AI & Computer Vision engineer. I train deep-learning vision models and put them into production on edge hardware — currently vehicle detection, traffic scene analysis and licence plate recognition inside an intelligent transportation platform, plus the LLM assistant that lets operators ask it questions in plain language.

Ankara, Turkey · LinkedIn · goktugg.gumuss@gmail.com

llm-traffic-assistant — a self-hosted RAG and tool-calling service. Ten JSON-schema tools whose schemas are generated from their function signatures, multi-turn sessions, SSE streaming, and a grounding checker that traces every number in an answer back to a tool result before returning it. Scored by a held-out evaluation set that includes questions it must refuse.

ocr-reads-or-guesses — does a vision-language model read the pixels, or rewrite them into what it expected? Shown KAVBAK, a guesser returns KAVŞAK: a well-formed, confident, invisible label error. 36,000 readings in Turkish say prior-pull scales with the model — primed with its domain, the 7B model rewrites clean, undegraded words 4.5% of the time, and scores 0.1% on licence plates because it returns the letter block and drops the digits. The CTC baseline never repaired once. Readings on Kaggle, with a notebook that rebuilds every table from them.

traffic-vision-pipeline — detect, track and count vehicles, with the frame budget measured stage by stage. Benchmarked on an RTX 5090: at batch 1 the pipeline is bound by fixed per-call overhead rather than by the model, and batching 16 frames is worth 1.6–5.7×. The counting logic is verified against scenes whose answer is known before the pipeline runs.

night-vehicle-detection — a controlled ablation asking whether darkened daytime images can stand in for real night data. They cannot: manufacturing night bought nothing (−1.3%, inside noise) while real night frames bought 28.6% at night, almost all of it recall. Weights on the Hub as G2mus/night-vehicle-yolov8s.

cat-irt-engine — a Rasch adaptive testing engine. Reaches the accuracy of a 45-item fixed exam in 27 items, and states the information bound that makes anything shorter impossible. No dependencies.

adaptive-huffman-delta — lossless compression for 16-bit sensor streams on microcontrollers. Delta coding in front of adaptive Huffman, header-only C++17 with no heap, no recursion and caller-owned buffers. 73% saved on slow sensor signals, and a worst case on incompressible input bounded at 2% growth rather than left to chance.

autonomous-vtol-uav — onboard detection, tracking and follow control for a VTOL UAV. First place at the 16th R&D Project Market.

Python C# C++ · PyTorch TensorFlow YOLOv8 Detectron2 OpenCV · llama.cpp Qdrant RAG function calling · FastAPI Flask .NET · Jetson STM32 ROS PX4 · PostgreSQL Docker Git

_{Every repository above documents what it measured, and what it did not. If a number appears in a README here, there is a script in the same repository that reproduces it.}