11/04/2026 updated


100 % available
Freelance Robotics Engineer | ML on edge hardware
Nicosia, Cyprus
Worldwide
B.Sc. Physics, TUMAbout me
I specialise in making machine learning run on real robots, on edge hardware, in real time. If your robot is reacting too slowly or missing what is right in front of it, email me what it does wrong and the hardware it runs on, and we can discuss what can be done about it.
AI AgentsArtificial IntelligenceComputer VisionAutonomous SystemC++ (Programming Language)Computational Science and EngineeringPython (Programming Language)Machine LearningRobotics Design and ProductionRobotic Automation SoftwareSimulationsRust (Programming Language)Robot Operating SystemMobile RobotsIndustrial Robot
Machine learning that runs on real robots, on edge hardware, in real time. That is most of what I do: perception, localization, and control code that lets a robot navigate somewhere hazardous, decide what it is looking at, and act on it, under real constraints on latency and power.
The hardest problems are usually at the edge case. The smallest object, the hardest frame, the one a weaker model misses, or the moment a robot moves too fast and loses what it was tracking. That is where most of my work actually happens.
I also build systems meant to run unsupervised. A pipeline is only useful if you can trust it to make the right call without you watching it, and that trust gets earned through engineering, not assumed.
For case studies, testimonials, and current offers: https://intellycode.dev
Experience
Professional experience0 - 5 years
Freelancer since2025
Languages
GermanFluentEnglishNative speakerGreekNative speakerRussianNative speaker
Project history
Built a cold outreach system trusted to act on its own, unsupervised. Maintains high level of personalization while running completely autonomously.
Got a neuromorphic gesture-recognition board driving a third-party robot. Built the hardware to supply the external processing unit, set up the networking and communication between the robot and the processing unit, and wrote the ROS layer turning recognized gestures into robot actions. Control ran on a timing-based approach and lidar-based obstacle avoidance kept it from colliding.
Built a detector and tracker for an early-stage deep-tech company's target object, so they could rely on the system precisely where it is hardest to trust: spotting the smallest, most distant instances, the ones a weaker model misses first. The new model beat the published baseline there by 16%, and the classifier I added kept that performance while telling the target apart from three classes of lookalikes, plus motion prediction so the tracker keeps a live estimate running between observations.