Umut Onur Yaşar
AI/ML Engineer · Deep Learning & Computer Vision
I work at the intersection of model efficiency and real-world deployment — designing, training, and optimizing neural networks with a focus on speed-accuracy trade-offs.
My background in Electrical and Electronics Engineering gives me a low-level intuition that complements high-level ML research: from CUDA kernel optimization to transformer architecture design, I can reason across the full stack.
Current Work
Knowledge Distillation for RT-DETR (Stanford CS229 — 2025) Compressing transformer-based object detectors without sacrificing accuracy. Teacher: RT-DETR-L (32M params) · Student: RT-DETR-S (17M params). Running a 12-configuration ablation grid on COCO comparing Logit-KD (KL divergence) and Feature-KD (encoder L2 + decoder cosine similarity) strategies.
Research Interests
- Model Compression — knowledge distillation, pruning, quantization
- Object Detection — transformer-based architectures (RT-DETR, DETR variants)
- Efficient Inference — CUDA optimization, edge deployment
- Representation Learning — feature alignment, intermediate supervision
Selected Projects
| Project | Stack | Description |
|---|---|---|
| RT-DETR Knowledge Distillation | PyTorch · CUDA · COCO | Teacher-student framework for real-time object detection |
| Real-Time Object Detection | Python · YOLO · OpenCV | Real-time detection pipeline |
| Signal Processing Toolbox | C++ · Qt · FFT | Filtering, Fourier analysis, waveform visualization |
| RF Propagation & Coverage Tool | Python | RF signal simulation and visualization |