Insights and experiences from AI/ML research and production deployments
My Master's thesis at JKU Linz (supervised by Prof. Sepp Hochreiter) introduced class-conditional separation loss into conditional diffusion models used as generative classifiers — improving out-of-distribution detection from 80.25% to 99.03% AUROC on CIFAR-10, an 18.8 percentage-point gain.
I went from a research background (diffusion models, computer vision) to shipping a production AI SaaS — Faultrix, an AI-powered construction quality control platform generating ÖNORM-compliant reports in under 1 minute. Here's what the gap between "AI research" and "AI product" actually looks like.
How I combined YOLOv8 feature extraction with conditional diffusion models to achieve 98.4% defect detection accuracy on inkjet-printed building components in a real-time production environment — part of the Zer0P project funded by the Government of Upper Austria.
The specific architecture choices, data engineering decisions, and production deployment tricks that contributed to 98.4% defect detection accuracy on inkjet-printed building components at PROFACTOR GmbH — not a tutorial, a post-mortem.