Currently an Artificial Intelligence Software Engineer at Baykar Technologies , I specialize in developing high-efficiency computer vision systems and automating the lifecycle of AI models through robust DevOps/MLOps practices.
| Category | Proficiency |
|---|---|
| Languages | C++23 (Factory/Decorator Patterns), Python, SQL, C#, Git, OOP |
| MLOps & DevOps | Docker, Nginx, CI/CD, AWS, Azure, Embedded Linux |
| Inference & Streaming | TensorRT, FFmpeg API (Sub-10ms latency), GStreamer, ZeroMQ |
| Deep Learning | YOLOv12, Mask R-CNN, Detectron2, Transformers, VLMs |
| GenAI & LLM | Qwen-3 (32B), RAG, Chain-of-Thought (CoT) Prompting, LangChain |
An intelligent CI/CD integrated agent that understands the semantic impact of code changes beyond simple diffs.
- Smart Triage: Classifies changes (SKIP, AUTO_APPROVE, CRITICAL) to optimize review costs and focus.
- Semantic Analysis: Detects ripple effects in code logic that standard linters miss.
- Policy Enforcement: Automatically applies custom review policies and SAST security gates using LLM-based reasoning.
An enterprise-level integration platform that automates the deployment of AI models for production labeling environments.
- Automation: Auto-detects new model implementations and generates Nuclio deployment files for zero-config scaling.
- Efficiency: Implements Lazy Initialization and Smart Resource Management, unloading idle models to optimize GPU memory.
- Orchestration: Manages multiple model types (Detectors, Trackers, Interactors) through a unified FastAPI interface.
- Designed and developed a reusable, scalable multimedia API in
$C++$ using the FFmpeg API. - Leveraged the Producer-Consumer Design Pattern to achieve sub-10ms real-time video streaming latency.
- Currently powering 4+ production projects.
- YOLOv12 Optimization: Reached SOTA precision by increasing mAP50 score from 98% to 99% via HMM post-processing.
-
Inference Speed: Achieved a massive 72% FPS boost on YOLOv8 and 17.5% on Faster R-CNN using TensorRT and
$C++23$ . - Segmentation Excellence: Improved Detectron2's Mask R-CNN mAP by 33% using a multi-channel fusion approach.
- LLM Reasoning: Automated internal software analysis by applying Few-shot and CoT prompting on the Qwen-3 32B model.
- Video SOTA: Enhanced MaskFreeVIS with an optical-flow fusion block, achieving a 4.4% mAP gain on video sequences.
- 📧 Email: [email protected]
- 💼 LinkedIn: yusufcicek
Building the future of autonomous systems, one optimized kernel at a time.