Robotics and AI engineer working on Physical AI: machines that sense the physical world and act in it. Co-founder and CTO of RoboRecs.
A benchmark forgives you. Hardware does not. Physical AI is the second kind of problem, and that single constraint shapes how I build.
→ Robustness beats benchmark scores. Two percent on a test set is not worth an unpredictable machine. Medical imaging taught me that one the hard way.
→ Most of the gap is measurement. Knowing what your sensor actually reports, and what your model silently assumes it reports. Unglamorous, and there is no way around it.
→ Embodiment changes the problem. An agent that can act has to reason about consequences it cannot undo. Once there is a body in the loop, perception and control stop being separate modules you can hand to separate teams.
→ The last mile is the work. Most of a project is spent getting from a paper that works to something that survives a Tuesday in a real building.
→ Design for the room you will actually be in. Cluttered, badly lit, partially observed, and different from yesterday. Building for the clean case and hardening later is how you end up with a demo.
Built to adapt, engineered to solve.
Get in touchThree products in production, across three different problems. RoboRecs and Buildly each replace something an expert does slowly by hand with something reliable and fast. Robotoplus works the other direction: it gets the physical robot itself out of the lab and onto a factory floor, then keeps it running after the sale.
Egocentric, multimodal human demonstration data for Physical AI
Humanoid robot foundation models need millions of hours of manipulation data. Roughly five thousand hours of open data exist. RoboRecs closes that gap with head- and wrist-mounted capture that records tasks from the operator's viewpoint, the same one the robot will have, across RGB, depth, IMU, audio, pose and force. EU-jurisdiction data, captured under GDPR and the AI Act.
Bills of quantities for the Bulgarian construction sector, in ten minutes
Construction cost estimation in Bulgaria still runs on spreadsheets and guesswork. Buildly generates a complete bill of quantities online from a library of more than 4,000 work items priced at real market rates, then carries it through to quotes and invoices. A different sector from robotics, and the same engineering problem: replacing an expert's slow manual process with something reliable and fast.
Humanoid and quadruped robots, sold, integrated and kept running across the Balkans
Most companies buying a humanoid robot are buying a research platform, not a production tool. Robotoplus selects, integrates and maintains AGIBOT humanoid and quadruped platforms for manufacturing, logistics and hospitality across Bulgaria, Romania, Greece and Cyprus, with ROS 2 engineering underneath for perception, navigation and pick-and-place, and one contract covering the hardware, the software and the support after the sale.
I am a robotics and AI engineer with an MSc from EPFL, focused on Physical AI. I build intelligent systems that perceive, reason and interact with the physical world. My experience spans robotics, computer vision, machine learning and real-world system deployment across medical imaging and autonomous systems.
Before that I spent a year at SamanTree Medical, on image quality for histological scanners. Building AI that a clinician will act on is a particular discipline: the model has to be right, and it has to be right in the same way every time, on a machine that has been sitting in a theatre since Monday.
I am currently co-founder and CTO of RoboRecs, building the egocentric human demonstration data that humanoid robot foundation models are starved of. What interests me is the gap between a research result and a system that still works on a deployment deadline, in a room nobody tidied first.
Day to day that means robot perception, computer vision and deep learning, mostly in Python and C++, and a lot of time spent on the measurement layer underneath all three.
Born and raised in Sofia, Bulgaria, with academic and practical experience across Switzerland, Mexico, Washington D.C., The Hague and China. I have taught mathematics, numerical analysis and robotics at EPFL, and ran robotics clubs for children for four consecutive years.
I founded the Association of Bulgarian Students in Lausanne (ABSL), the first organisation of its kind for Bulgarians in a Swiss academic environment, and served as its president for five consecutive terms, building a community of over 100 members from EPFL, UNIL and HEC Lausanne.
The work I want to be doing is the part between a result and a system: making something behave the same way on the tenth run as it did on the first.
Trained at institutions with international intake, strict standards, and work that had to actually run.
International robotics olympiads, a Mars rover competition, and a few things further afield.
Engineering competence applied to systems that have to work on real hardware.
Medical imaging, autonomous systems, and the transition from a research result to something that runs reliably.
Engineering solutions to real problems, from medical imaging through to autonomous robotics.
You learn as much about an engineer from what they do when nobody is grading it.
Always interested in connecting with engineers, founders and researchers building the next generation of Physical AI and intelligent robotics.