AI Research Scientists in Germany

Project: Defect Generation on Test-Bench Images of Metal Surfaces Environment: Automated Visual Inspection (AVI), Metallurgy & Manufacturing

• Objective & Implementation: Designed, architected, and trained Generative Adversarial Networks (Pix2PixHD / SPADE) for image-to-image transformation. Targeted generation of synthetic material defects (e.g., cracks, inclusions, scale) on rough metal surfaces under real test-bench lighting conditions for privacy-compliant and efficient dataset expansion (data augmentation).
• Technical Design: Implemented robust Generative AI and computer vision pipelines in Python and PyTorch. Used semantic segmentation approaches for mask-controlled defect synthesis and subsequent evaluation with EfficientDet object detection models.
• Business Impact: Massive dataset upscaling (10x) without time-consuming and costly physical test-bench runs, while significantly improving the detection performance of automated inspection systems.

Technologies & Skills Used: Python | PyTorch | SPADE | Pix2PixHD | EfficientDet | Machine Learning | Semantic Segmentation | Computer Vision

Conventional Paradigm Test (CPT) – AI Evaluation & LLM Research

Development of an experimental evaluation approach to examine “paradigmatic closure” in Large Language Models — that is, the question of how far LLMs can recognize the basic assumptions, values, and limits of the paradigms within which they generate answers.

Design and testing of an additional approach to classic AI benchmarks that does not primarily measure factual correctness or task performance, but instead examines a model’s ability to recognize alternative perspectives, make implicit assumptions visible, and reflect on the limits of its own answer or interpretation framework.

Focus areas: development of evaluation criteria and test questions · LLM evaluation and comparative model analysis · prompt and response analysis · qualitative classification of model answers · study of epistemic compression and value leakage · benchmark and literature research · development of structured assessment and analysis methods

As part of CPT, existing AI evaluation approaches and benchmarks were analyzed, and a minimalist test protocol was developed that classifies model answers by response patterns such as DIRECT, CLARIFY, PLURALIST, REFUSE, and META-AWARE. TruthfulQA was used as the basis for experimental application and comparison with existing reference answers.

Technologies & Methods: Large Language Models (LLMs) · Generative AI · Prompt Engineering · AI Evaluation · TruthfulQA · Benchmark Analysis · Human-in-the-Loop Evaluation · Qualitative Content Analysis · Research & Literature Review

Self-directed exploration of agentic AI development methods, taken from idea to a working, publicly usable application

• Built an interactive planning tool for modelling retirement withdrawals and tax strategy using an agentic AI (vibe coding) development approach – demonstrating self-directed investigation of new AI-assisted development methods
• Delivered live, tax-aware spending projections and adjustable user inputs; shipped as a free, install-free browser application built in Python, with attention to usability for non-technical users

• Research and develop models for chatbots, NLP and LLMs (e.g. Llama, Qwen, OpenAI)
• Enhance chatbots with RAG, in-context learning
• Supervised fine-tuning (PEFT, LoRA), Huggingface or Unsloth
• Advanced training methods: Test-time training, (transductive) active learning, reinforcement learning
• High-throughput serving with vLLM
• Apply embedding models (e.g. SentenceTransformers), similarity/vector search or vector DB or ranking (e.g. LlamaIndex, Faiss, LangChain)
• Generate and filter synthetic data, clustering
• Detect hallucinations
• Evaluate chatbot models (Rouge, BLEU, F1-Score, Recall, Precision)
• Visualization of experiments (matplotlib)

• Designed an automated model evaluation pipeline enabling LLMs to inspect each other for alignment issues
• Implemented a RAG system with iterative cross-examination for reliable results
• Automated generation of written summaries and hypotheses to support rapid iteration and hypothesis testing

• Designed and implemented a real-time perception pipeline using YOLOv7 on Time-of-Flight (ToF) sensor data, enabling live streaming, inference, and on-frame visualization for passenger detection.
• Fine-tuned and evaluated multiple state-of-the-art monocular depth estimation models for Automatic Passenger Counting (APC), and developed a custom hybrid depth model that improved depth accuracy in challenging scene regions.
• Demonstrated that model-generated depth maps outperform raw sensor depth for APC tasks across several datasets, contributing to measurable reductions in counting error.

• Train and optimize MAVAE/VAE models in PyTorch to detect anomalies in multivariate time-series sensor data.
• Design preprocessing workflows and evaluation pipelines to improve model accuracy and robustness.
• Benchmark MAVAE performance against baseline statistical and deep learning approaches, presenting comparative insights.
• Collaborate with research supervisors to refine hypotheses and translate experimental findings into deployable research outputs.

• Led a team of junior Research Engineers, providing mentorship, technical guidance, and career development support to foster their growth in deep learning and machine learning engineering.

• Technologies: GoLang, JavaScript, TypeScript, AWS, Terraform, Git
• Conception of AWS infrastructure and existing services
• Analysis of IAM accounts and roles
• Setup of Cost Explorer and CloudWatch monitoring
• Setup of DynamoDB and S3 persistence of collected information
• Reporting and cost calculation
• Conception of Terraform deployment

• Contributed to the Karakutu project, developing AI-driven tools to analyze news in Turkey.
• Assisted in web scraping, applied NER for entity extraction, and built interactive filtering interfaces (Vue.js, Plotly.js) for entity and location based search.
• Performed sentiment and content-shift analysis to detect editorial influence in modified news articles.