Dr.-Ing. Pedram Azad
Machine learning and performance-critical software — from research to production.
For more than twenty years I have built software where research turns into product: visual perception for humanoid robots, industrial machine vision, speech processing for the car cabin, and today agentic AI. What connects them is working through a problem at every level instead of staying on the surface.
PhD from the Dillmann group at Universität Karlsruhe (TH) / KIT. Since then a company of my own, a head of R&D role, and currently principal software engineer and ML/AI team lead at Cinemo.
Work
-
Machine learning
Neural models for speech and perception on embedded targets: training pipelines, model representation, and a custom inference framework I wrote in C++ that takes models from TensorFlow and PyTorch through its own exchange format. More recently, the architecture and evaluation of agentic systems — software that uses tools on its own.
-
Performance-critical software
Architecture and implementation in C/C++. Optimization starts with the choice of data structure and algorithm, ends at the machine level — SIMD down to inline assembly, OpenMP, CUDA — and can shape the architecture itself. The aim is components that are powerful, robust and easy to use.
-
Computer vision
Computer vision in humanoid robotics and machine vision in industrial automation. Object detection, 6-DoF pose estimation, markerless human motion capture, camera calibration, 3D scanning and measurement. Computer vision is the subject of my dissertation, and the thread running through fifteen years of industry projects.
-
Speech and signal processing
Echo cancellation, noise reduction, voice activity and wake word detection for the car cabin — a domain built up from scratch taken through to pre-certification, including hands-free telephony under ITU-T P.1100, P.1110 and P.1120. Building up the most relevant certification tests in-house increased productivity substantially and was what built the domain knowledge in the first place.
Background
- since 2019
-
Cinemo, Karlsruhe
Senior software engineer, principal since 2024, and since 2026 also ML/AI team lead: together with a colleague I lead the AI Competence Center, a cross-departmental team of about ten. Agentic systems, architecture of an in-car voice assistant, and in-cabin speech processing (speech enhancement and wake word detection).
- 2014–2019
-
TECXIPIO, Karlsruhe
Senior researcher and developer, from 2017 head of research and development. Co-applicant and technical lead on two funded research projects: media identification and video super resolution.
- 2009–2022
-
Keyetech, Karlsruhe
Founder and managing director of my own company for industrial machine vision software, full-time until 2014 and alongside employment after that. Customer projects in printing, packaging machinery, special-purpose machinery and laser technology: 3D measurement, template-free print inspection, robot-guided grasping. Highly optimized libraries of my own for image processing and edge-based object recognition.
- 2008
-
PhD, University of Karlsruhe (TH)
Dissertation Visual Perception for Manipulation and Imitation in Humanoid Robots, published as a monograph by Springer.
- 2005–2011
-
University of Karlsruhe (TH) / KIT
Research associate at the Institute for Anthropomatics, Dillmann group. Visual perception for the humanoid robot ARMAR-III, European collaborative projects, teaching. From 2009 part-time as a postdoc, alongside my company.
- 2004
-
ATR Institute International, Kyoto
Research stay as a PhD student at the Computational Neuroscience Laboratories, Department of Humanoid Robotics and Computational Neuroscience.
- 2003
-
Diploma in computer science, University of Karlsruhe (TH)
Dipl.-Inform., with a diploma thesis on a projector-based 3D scanner for dynamic scenes.
- 2000–2004
-
RapidSolution Software, Karlsruhe
Systems developer, later freelance: network components in C++ and Java, object-oriented software architecture.
Teaching
During my years at the university I lectured and ran exercise classes, designed written exams and examined a large number of students orally.
Explaining taught me more about the subject than writing papers did. A method is only properly understood once it can be presented without scaffolding — as simple as possible, but no simpler.
-
Robotics III — Sensors and Perception in Robotics
Lecturer over several years, jointly with a colleague.
-
Cognitive Systems
Teaching assistant over several years, and responsible for the image processing part of the lecture.
-
Examining and supervision
Oral examinations, co-design of written exams, supervision of student theses.
-
Textbooks
First author of a book on computer vision, appeared in German and English, used as course literature.
Writing
My work on stereo-based object recognition, pose estimation and markerless human motion capture was carried out around the humanoid robot ARMAR-III and published at the robotics conferences, among them ICRA, IROS and Humanoids. 44 publications with more than 2,000 citations according to Semantic Scholar. The full list with BibTeX entries is on the publications page.
Software
Integrating Vision Toolkit
A platform-independent computer vision library in C++ with interfaces to Qt and OpenCV, which I wrote as sole author during my doctorate and released under the 3-clause BSD license. It carried my own research, served as the running practical example in two textbooks, and was used in a number of industry projects.
doIT Software Award 2006 of the MFG Foundation Baden-Württemberg, tenth place, jointly with Prof. R. Dillmann.