Furkan Küçük
Machine Learning and Artificial Intelligence Engineer/Researcher
Machine learning engineer and researcher with 5+ years taking computer vision, diffusion and agentic LLM systems from research into production.
Based in Lausanne, Switzerland. Open to ML research and applied AI roles.
About
I work on machine learning models and the systems that put them to use: computer vision, diffusion models, vision-language models and agentic LLM systems. Since 2020 I have taken them from research into production at DataBoss, Poliark, EPAM and ikas.
At ikas I directed AI integration across teams, shipped 6+ image generation and editing tools and built two agentic systems. Before that I worked with Zalando on a diffusion-based virtual try-on pipeline at EPAM, built LLM agents that work with graph models and an optimizer for architectural design as Head of AI at Poliark, and spent almost four years on real-time video analytics with TensorRT and DeepStream at DataBoss. I use LLMs as orchestrators that call deterministic tools, not as end-to-end solutions.
My MSc thesis at METU (GPA 3.93/4.00, completed September 2026) is KAVA: a video codec that fits an 81k-parameter adapter to each video on a frozen 14B video diffusion model, under review at ICLR 2027. I live in Lausanne and hold a Swiss work permit.
Experience
– 8 mos
Lead AI Solutions Architect
ikas · Türkiye
- Directed AI integration for products and internal workflows across multiple teams
- Designed and deployed 6+ commercial-grade computer vision tools for image generation/editing
- Built 2 agentic systems for documentation and chatbot automation
- AI Integration
- Computer Vision
- Agentic Systems
- MLOps
- LLMs
- Multi-Modal Models
– 6 mos
Lead Data Scientist
EPAM Systems · Türkiye
- Led EPAM's Computer Vision initiatives in Turkey
- Collaborated with Zalando on production-grade VTOn pipeline, advancing base generator and generative super-resolution models in a 3-component architecture
- Conducted multi-node training and deployment on Kubernetes clusters
- Computer Vision
- Diffusion Models
- Generative Super-Resolution
- VTOn
- MLOps
- Kubernetes
- Multi-Node Training
– 1 yr 2 mos
Head of Artificial Intelligence
Poliark Inc. · Türkiye
- Developed LLM-based agents and integrated graph models, diffusion models and MIQP optimizations
- Built custom ControlNet and LoRA pipelines for high-accuracy visual outputs in architectural design
- Engineered floor plan vectorization from CAD images via segmentation, graph processing and optimization
- Trained task-specific LLMs for architectural design and space planning
- Managed GPU/CPU deployment pipelines across AWS and GCP
- LLM Agents
- Computer Vision
- ControlNet
- LoRA
- Instance Segmentation
- Graph Processing
- MIQP
- AWS
- GCP
- MLOps
– 3 yrs 9 mos
Machine Learning Researcher in Computer Vision
DataBoss Security & Analytics · Türkiye
- Conducted research on video object detection, image enhancement, super-resolution, dehazing, and OCR
- Built distributed real-time CV deployment systems utilizing TensorRT and DeepStream
- Led MLOps infrastructure buildout with DVC, experiment tracking and orchestration
- Contributed to design, development and deployment of from-scratch distributed inference architecture; conducted multi-node training
- Improved object detection and super-resolution performance by 80% and 30% respectively
- Computer Vision
- TensorRT
- DeepStream
- MLOps
- Object Detection
- OCR
- Super Resolution
- Dehazing
- Multi-Node Training
- Distributed Inference
– 7 mos
Research Engineer
TOBB University of Economics and Technology · Türkiye
- Led "Collaborative Caching in 5G Networks" with HAVELSAN, resulting in patents in EU, US and Turkey
- Built LTE simulations using NS-3
- 5G Networks
- NS-3
- Research
- Patent Development
- Probability Theory
- Caching Algorithms
Skills
- Programming Languages
- Python
- C/C++
- CUDA (GPU)
- TypeScript
- Java
- MATLAB
- LaTeX
- ML Inference & Deployment
- TensorRT
- ONNX
- DeepStream
- Docker
- TensorBoard
- ML/AI Frameworks
- PyTorch
- Hugging Face Transformers
- PEFT
- Torchscript
- OpenCV
- MMCV
- Keras
- Scikit-Learn
- MLOps & Cloud
- AWS
- GCP
- MLFlow
- DVC
- W&B
- ClearML
- Git
- Unity
- Roboflow
- CVAT
- LLMs & Orchestration
- LLMs
- vLLM
- LangChain
- LangGraph
- Langsmith
- Langflow
- AI-SDK
- Mastra
- n8n
- Data Streaming
- Kafka
- Kafka Connect
- RedPanda
- gRPC
- RTSP
- InfluxDB
- QuestDB
- Grafana
- Data Science Libraries
- NumPy
- Pandas
- SciPy
- sci-opt
- Matplotlib
- Seaborn
- Specializations
- LLMs
- AI Agents
- Computer Vision
- Object Detection
- OCR
- Super Resolution
- Domain Adaptation
- Diffusion Models
- ControlNet
- LoRA
- Distributed HPC (training and inference)
- Languages
- Turkish Native
- English TOEFL iBT 98/120
- French Beginner, learning
Projects
3 projects
Eulerian Video Magnification in CUDA
Hand-written CUDA C++ implementation of MIT's Eulerian Video Magnification, up to ~1,650× faster than the NumPy reference implementation on an RTX 3090 (FP16, kernel time). Six backends (CUDA, Metal, Vulkan, OpenCL, PyTorch, NumPy).
- CUDA C++
- Metal
- Vulkan
- OpenCL
- PyTorch
- pybind11
- GPU Benchmarking
ResuVL
Three fine-tuned vision-language models, ResuVL-S, ResuVL-M and ResuVL-L (0.26B, 0.51B and 0.87B parameters), that read a CV page image and write its content as JSON, with no PDF text layer. On 204 held-out pages ResuVL-L scores 0.845 field F1 and reads a page in about a minute on a laptop CPU. Its SFT, DPO and GRPO experiments are documented in the repository.
Models on Hugging Face- Vision-language models
- Fine-tuning
- Knowledge distillation
- Document AI
- PyTorch
Video Decomposition Prior, re-implemented
PyTorch re-implementation of the ICLR 2024 paper "Video Decomposition Prior: Editing Videos Layer by Layer", with relighting, dehazing and segmentation results. Built with Alper Bahçekapılı for METU CENG 796 Deep Generative Models.
- PyTorch
- Video editing
- Generative models
- Paper re-implementation
Research
2026Paper
KAVA: Keyframe-Anchored Video Adapters for Extreme Video Compression
Under review at the International Conference on Learning Representations, ICLR 2027
F. Küçük, R. G. Cinbiş
2022Patent
A Use Frequency-based Cooperative Caching Method for Multi-layer Network Structures
Patent granted in Turkey, the United States and the European Union
F. Küçük, T. Girici, A. M. Yuksel, M. K. Cetin
TR WO2020142046A1 US US11343350B2 EU EP3906658NWA1
2019Paper
Collaborative Hierarchical Caching in ICN-Based 5G Networks
Third International Balkan Conference on Communications and Networking, BALKANCOM'19
F. Küçük, T. Girici, A. M. Yuksel, M. K. Cetin
2017Paper
Fair Beam Allocation in Millimeter-Wave Multiuser Transmission
WSA 2017; 21th International ITG Workshop on Smart Antennas, pp. 1–5, Berlin, Germany
F. Karababa, F. Küçük, T. Girici
2023Conference
CVPR '23 Vision-based InduStrial InspectiON
1st workshop on Vision-based InduStrial InspectiON, in collaboration with Apple
Vancouver, Canada · January 2023 – June 2023
Education
– 4 yrs 9 mos
Master of Science in Computer Engineering
Middle East Technical University · Türkiye
GPA: 3.93/4.00 (Summa Cum Laude)
- One-year preparatory program in undergraduate Computer Engineering (Operating Systems, Algorithms, Data Structures, Discrete Computational Structures)
- Research on advanced Computer Vision techniques under Asst. Prof. R. G. Cinbiş
- Focus on Domain Adaptation, Generative models, and compression
- From 11/2025: full-time thesis research on extreme video compression with frozen video-diffusion models (KAVA, under review at ICLR 2027)
- Computer Vision
- Domain Adaptation
- Video Compression
- Diffusion Models
– 4 yrs 9 mos
Bachelor of Engineering in Electrical and Electronics Engineering
TOBB ETU · Türkiye
GPA: 3.27/4.00
- Contributed to radar systems and 5G communication networks research
- Collaborated with Dr. T. Girici and Dr. A.M. Yüksel Turgut on caching algorithms
- Strong foundation in electrical engineering and telecommunications
- 5G Networks
- Radar Systems
- Telecommunications
- Caching Algorithms
- Probability Theory
He designed and built our generative AI solutions from the ground up … What I valued most was his ability to turn cutting-edge research into clean, production-ready systems that the rest of the engineering team could build on.
Furkan is a strong Data Scientist with deep and mature expertise in computer vision, consistently delivering high-quality technical solutions … he successfully helped manage and coordinate part of our Turkey-based team, demonstrating ownership, reliability, and natural leadership skills.
Get in touch
The quickest way to reach me is email.
Based in Lausanne, Switzerland. Open to ML research and applied AI roles.
Turkish nationality, Swiss work permit