About

I’m a Data Scientist at Wayve (autonomous driving) and Co-Founder & Full-Stack Developer at Fletza. Previously, I completed a PhD in Deep Learning at the University of Stuttgart, in cooperation with Robert Bosch GmbH, advised by Prof. Dr.-Eng. Bin Yang and co-supervised by Dr. rer. nat. Jens Mehnert.

My main research contribution is Mixture-of-Depths (MoD) for CNNs (ACCV 2024), achieving up to 25% CPU / 15% GPU inference speed-ups while maintaining accuracy across ImageNet, Cityscapes, and Pascal VOC.

For collaborations, technical discussions, or talks and workshops, reach out via email. My CV is available here.

Highlights

• Efficient CNN inference: MoD for CNNs (ACCV 2024), up to 25% CPU / 15% GPU speed-ups while maintaining accuracy
• Publications & IP: 5 publications and 5 patent applications (1 published)
• Applied ML: Wayve (autonomous driving), Bosch (industrial PhD), Quality Match (data quality)
• Talks & workshops: AI/LLM fundamentals, prompt engineering, and responsible AI for workplace settings (200+ participants)

Current work

Wayve - Data Scientist

• ML for autonomous driving with a focus on dataset quality and scalable data curation
• Developing evaluation workflows and quality signals to support training and validation at scale

Fletza - Co-Founder & Full-Stack Developer

• Platform that transforms lecture materials (PDF, Word, PPTX) into flashcards, quizzes, and exam questions
• SM-2 spaced repetition and exam simulations with automatic feedback
• Learn more at www.fletza.com

Selected publication

• Mixture-of-Depths (MoD) for CNNs - ACCV 2024 (ArXiv)
  Authors: Rinor Cakaj, Jens Mehnert, Bin Yang
  Contribution: Channel-selective inference for CNNs to reduce compute while preserving accuracy.
  Selected results:
  • ImageNet: ResNet75-MoD matches ResNet50 with 25% CPU / 15% GPU speed-up; ResNet86-MoD +0.45% accuracy with 6% CPU / 5% GPU speed-up
  • Cityscapes: FCN-ResNet86-MoD +0.95% mIoU at similar cost
  • Pascal VOC: Faster-RCNN-ResNet86-MoD +0.37% mAP (+0.4% AP50) with 10% CPU speed-up

Talks & Workshops

Topics I cover

• AI and LLM fundamentals for workplace settings (capabilities, limitations, risks)
• Prompt engineering in practice (structure, iteration, verification)
• Responsible AI use: data usage, monitoring, and governance implications
• Scenario-based exercises for critical evaluation of AI outputs

Previous experience

• Quality Match GmbH - Solution Architect: AI-driven data quality methods, Nano-Task Trees for annotation, LLM-based nano-task design
• Robert Bosch GmbH - Industrial PhD (Deep Learning): PhD in cooperation with Bosch; patents related to methods developed during the PhD
• KPMG - Working Student (CIO Consulting - Digital Strategy)
• Segura & Jesberger GmbH - Working Student (Family Office)

Background

• M.Sc. Mathematics (interdisciplinary Informatics), TU Darmstadt
• B.Sc. Mathematics with Economics (bilingual), TU Darmstadt

Personal projects (selected)

• Stock Index Trend Prediction: end-to-end pipeline in Python/PyTorch (data collection, validation, training, evaluation)
• Electronic Access Control System: Raspberry Pi + RFID-based access control with GUI for a local swimming club