đź’ˇ Projects
A summary of the projects I have worked on over the last few years.
IndEgo: Industrial Egocentric AI Research (2024-Present)
IndEgo is an internal research initiative spanning multiple federally and internationally funded projects. It explores how egocentric computer vision and AI can be applied in industrial environments to enhance human–machine collaboration. The goal is to enable digital and robotic assistants that understand tasks from a first-person perspective, learn from human demonstrations, and support workers in complex, real-world scenarios.
DAADEM (Starting in late 2025)
Digitalisierte, assistierte und automatisierte Demontage
Goal: Digitize and partially automate the disassembly of end-of-life vehicle parts (e.g., engine blocks) to support recycling, reuse, and sustainable manufacturing. DAADEM turns egocentric/robotic vision research into a real-world, funded application for sustainable, automated disassembly in industry.
Approach:
- Combine multimodal data capture (ego/exo cameras, depth, audio, robotics).
- Use AI for part recognition, localisation, and quality assessment.
- Apply AR-based worker guidance and robotic assistance for adaptive disassembly.
- Create a “data-supplying structure”- a platform to exchange standardised industrial and AI data between SMEs, OEMs, and startups.
Outcome: A demonstrator showing KI- and AR-guided disassembly of a motorblock, forming the basis for future robot-based automation and circular-economy integration.
Funded by the Fraunhofer Gesellschaft.
Steri.Bot (2025-2028)
Collaborative AI Robotics for Safer, Smarter Healthcare Workflows
Goal: Steri.Bot aims to reduce physical strain and infection exposure through an intelligent robot assistant that collaborates with humans via natural language and visual understanding.
The project lies at the intersection of egocentric AI, robotic manipulation, and multimodal learning, areas central to my ongoing research.
- Egocentric and Exocentric Data for Learning by Demonstration
- Uses *Project Aria smart glasses to capture *egocentric multimodal data (video, gaze, audio, motion) from skilled staff.
- Combines this with exocentric views from the Cir.Log camera system for multi-view imitation learning in grasping and process understanding.
- Vision–Language-Robotics Integration
A specialised domain-adapted language mode (Steri.GPT) and multimodal vision-language models (e.g. LLaVA, DINOv2) power the robot’s perception and reasoning:
- Understands and executes spoken commands (“SteriBot, bring me the basic tray from the top shelf”)
- Detects and explains manipulation errors
- Requests human help in ambiguous cases (“Please separate these instruments”)
- Multi-Agent Cognitive Architecture
- TU Berlin develops a retrieval-augmented multi-agent system (MAS) linking hospital information systems (KIS/IMS), robotic controllers, and NLP interfaces.
- Enables dynamic task allocation, context reasoning, and interactive dialogue between humans, robots, and digital infrastructure.
- Continual and Active Learning in the Real World
- The ML pipeline improves continuously via data-centric continual-learning pipelines, managing 100–200 TB of multimodal data.
- Data privacy and GDPR compliance are ensured through Fraunhofer-hosted secure cloud infrastructure and automated anonymization (face blurring, voice transcription).
Funded by the German Agency BMFTR.
CirculAIRe (Under Review, 2026)
Circular Value Creation in White Goods via LCA and AI-driven 9Rs
Summary: The traditional “take-make-dispose” model for household appliances is economically inefficient and environmentally unsustainable. Project CirculAIRe is a large-scale, international R&D initiative designed to solve this challenge by creating an intelligent, data-driven ecosystem that makes the circular economy for white goods profitable and scalable. While the project is broad, my core research focuses on developing the cutting-edge AI and computer vision technologies that serve as the “brain” and “eyes” of this new system.
KIKERP (2023-2026)
AI-based identification and classification of (used) electrical appliances for robotic process automation in circular economy-oriented digital management ecosystems.
Summary: An AI assistance system that evaluates refrigerators, washing machines and similar appliances that have reached the end of their useful life to determine the most effective next step: Are they suitable for repair and refurbishment, or do they need to be recycled instead? Incorporated into a cloud-based management platform, the system identifies household appliances based on images, helping to determine their quality and price so the used and refurbished products can be sold on e-commerce portals.
Funded by BMFTR and DLR Projektträger, under the KI4KMU program #16IS23055C.
More Information: Fraunhofer News – AI-Driven Lifecycle Management
MRO 2 (2023-2024)
Maintenance, Repair and Overhaul
Summary: In the project, new dynamic process chains are being developed in which the value creation steps are specified through digitisation. Each component is to pass through an individual repair chain. This development will be demonstrated using the example of gas turbine blades, which can be operated at higher temperatures or with longer operating intervals after repair. My contribution included developing an MLOps based framework for dataset curation and continual model training and deployment.
Funded by BAM and the EU European Regional Development Fund.
EIBA (2019-2022)
Sensory recording, automated identification and evaluation of old parts based on product data as well as information about previous deliveries.
Summary: The aim of the project was to develop an ML-based solution for identifying and assessing the condition of old parts. Using multimodal and multi-view AI methods, the system should be able to see and recognise products and compare them with other available information. I worked as an intern on the project and also did my Master Thesis on the topic of Green Incremental Learning, bringing together Green AI and Continual Learning.
This research has been funded by BMFTR in the framework of ReziProK (project ID 033R226).
I have also worked on industry projects, which are covered by NDAs, and cannot be shared in detail publicly. This includes SMEs and large organisations across Europe, spanning automotive, electronics, energy, and white goods sectors.