Peter Schillen-Senior AI, Machine Learning & Computer Vision Expert

Project: Defect generation on inspection images of metal surfaces

• Environment:* Automated Visual Inspection (AVI), Metallurgy & Manufacturing

• Goal & implementation: Concept, architecture, and training of 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 inspection-light conditions for privacy-compliant and efficient dataset expansion (Data Augmentation).

• Technical design: Implementation of robust Generative AI and Computer Vision pipelines in Python and PyTorch. Use of semantic segmentation approaches for mask-guided defect synthesis and downstream evaluation with EfficientDet object detection models.

• Business impact: Massive dataset upscaling (factor of 10x) without time- and cost-intensive physical inspection runs, while at the same time drastically improving the detection performance of automated inspection systems.

Technologies & skills used: Python | PyTorch | SPADE | Pix2PixHD | EfficientDet | Machine Learning | Semantic Segmentation | Computer Vision

Project: Synthetic data for surface defects

• Environment:* Industrial Vision, Surface Quality Inspection

• Goal & implementation: Development of a conditioned generation pipeline (via ControlNet and latent-based inpainting methods) for the targeted synthesis of variable surface defects (e.g. scratches, pores, coating defects) on flawless product photos with exact mathematical control over position, size, and defect class.

• Technical design: Concept and fine-tuning of generative AI pipeline architectures (Stable Diffusion Inpainting, ControlNet) in PyTorch and Python. Image processing with OpenCV, mask generation and evaluation with Mask R-CNN / Semantic Segmentation. Development of APIs and tooling GUIs with Qt and cloud scaling on Amazon Web Services (AWS).

• Business impact: Savings of an estimated 80% of physical scrap accumulation for data collection; drastic increase in detection accuracy (Recall) for rare, critical defect classes in industrial quality assurance (Quality Control Engineering).

Technologies & skills used: Python | PyTorch | Stable Diffusion Inpainting | ControlNet | Mask-RCNN | OpenCV | Semantic Segmentation | Machine Learning | Amazon Web Services (AWS) | Qt (Software) | API Development | Quality Control Engineer

Project: Generation of synthetic defects on weld seams

• Environment:* Non-Destructive Testing (NDT), Heavy Industry

• Goal & implementation: Development of a physics- and edge-guided generation pipeline for the synthesis of realistic weld seam defects (e.g. lack of fusion, pores, crater cracks) on optical images. Automatic balancing and targeted multiplication of the training dataset for downstream segmentation and detection models.

• Technical design: Training and orchestration of generative deep learning architectures (GANs, Latent Diffusion Models, Transformers) with PyTorch and Python. Image processing and feature extraction with OpenCV, as well as UI development in Qt.

• Business impact: Complete balancing of the previously heavily distorted dataset (Data Imbalance Solved) and significant improvement in the detection accuracy of downstream AI models in industrial quality inspection.

Technologies & skills used: Python | PyTorch | GANs | Latent Diffusion Models | Transformers | Computer Vision | OpenCV | Data Science | Qt (Software)

Project: Web-based Training Platform for Generative AI

• Environment:* Cloud & SaaS

• Goal & Implementation: Concept and development of a scalable web framework for fully automated management of complex datasets and running distributed fine-tuning jobs in the cloud.

• Technical Design: Backend and fullstack architecture based on Python and Django. Integration of PyTorch for model training and orchestration of high-performance cloud infrastructure on Amazon Web Services (AWS SageMaker, Elastic Beanstalk, EC2, S3).

• Business Impact: Significant reduction in setup and training times when adapting generative models through a modular user and project management system.

Technologies & Skills Used: Python | PyTorch | Django | AWS (SageMaker, Elastic Beanstalk, EC2, S3) | SaaS Architecture | Data Engineering

Project: Simulation of a Vision Solution for Robot-Based Sewing & Fabric Processing

• Environment:* Industrial Automation & Robotics

• Goal & Implementation: Development of a physically accurate fabric simulation for highly complex deformable materials using mathematical optimization and systems simulation. Implementation of a parallel on-demand cloud rendering pipeline on AWS for scalable synthetic data generation.

• Technical Design: Automated control and scripting of simulations using Python and the Blender API as part of a flexible system architecture.

• Business Impact: Enabled efficient training of robust vision models for robots without complex manual test setups or costly physical test series.

Technologies & Skills Used: Python | Blender (API) | Amazon Web Services (AWS) | Systems Simulation | Mathematical Optimization | Software Development (general)

Technical Leadership & Strategy

• Technical Lead: Technical leadership of an engineering company with a strategic focus on innovative generative AI solutions and synthetic data generation.
• Client Projects: End-to-end design, development, and scalable deployment of industrial computer vision and AI systems for enterprise clients.

Generative AI & Data Architecture

• Generative AI Tools: Architecture and scaling of proprietary generative AI tools for the automated creation of synthetic training data.
• Efficiency & Impact: Significant reduction in data annotation costs and faster model training cycles through the targeted use of synthetic datasets.

Machine Learning, Deep Learning & Innovation

• Deep Learning Innovations: Development of new deep learning methods (CNNs, Transformer) for precise information extraction from highly complex laser images as well as for trinamiX material recognition and skin detection using PyTorch and TensorFlow.
• Architecture Optimization: Continuous evaluation and performance improvement of neural network architectures for demanding computer vision and classification tasks.
• Technical Leadership: Technical contact person and mentor for all machine learning and deep learning topics across the entire company.

Data Engineering, Pipelines & Spectroscopy

• End-to-End Data Pipelines: Design and implementation of complete data preparation and modeling pipelines - including hardware specification for data collection, automated data processing, metric evaluation, and systematic testing.
• Spectroscopy & Chemometrics: Analysis of NIR spectra and development of chemometric algorithms for accurate material and spectral data analysis.

Embedded Systems, Prototyping & Cloud

• Embedded & Edge Deployment: Responsibility for the software development of the first spectrometer prototype and implementation of resource-efficient algorithms (C++, Python) for use on embedded systems and mobile devices (e.g. smartphones).
• Cloud Integration: Connection, scaling, and management of data science pipelines in the AWS cloud.

Project: Development of Mobile FaceAuth & Material Classification App

• Context:* Fintech, Security & Biometrics

• Goal & Implementation: Porting and hardware-close optimization of highly complex deep learning and computer vision models for face recognition, object detection, and material classification on mobile devices (Android).

• Technical Design: High-performance development in C++ with direct integration of OpenGL ES for real-time graphics and image processing on mobile graphics chips. Optimization of inference times for strict edge and latency requirements.

• Business Impact: Enabled highly secure, real-time payment and authentication processes directly on edge devices without cloud latency and while meeting the highest security standards.

Technologies & Skills Used: C++ | Computer Vision | Object Detection | OpenGL ES | Android | Edge AI & Biometrics

Project: NIR spectrometer development (PbS cells) & mobile integration

• Environment:* Spectroscopy & Mobile Devices

• Goal & implementation: Hardware-close control and signal processing for infrared sensors (PbS cells). Development of robust chemometric evaluation algorithms for NIR spectra, as well as seamless connection to a mobile Android app and a scalable database backend.

• Technical design: High-performance signal processing and data transformations in C++ and Python. Architecture of the interfaces and backend for efficient transfer and storage of spectral data on mobile devices.

• Business impact: Successful commercialization of a mobile, handheld spectrometer for material identification in seconds directly on site.

Technologies & skills used: Python | C++ | Data transformation | Database backend | Android | Embedded Systems | Signal processing

Algorithm development, image processing & computer vision

• 3D reconstruction & computer vision: Development of physically motivated algorithms and pipelines based on C++, Python, and OpenCV for precise reconstruction of 3D point clouds from projected laser points.
• Machine learning & AI: Implementation of models for object detection and semantic segmentation using TensorFlow.
• Software development & GUI: Creation of high-performance software solutions, interfaces (APIs), and user interfaces using C++ and Qt, as well as visualizations in OpenGL.

Simulation, Cloud & Tooling

• Physics simulation: Modeling and simulation of physical effects, as well as provision of custom simulation tools.
• Cloud & DevOps: Deployment and integration of solutions in cloud environments (AWS, Microsoft Azure) using modern version control (Git).