled for the experiments and the PyTorch setup used to run them on CPU and on CUDA.