Cenk M. Yetis, PhD
Staff AI/ML Engineer · Edge AI & Efficient Models
I build compact neural models — audio, image and video codecs, and speech models — that run on-device within real memory, latency and power budgets.
Download CV · LinkedIn · GitHub
30 peer-reviewed papers, 1400+ citations · Google Scholar
My career has one thread: getting the most out of constrained systems. For 15+ years I designed wireless and signal-processing algorithms in industry and research labs, where bandwidth, latency and hardware limits were the whole problem. Today I apply that discipline to deep learning: knowledge distillation, quantization and pruning for edge devices, neural codecs, and hybrid edge–cloud systems that route between small and large language models. I'm at my best where model design meets deployment constraints.
3% PSNR loss under channel errors, against 21% for a learned-compression baseline, with an end-to-end learned transmission model. Related ideas appeared in a Huawei satellite image-sharing feature.
1/8 the compute of brute-force search, at equal performance, for deep-learning-based beam selection (Lund University / WASP).
Transformer-based neural audio codecs for low-bitrate transmission over unreliable links, plus edge–cloud SLM/LLM orchestration (Huawei).
Compact audio and speech models for edge AI using knowledge distillation and quantization (Analog Devices).
What I do
Edge AI & model efficiency: knowledge distillation, quantization, pruning, tinyML, TorchScript/LibTorch deployment, hardware–software co-design
Audio, speech & video: neural audio and video codecs, speech-to-speech, TTS-based synthetic speech data, keyword spotting, acoustic event detection
LLM & agentic systems: edge–cloud SLM/LLM orchestration, multimodal RAG, LangChain/LangGraph prototypes
Signal processing & wireless: 15+ years of PHY-layer algorithms (beamforming, channel estimation, semantic communication), PhD in Communication Systems